Re: [HACKERS] cheaper snapshots redux
On Tue, Sep 13, 2011 at 7:49 AM, Amit Kapila amit.kap...@huawei.com wrote:
Yep, that's pretty much what it does, although xmax is actually
defined as the XID *following* the last one that ended, and I think
xmin needs to also be in xip, so in this case you'd actually end up
with xmin = 15, xmax = 22, xip = { 15, 16, 17, 19 }. But you've got
the basic idea of it.
Shouldn't Xmax be 21 okay as current check in TupleVisibility indicate if
XID is greater than equal to Xmax then it returns tuple is not visible.
No, that's not OK. You stipulated 21 as committed, so it had better be visible.
In particular, if someone with proc-xmin = InvalidTransactionId is
taking a snapshot while you're computing RecentGlobalXmin, and then
stores a proc-xmin less than your newly-computed RecentGlobalXmin,
you've got a problem.
I am assuming here you are reffering to take a snapshot means it has to be
updated in shared memory because otherwise no need to refer proc with your
new design.
Session-1
Updating RecentGlobalXmin during GetSnapshotData() using shared memory copy
of snapshot and completed transactions as RecentGlobalXmin can be updated if
we get xmin.
Session-2
Getting Snapshot to update in shared memory, here it needs to go through
procarray.
Now when it is going through procarray using proclock it can be case that
proc of Session-1 has InvalidTransId, so we will ignore it and go through
remaining session procs.
Now normally Session-1 proc should not get lesser xmin as compare to other
session procs but incase it has got his copy from shared memory ring buffer
before other session procs then it can be lower and which can cause a
problem.
It's not one extra read - you'd have to look at every PGPROC.
If the above explanation is right then is this the reason that to update
RecentGlobalXmin, it has to go through every PGPROC.
Your explanation isn't very clear to me. But I will post the patch
once I have some of these details sorted out.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Yep, that's pretty much what it does, although xmax is actually
defined as the XID *following* the last one that ended, and I think
xmin needs to also be in xip, so in this case you'd actually end up
with xmin = 15, xmax = 22, xip = { 15, 16, 17, 19 }. But you've got
the basic idea of it.
Shouldn't Xmax be 21 okay as current check in TupleVisibility indicate if
XID is greater than equal to Xmax then it returns tuple is not visible.
In particular, if someone with proc-xmin = InvalidTransactionId is
taking a snapshot while you're computing RecentGlobalXmin, and then
stores a proc-xmin less than your newly-computed RecentGlobalXmin,
you've got a problem.
I am assuming here you are reffering to take a snapshot means it has to be
updated in shared memory because otherwise no need to refer proc with your
new design.
Session-1
Updating RecentGlobalXmin during GetSnapshotData() using shared memory copy
of snapshot and completed transactions as RecentGlobalXmin can be updated if
we get xmin.
Session-2
Getting Snapshot to update in shared memory, here it needs to go through
procarray.
Now when it is going through procarray using proclock it can be case that
proc of Session-1 has InvalidTransId, so we will ignore it and go through
remaining session procs.
Now normally Session-1 proc should not get lesser xmin as compare to other
session procs but incase it has got his copy from shared memory ring buffer
before other session procs then it can be lower and which can cause a
problem.
It's not one extra read - you'd have to look at every PGPROC.
If the above explanation is right then is this the reason that to update
RecentGlobalXmin, it has to go through every PGPROC.
***
This e-mail and attachments contain confidential information from HUAWEI,
which is intended only for the person or entity whose address is listed
above. Any use of the information contained herein in any way (including,
but not limited to, total or partial disclosure, reproduction, or
dissemination) by persons other than the intended recipient's) is
prohibited. If you receive this e-mail in error, please notify the sender by
phone or email immediately and delete it!
-Original Message-
From: Robert Haas [mailto:robertmh...@gmail.com]
Sent: Monday, September 12, 2011 9:31 PM
To: Amit Kapila
Cc: pgsql-hackers@postgresql.org
Subject: Re: [HACKERS] cheaper snapshots redux
On Mon, Sep 12, 2011 at 11:07 AM, Amit Kapila amit.kap...@huawei.com
wrote:
If you know what transactions were running the last time a snapshot
summary
was written and what transactions have ended since then, you can work
out
the new xmin on the fly. I have working code for this and it's actually
quite simple.
I believe one method to do same is as follows:
Let us assume at some point of time the snapshot and completed XID list is
somewhat as follows:
Snapshot
{ Xmin 5, Xip[] 8 10 12, Xmax - 15 }
Committed XIDS 8, 10 , 12, 18, 20, 21
So it means 16,17,19 are running transactions. So it will behave as
follows:
{ Xmin 16, Xmax 21, Xip[] 17,19 }
Yep, that's pretty much what it does, although xmax is actually
defined as the XID *following* the last one that ended, and I think
xmin needs to also be in xip, so in this case you'd actually end up
with xmin = 15, xmax = 22, xip = { 15, 16, 17, 19 }. But you've got
the basic idea of it.
But if we do above way to calculate Xmin, we need to check in existing Xip
array and committed Xid array to find Xmin. Wont this cause reasonable
time
even though it is outside lock time if Xip and Xid are large.
Yes, Tom raised this concern earlier. I can't answer it for sure
without benchmarking, but clearly xip[] can't be allowed to get too
big.
Because GetSnapshotData() computes a new value for RecentGlobalXmin by
scanning the ProcArray. This isn't costing a whole lot extra right now
because the xmin and xid fields are normally in the same cache line, so
once you've looked at one of them it doesn't cost that much extra to
look at the other. If, on the other hand, you're not looking at (or even
locking) the
ProcArray, then doing so just to recomputed RecentGlobalXmin sucks.
Yes, this is more time as compare to earlier, but if our approach to
calculate Xmin is like above point, then one extra read outside lock
should
not matter. However if for above point approach is different then it will
be
costlier.
It's not one extra read - you'd have to look at every PGPROC. And it
is not outside a lock, either. You definitely need locking around
computing RecentGlobalXmin; see src/backend/access/transa/README. In
particular, if someone with proc-xmin = InvalidTransactionId is
taking a snapshot while you're computing RecentGlobalXmin, and then
stores a proc-xmin less than your newly-computed RecentGlobalXmin,
you've got a problem. That can't happen right now because no
transactions can commit while
Re: [HACKERS] cheaper snapshots redux
On Sun, Sep 11, 2011 at 11:08 PM, Amit Kapila amit.kap...@huawei.com wrote: In the approach mentioned in your idea, it mentioned that once after taking snapshot, only committed XIDs will be updated and sometimes snapshot itself. So when the xmin will be updated according to your idea as snapshot will not be updated everytime so xmin cannot be latest always. If you know what transactions were running the last time a snapshot summary was written and what transactions have ended since then, you can work out the new xmin on the fly. I have working code for this and it's actually quite simple. RecentGlobalXmin doesn't need to be completely up to date, and in fact recomputing it on every snapshot becomes prohibitively expensive with this approach. I'm still struggling with the best way to handle that. RecentGlobalXmin and RecentXmin are mostly updated with snapshots xmin and that too outside ProcArrayLock, so why it will be expensive if you have updated xmin. Because GetSnapshotData() computes a new value for RecentGlobalXmin by scanning the ProcArray. This isn't costing a whole lot extra right now because the xmin and xid fields are normally in the same cache line, so once you've looked at one of them it doesn't cost that much extra to look at the other. If, on the other hand, you're not looking at (or even locking) the ProcArray, then doing so just to recompute RecentGlobalXmin sucks. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Mon, Sep 12, 2011 at 11:07 AM, Amit Kapila amit.kap...@huawei.com wrote:
If you know what transactions were running the last time a snapshot summary
was written and what transactions have ended since then, you can work out
the new xmin on the fly. I have working code for this and it's actually
quite simple.
I believe one method to do same is as follows:
Let us assume at some point of time the snapshot and completed XID list is
somewhat as follows:
Snapshot
{ Xmin – 5, Xip[] – 8 10 12, Xmax - 15 }
Committed XIDS – 8, 10 , 12, 18, 20, 21
So it means 16,17,19 are running transactions. So it will behave as follows:
{ Xmin – 16, Xmax – 21, Xip[] – 17,19 }
Yep, that's pretty much what it does, although xmax is actually
defined as the XID *following* the last one that ended, and I think
xmin needs to also be in xip, so in this case you'd actually end up
with xmin = 15, xmax = 22, xip = { 15, 16, 17, 19 }. But you've got
the basic idea of it.
But if we do above way to calculate Xmin, we need to check in existing Xip
array and committed Xid array to find Xmin. Won’t this cause reasonable time
even though it is outside lock time if Xip and Xid are large.
Yes, Tom raised this concern earlier. I can't answer it for sure
without benchmarking, but clearly xip[] can't be allowed to get too
big.
Because GetSnapshotData() computes a new value for RecentGlobalXmin by
scanning the ProcArray. This isn't costing a whole lot extra right now
because the xmin and xid fields are normally in the same cache line, so
once you've looked at one of them it doesn't cost that much extra to
look at the other. If, on the other hand, you're not looking at (or even
locking) the
ProcArray, then doing so just to recomputed RecentGlobalXmin sucks.
Yes, this is more time as compare to earlier, but if our approach to
calculate Xmin is like above point, then one extra read outside lock should
not matter. However if for above point approach is different then it will be
costlier.
It's not one extra read - you'd have to look at every PGPROC. And it
is not outside a lock, either. You definitely need locking around
computing RecentGlobalXmin; see src/backend/access/transa/README. In
particular, if someone with proc-xmin = InvalidTransactionId is
taking a snapshot while you're computing RecentGlobalXmin, and then
stores a proc-xmin less than your newly-computed RecentGlobalXmin,
you've got a problem. That can't happen right now because no
transactions can commit while RecentGlobalXmin is being computed, but
the point here is precisely to allow those operations to (mostly) run
in parallel.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
If you know what transactions were running the last time a snapshot summary
was written and what transactions have ended since then, you can work out
the new xmin on the fly. I have working code for this and it's actually
quite simple.
I believe one method to do same is as follows:
Let us assume at some point of time the snapshot and completed XID list is
somewhat as follows:
Snapshot
{
Xmin - 5
Xip[] - 8 10 12
Xmax - 15
}
Committed XIDS - 8, 10 , 12, 18, 20, 21
So it means 16,17,19 are running transactions. So it will behave as follows:
Xmin - 16
Xmax - 21
Xip[] - 17,19
But if we do above way to calculate Xmin, we need to check in existing Xip
array and committed Xid array to find Xmin. Won't this cause reasonable time
even though it is outside lock time if Xip and Xid are large.
Because GetSnapshotData() computes a new value for RecentGlobalXmin by
scanning the ProcArray. This isn't costing a whole lot extra right now
because the xmin and xid fields are normally in the same cache line, so
once you've looked at one of them it doesn't cost that much extra to
look at the other. If, on the other hand, you're not looking at (or even
locking) the
ProcArray, then doing so just to recomputed RecentGlobalXmin sucks.
Yes, this is more time as compare to earlier, but if our approach to
calculate Xmin is like above point, then one extra read outside lock should
not matter. However if for above point approach is different then it will be
costlier.
***
This e-mail and attachments contain confidential information from HUAWEI,
which is intended only for the person or entity whose address is listed
above. Any use of the information contained herein in any way (including,
but not limited to, total or partial disclosure, reproduction, or
dissemination) by persons other than the intended recipient's) is
prohibited. If you receive this e-mail in error, please notify the sender by
phone or email immediately and delete it!
-Original Message-
From: Robert Haas [mailto:robertmh...@gmail.com]
Sent: Monday, September 12, 2011 7:39 PM
To: Amit Kapila
Cc: pgsql-hackers@postgresql.org
Subject: Re: [HACKERS] cheaper snapshots redux
On Sun, Sep 11, 2011 at 11:08 PM, Amit Kapila amit.kap...@huawei.com
wrote:
In the approach mentioned in your idea, it mentioned that once after
taking snapshot, only committed XIDs will be updated and sometimes
snapshot
itself.
So when the xmin will be updated according to your idea as snapshot will
not be updated everytime so xmin cannot be latest always.
If you know what transactions were running the last time a snapshot
summary was written and what transactions have ended since then, you
can work out the new xmin on the fly. I have working code for this
and it's actually quite simple.
RecentGlobalXmin doesn't need to be completely up to date, and in fact
recomputing it on every snapshot becomes prohibitively expensive with this
approach. I'm still struggling with the best way to handle that.
RecentGlobalXmin and RecentXmin are mostly updated with snapshots xmin
and that too outside ProcArrayLock, so why it will be expensive if you
have
updated xmin.
Because GetSnapshotData() computes a new value for RecentGlobalXmin by
scanning the ProcArray. This isn't costing a whole lot extra right
now because the xmin and xid fields are normally in the same cache
line, so once you've looked at one of them it doesn't cost that much
extra to look at the other. If, on the other hand, you're not looking
at (or even locking) the ProcArray, then doing so just to recompute
RecentGlobalXmin sucks.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
Re: [HACKERS] cheaper snapshots redux
4. Won't it effect if we don't update xmin everytime and just noting the committed XIDs. The reason I am asking is that it is used in tuple visibility check so with new idea in some cases instead of just returning from begining by checking xmin it has to go through the committed XID list. I understand that there may be less cases or the improvement by your idea can supesede this minimal effect. However some cases can be defeated. The snapshot xmin has to be up to date. I'm not planning to break that because it would be wrong. In the approach mentioned in your idea, it mentioned that once after taking snapshot, only committed XIDs will be updated and sometimes snapshot itself. So when the xmin will be updated according to your idea as snapshot will not be updated everytime so xmin cannot be latest always. RecentGlobalXmin doesn't need to be completely up to date, and in fact recomputing it on every snapshot becomes prohibitively expensive with this approach. I'm still struggling with the best way to handle that. RecentGlobalXmin and RecentXmin are mostly updated with snapshots xmin and that too outside ProcArrayLock, so why it will be expensive if you have updated xmin. With Regards, Amit Kapila. *** This e-mail and attachments contain confidential information from HUAWEI, which is intended only for the person or entity whose address is listed above. Any use of the information contained herein in any way (including, but not limited to, total or partial disclosure, reproduction, or dissemination) by persons other than the intended recipient's) is prohibited. If you receive this e-mail in error, please notify the sender by phone or email immediately and delete it! -Original Message- From: Robert Haas [mailto:robertmh...@gmail.com] Sent: Thursday, September 08, 2011 7:50 PM To: Amit Kapila Cc: pgsql-hackers@postgresql.org Subject: Re: [HACKERS] cheaper snapshots redux On Tue, Sep 6, 2011 at 11:06 PM, Amit Kapila amit.kap...@huawei.com wrote: 1. With the above, you want to reduce/remove the concurrency issue between the GetSnapshotData() [used at begining of sql command execution] and ProcArrayEndTransaction() [used at end transaction]. The concurrency issue is mainly ProcArrayLock which is taken by GetSnapshotData() in Shared mode and by ProcArrayEndTransaction() in X mode. There may be other instances for similar thing, but this the main thing which you want to resolve. Yep. 2. You want to resolve it by using ring buffer such that readers don't need to take any lock. Yep. Actually, they're still going to need some spinlocks at least in the first go round, to protect the pointers. I'm hoping those can eventually be eliminated on machines with 8-byte atomic reads using appropriate memory barrier primitives. 1. 2 Writers; Won't 2 different sessions who try to commit at same time will get the same write pointer. I assume it will be protected as even indicated in one of your replies as I understood? Yes, commits have to be serialized. No way around that. The best we'll ever be able to do is shorten the critical section. 2. 1 Reader, 1 Writter; It might be case that some body has written a new snapshot and advanced the stop pointer and at that point of time one reader came and read between start pointer and stop pointer. Now the reader will see as follows: snapshot, few XIDs, snapshot So will it handle this situation such that it will only read latest snapshot? In my prototype implementation that can't happen because the start and stop pointers are protected by a single spinlock and are moved simultaneously. But I think we can get rid on machines with 8-byte atomic writes of that and just move the stop pointer first and then the start pointer. If you get more than one snapshot in the middle you just ignore the first part of the data you read and start with the beginning of the last snapshot. 3. How will you detect overwrite. If the write pointer is greater than the start pointer by more than the ring size, you've wrapped. 4. Won't it effect if we don't update xmin everytime and just noting the committed XIDs. The reason I am asking is that it is used in tuple visibility check so with new idea in some cases instead of just returning from begining by checking xmin it has to go through the committed XID list. I understand that there may be less cases or the improvement by your idea can supesede this minimal effect. However some cases can be defeated. The snapshot xmin has to be up to date. I'm not planning to break that because it would be wrong. RecentGlobalXmin doesn't need to be completely up to date, and in fact recomputing it on every snapshot becomes prohibitively expensive with this approach. I'm still struggling with the best way to handle that. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise
Re: [HACKERS] cheaper snapshots redux
4. Won't it effect if we don't update xmin everytime and just noting the committed XIDs. The reason I am asking is that it is used in tuple visibility check so with new idea in some cases instead of just returning from begining by checking xmin it has to go through the committed XID list. I understand that there may be less cases or the improvement by your idea can supesede this minimal effect. However some cases can be defeated. The snapshot xmin has to be up to date. I'm not planning to break that because it would be wrong. In the approach mentioned in your idea, it mentioned that once after taking snapshot, only committed XIDs will be updated and sometimes snapshot itself. So when the xmin will be updated according to your idea. RecentGlobalXmin doesn't need to be completely up to date, and in fact recomputing it on every snapshot becomes prohibitively expensive with this approach. I'm still struggling with the best way to handle that. RecentGlobalXmin and RecentXmin are mostly updated with snapshots xmin and that too outside ProcArrayLock, so why it will be expensive if you have updated xmin. With Regards, Amit Kapila. *** This e-mail and attachments contain confidential information from HUAWEI, which is intended only for the person or entity whose address is listed above. Any use of the information contained herein in any way (including, but not limited to, total or partial disclosure, reproduction, or dissemination) by persons other than the intended recipient's) is prohibited. If you receive this e-mail in error, please notify the sender by phone or email immediately and delete it!
Re: [HACKERS] cheaper snapshots redux
On Tue, Sep 6, 2011 at 11:06 PM, Amit Kapila amit.kap...@huawei.com wrote: 1. With the above, you want to reduce/remove the concurrency issue between the GetSnapshotData() [used at begining of sql command execution] and ProcArrayEndTransaction() [used at end transaction]. The concurrency issue is mainly ProcArrayLock which is taken by GetSnapshotData() in Shared mode and by ProcArrayEndTransaction() in X mode. There may be other instances for similar thing, but this the main thing which you want to resolve. Yep. 2. You want to resolve it by using ring buffer such that readers don't need to take any lock. Yep. Actually, they're still going to need some spinlocks at least in the first go round, to protect the pointers. I'm hoping those can eventually be eliminated on machines with 8-byte atomic reads using appropriate memory barrier primitives. 1. 2 Writers; Won't 2 different sessions who try to commit at same time will get the same write pointer. I assume it will be protected as even indicated in one of your replies as I understood? Yes, commits have to be serialized. No way around that. The best we'll ever be able to do is shorten the critical section. 2. 1 Reader, 1 Writter; It might be case that some body has written a new snapshot and advanced the stop pointer and at that point of time one reader came and read between start pointer and stop pointer. Now the reader will see as follows: snapshot, few XIDs, snapshot So will it handle this situation such that it will only read latest snapshot? In my prototype implementation that can't happen because the start and stop pointers are protected by a single spinlock and are moved simultaneously. But I think we can get rid on machines with 8-byte atomic writes of that and just move the stop pointer first and then the start pointer. If you get more than one snapshot in the middle you just ignore the first part of the data you read and start with the beginning of the last snapshot. 3. How will you detect overwrite. If the write pointer is greater than the start pointer by more than the ring size, you've wrapped. 4. Won't it effect if we don't update xmin everytime and just noting the committed XIDs. The reason I am asking is that it is used in tuple visibility check so with new idea in some cases instead of just returning from begining by checking xmin it has to go through the committed XID list. I understand that there may be less cases or the improvement by your idea can supesede this minimal effect. However some cases can be defeated. The snapshot xmin has to be up to date. I'm not planning to break that because it would be wrong. RecentGlobalXmin doesn't need to be completely up to date, and in fact recomputing it on every snapshot becomes prohibitively expensive with this approach. I'm still struggling with the best way to handle that. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
I wanted to clarify my understanding and have some doubts. What I'm thinking about instead is using a ring buffer with three pointers: a start pointer, a stop pointer, and a write pointer. When a transaction ends, we advance the write pointer, write the XIDs or a whole new snapshot into the buffer, and then advance the stop pointer. If we wrote a whole new snapshot, we advance the start pointer to the beginning of the data we just wrote. Someone who wants to take a snapshot must read the data between the start and stop pointers, and must then check that the write pointer hasn't advanced so far in the meantime that the data they read might have been overwritten before they finished reading it. Obviously, that's a little risky, since we'll have to do the whole thing over if a wraparound occurs, but if the ring buffer is large enough it shouldn't happen very often. Clarification -- 1. With the above, you want to reduce/remove the concurrency issue between the GetSnapshotData() [used at begining of sql command execution] and ProcArrayEndTransaction() [used at end transaction]. The concurrency issue is mainly ProcArrayLock which is taken by GetSnapshotData() in Shared mode and by ProcArrayEndTransaction() in X mode. There may be other instances for similar thing, but this the main thing which you want to resolve. 2. You want to resolve it by using ring buffer such that readers don't need to take any lock. Is my above understanding correct? Doubts 1. 2 Writers; Won't 2 different sessions who try to commit at same time will get the same write pointer. I assume it will be protected as even indicated in one of your replies as I understood? 2. 1 Reader, 1 Writter; It might be case that some body has written a new snapshot and advanced the stop pointer and at that point of time one reader came and read between start pointer and stop pointer. Now the reader will see as follows: snapshot, few XIDs, snapshot So will it handle this situation such that it will only read latest snapshot? 3. How will you detect overwrite. 4. Won't it effect if we don't update xmin everytime and just noting the committed XIDs. The reason I am asking is that it is used in tuple visibility check so with new idea in some cases instead of just returning from begining by checking xmin it has to go through the committed XID list. I understand that there may be less cases or the improvement by your idea can supesede this minimal effect. However some cases can be defeated. -- With Regards, Amit Kapila. *** This e-mail and attachments contain confidential information from HUAWEI, which is intended only for the person or entity whose address is listed above. Any use of the information contained herein in any way (including, but not limited to, total or partial disclosure, reproduction, or dissemination) by persons other than the intended recipient's) is prohibited. If you receive this e-mail in error, please notify the sender by phone or email immediately and delete it! -Original Message- From: pgsql-hackers-ow...@postgresql.org [mailto:pgsql-hackers-ow...@postgresql.org] On Behalf Of Robert Haas Sent: Sunday, August 28, 2011 7:17 AM To: Gokulakannan Somasundaram Cc: pgsql-hackers@postgresql.org Subject: Re: [HACKERS] cheaper snapshots redux On Sat, Aug 27, 2011 at 1:38 AM, Gokulakannan Somasundaram gokul...@gmail.com wrote: First i respectfully disagree with you on the point of 80MB. I would say that its very rare that a small system( with 1 GB RAM ) might have a long running transaction sitting idle, while 10 million transactions are sitting idle. Should an optimization be left, for the sake of a very small system to achieve high enterprise workloads? With the design where you track commit-visbility sequence numbers instead of snapshots, you wouldn't need 10 million transactions that were all still running. You would just need a snapshot that had been sitting around while 10 million transactions completed meanwhile. That having been said, I don't necessarily think that design is doomed. I just think it's going to be trickier to get working than the design I'm now hacking on, and a bigger change from what we do now. If this doesn't pan out, I might try that one, or something else. Second, if we make use of the memory mapped files, why should we think, that all the 80MB of data will always reside in memory? Won't they get paged out by the operating system, when it is in need of memory? Or do you have some specific OS in mind? No, I don't think it will all be in memory - but that's part of the performance calculation. If you need to check on the status of an XID and find that you need to read a page of data in from disk, that's going to be many orders of magnitude slower than anything we do with s snapshot now. Now, if you gain enough elsewhere, it could still
Re: [HACKERS] cheaper snapshots redux
No, I don't think it will all be in memory - but that's part of the performance calculation. If you need to check on the status of an XID and find that you need to read a page of data in from disk, that's going to be many orders of magnitude slower than anything we do with s snapshot now. Now, if you gain enough elsewhere, it could still be a win, but I'm not going to just assume that. I was just suggesting this, because the memory costs have come down a lot(as you may know) and people can afford to buy more memory in enterprise scenario. We may not need to worry about MBs of memory, especially with the cloud computing being widely adopted, when we get scalability. Gokul.
Re: [HACKERS] cheaper snapshots redux
On Sun, Aug 28, 2011 at 4:33 AM, Gokulakannan Somasundaram gokul...@gmail.com wrote: No, I don't think it will all be in memory - but that's part of the performance calculation. If you need to check on the status of an XID and find that you need to read a page of data in from disk, that's going to be many orders of magnitude slower than anything we do with s snapshot now. Now, if you gain enough elsewhere, it could still be a win, but I'm not going to just assume that. I was just suggesting this, because the memory costs have come down a lot(as you may know) and people can afford to buy more memory in enterprise scenario. We may not need to worry about MBs of memory, especially with the cloud computing being widely adopted, when we get scalability. The proof of the pudding is in the eating, so let me finish coding up this approach and see how it works. Then we can decide where to go next... -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Sat, Aug 27, 2011 at 1:38 AM, Gokulakannan Somasundaram gokul...@gmail.com wrote: First i respectfully disagree with you on the point of 80MB. I would say that its very rare that a small system( with 1 GB RAM ) might have a long running transaction sitting idle, while 10 million transactions are sitting idle. Should an optimization be left, for the sake of a very small system to achieve high enterprise workloads? With the design where you track commit-visbility sequence numbers instead of snapshots, you wouldn't need 10 million transactions that were all still running. You would just need a snapshot that had been sitting around while 10 million transactions completed meanwhile. That having been said, I don't necessarily think that design is doomed. I just think it's going to be trickier to get working than the design I'm now hacking on, and a bigger change from what we do now. If this doesn't pan out, I might try that one, or something else. Second, if we make use of the memory mapped files, why should we think, that all the 80MB of data will always reside in memory? Won't they get paged out by the operating system, when it is in need of memory? Or do you have some specific OS in mind? No, I don't think it will all be in memory - but that's part of the performance calculation. If you need to check on the status of an XID and find that you need to read a page of data in from disk, that's going to be many orders of magnitude slower than anything we do with s snapshot now. Now, if you gain enough elsewhere, it could still be a win, but I'm not going to just assume that. As I play with this, I'm coming around to the conclusion that, in point of fact, the thing that's hard about snapshots has a great deal more to do with memory than it does with CPU time. Sure, using the snapshot has to be cheap. But it already IS cheap. We don't need to fix that problem; we just need to not break it. What's not cheap is constructing the snapshot - principally because of ProcArrayLock, and secondarily because we're grovelling through fairly large amounts of shared memory to get all the XIDs we need. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Thu, Aug 25, 2011 at 6:24 PM, Jim Nasby j...@nasby.net wrote: On Aug 25, 2011, at 8:24 AM, Robert Haas wrote: My hope (and it might turn out that I'm an optimist) is that even with a reasonably small buffer it will be very rare for a backend to experience a wraparound condition. For example, consider a buffer with ~6500 entries, approximately 64 * MaxBackends, the approximate size of the current subxip arrays taken in aggregate. I hypothesize that a typical snapshot on a running system is going to be very small - a handful of XIDs at most - because, on the average, transactions are going to commit in *approximately* increasing XID order and, if you take the regression tests as representative of a real workload, only a small fraction of transactions will have more than one XID. So BTW, there's a way to actually gather some data on this by using PgQ (part of Skytools and used by Londiste). PgQ works by creating ticks at regular intervals, where a tick is basically just a snapshot of committed XIDs. Presumably Slony does something similar. I can provide you with sample data from our production systems if you're interested. Yeah, that would be great. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Thu, Aug 25, 2011 at 6:29 PM, Jim Nasby j...@nasby.net wrote: Actually, I wasn't thinking about the system dynamically sizing shared memory on it's own... I was only thinking of providing the ability for a user to change something like shared_buffers and allow that change to take effect with a SIGHUP instead of requiring a full restart. I agree. That would be awesome. Sadly, I don't have time to work on it. :-( -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Tue, Aug 23, 2011 at 5:25 AM, Robert Haas robertmh...@gmail.com wrote: I've been giving this quite a bit more thought, and have decided to abandon the scheme described above, at least for now. It has the advantage of avoiding virtually all locking, but it's extremely inefficient in its use of memory in the presence of long-running transactions. For example, if there's an open transaction that's been sitting around for 10 million transactions or so and has an XID assigned, any new snapshot is going to need to probe into the big array for any XID in that range. At 8 bytes per entry, that means we're randomly accessing about ~80MB of memory-mapped data. That seems problematic both in terms of blowing out the cache and (on small machines) possibly even blowing out RAM. Nor is that the worst case scenario: a transaction could sit open for 100 million transactions. First i respectfully disagree with you on the point of 80MB. I would say that its very rare that a small system( with 1 GB RAM ) might have a long running transaction sitting idle, while 10 million transactions are sitting idle. Should an optimization be left, for the sake of a very small system to achieve high enterprise workloads? Second, if we make use of the memory mapped files, why should we think, that all the 80MB of data will always reside in memory? Won't they get paged out by the operating system, when it is in need of memory? Or do you have some specific OS in mind? Thanks, Gokul.
Re: [HACKERS] cheaper snapshots redux
On Thu, Aug 25, 2011 at 1:55 AM, Markus Wanner mar...@bluegap.ch wrote: One difference with snapshots is that only the latest snapshot is of any interest. Theoretically, yes. But as far as I understood, you proposed the backends copy that snapshot to local memory. And copying takes some amount of time, possibly being interrupted by other backends which add newer snapshots... Or do you envision the copying to restart whenever a new snapshot arrives? My hope (and it might turn out that I'm an optimist) is that even with a reasonably small buffer it will be very rare for a backend to experience a wraparound condition. For example, consider a buffer with ~6500 entries, approximately 64 * MaxBackends, the approximate size of the current subxip arrays taken in aggregate. I hypothesize that a typical snapshot on a running system is going to be very small - a handful of XIDs at most - because, on the average, transactions are going to commit in *approximately* increasing XID order and, if you take the regression tests as representative of a real workload, only a small fraction of transactions will have more than one XID. So it seems believable to think that the typical snapshot on a machine with max_connections=100 might only be ~10 XIDs, even if none of the backends are read-only. So the backend taking a snapshot only needs to be able to copy ~64 bytes of information from the ring buffer before other backends write ~27k of data into that buffer, likely requiring hundreds of other commits. That seems vanishingly unlikely; memcpy() is very fast. If it does happen, you can recover by retrying, but it should be a once-in-a-blue-moon kind of thing. I hope. Now, as the size of the snapshot gets bigger, things will eventually become less good. For example if you had a snapshot with 6000 XIDs in it then every commit would need to write over the previous snapshot and things would quickly deteriorate. But you can cope with that situation using the same mechanism we already use to handle big snapshots: toss out all the subtransaction IDs, mark the snapshot as overflowed, and just keep the toplevel XIDs. Now you've got at most ~100 XIDs to worry about, so you're back in the safety zone. That's not ideal in the sense that you will cause more pg_subtrans lookups, but that's the price you pay for having a gazillion subtransactions floating around, and any system is going to have to fall back on some sort of mitigation strategy at some point. There's no useful limit on the number of subxids a transaction can have, so unless you're prepared to throw an unbounded amount of memory at the problem you're going to eventually have to punt. It seems to me that the problem case is when you are just on the edge. Say you have 1400 XIDs in the snapshot. If you compact the snapshot down to toplevel XIDs, most of those will go away and you won't have to worry about wraparound - but you will pay a performance penalty in pg_subtrans lookups. On the other hand, if you don't compact the snapshot, it's not that hard to imagine a wraparound occurring - four snapshot rewrites could wrap the buffer. You would still hope that memcpy() could finish in time, but if you're rewriting 1400 XIDs with any regularity, it might not take that many commits to throw a spanner into the works. If the system is badly overloaded and the backend trying to take a snapshot gets descheduled for a long time at just the wrong moment, it doesn't seem hard to imagine a wraparound happening. Now, it's not hard to recover from a wraparound. In fact, we can pretty easily guarantee that any given attempt to take a snapshot will suffer a wraparound at most once. The writers (who are committing) have to be serialized anyway, so anyone who suffers a wraparound can just grab the same lock in shared mode and retry its snapshot. Now concurrency decreases significantly, because no one else is allowed to commit until that guy has got his snapshot, but right now that's true *every time* someone wants to take a snapshot, so falling back to that strategy occasionally doesn't seem prohibitively bad. However, you don't want it to happen very often, because even leaving aside the concurrency hit, it's double work: you have to try to take a snapshot, realize you've had a wraparound, and then retry. It seems pretty clear that with a big enough ring buffer the wraparound problem will become so infrequent as to be not worth worrying about. I'm theorizing that even with a quite small ring buffer the problem will still be infrequent enough not to worry about, but that might be optimistic. I think I'm going to need some kind of test case that generates very large, frequently changing snapshots. Of course even if wraparound turns out not to be a problem there are other things that could scuttle this whole approach, but I think the idea has enough potential to be worth testing. If the whole thing crashes and burns I hope I'll at least learn enough along the
Re: [HACKERS] cheaper snapshots redux
Robert, On 08/25/2011 03:24 PM, Robert Haas wrote: My hope (and it might turn out that I'm an optimist) is that even with a reasonably small buffer it will be very rare for a backend to experience a wraparound condition. It certainly seems less likely than with the ring-buffer for imessages, yes. Note, however, that for imessages, I've also had the policy in place that a backend *must* consume its message before sending any. And that I took great care for all receivers to consume their messages as early as possible. None the less, I kept incrementing the buffer size (to multiple megabytes) to make this work. Maybe I'm overcautious because of that experience. - a handful of XIDs at most - because, on the average, transactions are going to commit in *approximately* increasing XID order This assumption quickly turns false, if you happen to have just one long-running transaction, I think. Or in general, if transaction duration varies a lot. So the backend taking a snapshot only needs to be able to copy ~64 bytes of information from the ring buffer before other backends write ~27k of data into that buffer, likely requiring hundreds of other commits. You said earlier, that only the latest snapshot is required. It takes only a single commit for such a snapshot to not be the latest anymore. Instead, if you keep around older snapshots for some time - as what your description here implies - readers are free to copy from those older snapshots while other backends are able to make progress concurrently (writers or readers of other snapshots). However, that either requires keeping track of readers of a certain snapshot (reference counting) or - as I understand your description - you simply invalidate all concurrent readers upon wrap-around, or something. That seems vanishingly unlikely; Agreed. Now, as the size of the snapshot gets bigger, things will eventually become less good. Also keep configurations with increased max_connections in mind. With that, we not only the snapshots get bigger, but more processes have to share CPU time, on avg. making memcpy slower for a single process. Of course even if wraparound turns out not to be a problem there are other things that could scuttle this whole approach, but I think the idea has enough potential to be worth testing. If the whole thing crashes and burns I hope I'll at least learn enough along the way to design something better... That's always a good motivation. In that sense: happy hacking! Regards Markus Wanner -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Thu, Aug 25, 2011 at 10:19 AM, Markus Wanner mar...@bluegap.ch wrote: Note, however, that for imessages, I've also had the policy in place that a backend *must* consume its message before sending any. And that I took great care for all receivers to consume their messages as early as possible. None the less, I kept incrementing the buffer size (to multiple megabytes) to make this work. Maybe I'm overcautious because of that experience. What's a typical message size for imessages? - a handful of XIDs at most - because, on the average, transactions are going to commit in *approximately* increasing XID order This assumption quickly turns false, if you happen to have just one long-running transaction, I think. Or in general, if transaction duration varies a lot. Well, one long-running transaction that only has a single XID is not really a problem: the snapshot is still small. But one very old transaction that also happens to have a large number of subtransactions all of which have XIDs assigned might be a good way to stress the system. So the backend taking a snapshot only needs to be able to copy ~64 bytes of information from the ring buffer before other backends write ~27k of data into that buffer, likely requiring hundreds of other commits. You said earlier, that only the latest snapshot is required. It takes only a single commit for such a snapshot to not be the latest anymore. Instead, if you keep around older snapshots for some time - as what your description here implies - readers are free to copy from those older snapshots while other backends are able to make progress concurrently (writers or readers of other snapshots). However, that either requires keeping track of readers of a certain snapshot (reference counting) or - as I understand your description - you simply invalidate all concurrent readers upon wrap-around, or something. Each reader decides which data he needs to copy from the buffer, and then copies it, and then checks whether any of it got overwritten before the copy was completed. So there's a lively possibility that the snapshot that was current when the reader began copying it will no longer be current by the time he finishes copying it, because a commit has intervened. That's OK: it just means that, effectively, the snapshot is taken at the moment the start and stop pointers are read, and won't take into account any commits that happen later, which is exactly what a snapshot is supposed to do anyway. There is a hopefully quite small possibility that by the time the reader finishes copying it so much new data will have been written to the buffer that it will have wrapped around and clobbered the portion the reader was interested in. That needs to be rare. Now, as the size of the snapshot gets bigger, things will eventually become less good. Also keep configurations with increased max_connections in mind. With that, we not only the snapshots get bigger, but more processes have to share CPU time, on avg. making memcpy slower for a single process. Right. I'm imagining making the default buffer size proportional to max_connections. Of course even if wraparound turns out not to be a problem there are other things that could scuttle this whole approach, but I think the idea has enough potential to be worth testing. If the whole thing crashes and burns I hope I'll at least learn enough along the way to design something better... That's always a good motivation. In that sense: happy hacking! Thanks. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Robert Haas robertmh...@gmail.com writes: Well, one long-running transaction that only has a single XID is not really a problem: the snapshot is still small. But one very old transaction that also happens to have a large number of subtransactions all of which have XIDs assigned might be a good way to stress the system. That's a good point. If the ring buffer size creates a constraint on the maximum number of sub-XIDs per transaction, you're going to need a fallback path of some sort. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Robert, On 08/25/2011 04:48 PM, Robert Haas wrote: What's a typical message size for imessages? Most message types in Postgres-R are just a couple bytes in size. Others, especially change sets, can be up to 8k. However, I think you'll have an easier job guaranteeing that backends consume their portions of the ring-buffer in time. Plus wrap-around isn't that much of a problem in your case. (I couldn't drop imessage, but had to let senders wait). Well, one long-running transaction that only has a single XID is not really a problem: the snapshot is still small. But one very old transaction that also happens to have a large number of subtransactions all of which have XIDs assigned might be a good way to stress the system. Ah, right, that's why its a list of transactions in progress and not a list of completed transactions in SnapshotData... good. Each reader decides which data he needs to copy from the buffer, and then copies it, and then checks whether any of it got overwritten before the copy was completed. So there's a lively possibility that the snapshot that was current when the reader began copying it will no longer be current by the time he finishes copying it, because a commit has intervened. That's OK: it just means that, effectively, the snapshot is taken at the moment the start and stop pointers are read, and won't take into account any commits that happen later, which is exactly what a snapshot is supposed to do anyway. Agreed, that makes sense. Thanks for explaining. There is a hopefully quite small possibility that by the time the reader finishes copying it so much new data will have been written to the buffer that it will have wrapped around and clobbered the portion the reader was interested in. That needs to be rare. Yeah. Regards Markus Wanner -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Tom, On 08/25/2011 04:59 PM, Tom Lane wrote: That's a good point. If the ring buffer size creates a constraint on the maximum number of sub-XIDs per transaction, you're going to need a fallback path of some sort. I think Robert envisions the same fallback path we already have: subxids.overflowed. Regards Markus Wanner -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Thu, Aug 25, 2011 at 11:15 AM, Markus Wanner mar...@bluegap.ch wrote: On 08/25/2011 04:59 PM, Tom Lane wrote: That's a good point. If the ring buffer size creates a constraint on the maximum number of sub-XIDs per transaction, you're going to need a fallback path of some sort. I think Robert envisions the same fallback path we already have: subxids.overflowed. I have a slightly more nuanced idea, but basically yes. The trouble is that if you're keeping the snapshot around and updating it (rather than scanning the ProcArray each time) you need some sort of mechanism for the snapshot to eventually un-overflow. Otherwise, the first overflow leaves you in the soup for the entire lifetime of the cluster. What I have in mind is to store the highest subxid that has been removed from the snapshot, or InvalidTransactonId if we know the snapshot is complete. Whenever the highest removed subxid falls behind xmin, we can reset it to InvalidTransactionId. It would be sensible for clients to store the exact value of highest_removed_subxid in their snapshots as well, instead of just a Boolean flag. A pg_subtrans lookup is needed only for XIDs which are greater than xmin and less than or equal to highest_removed_subxid. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Aug 25, 2011, at 8:24 AM, Robert Haas wrote: My hope (and it might turn out that I'm an optimist) is that even with a reasonably small buffer it will be very rare for a backend to experience a wraparound condition. For example, consider a buffer with ~6500 entries, approximately 64 * MaxBackends, the approximate size of the current subxip arrays taken in aggregate. I hypothesize that a typical snapshot on a running system is going to be very small - a handful of XIDs at most - because, on the average, transactions are going to commit in *approximately* increasing XID order and, if you take the regression tests as representative of a real workload, only a small fraction of transactions will have more than one XID. So BTW, there's a way to actually gather some data on this by using PgQ (part of Skytools and used by Londiste). PgQ works by creating ticks at regular intervals, where a tick is basically just a snapshot of committed XIDs. Presumably Slony does something similar. I can provide you with sample data from our production systems if you're interested. -- Jim C. Nasby, Database Architect j...@nasby.net 512.569.9461 (cell) http://jim.nasby.net -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Aug 22, 2011, at 6:22 PM, Robert Haas wrote: With respect to a general-purpose shared memory allocator, I think that there are cases where that would be useful to have, but I don't think there are as many of them as many people seem to think. I wouldn't choose to implement this using a general-purpose allocator even if we had it, both because it's undesirable to allow this or any subsystem to consume an arbitrary amount of memory (nor can it fail... especially in the abort path) and because a ring buffer is almost certainly faster than a general-purpose allocator. We have enough trouble with palloc overhead already. That having been said, I do think there are cases where it would be nice to have... and it wouldn't surprise me if I end up working on something along those lines in the next year or so. It turns out that memory management is a major issue in lock-free programming; you can't assume that it's safe to recycle an object once the last pointer to it has been removed from shared memory - because someone may have fetched the pointer just before you removed it and still be using it to examine the object. An allocator with some built-in capabilities for handling such problems seems like it might be very useful Actually, I wasn't thinking about the system dynamically sizing shared memory on it's own... I was only thinking of providing the ability for a user to change something like shared_buffers and allow that change to take effect with a SIGHUP instead of requiring a full restart. I agree that we'd have to be very careful with allowing the code to start changing shared memory size on it's own... -- Jim C. Nasby, Database Architect j...@nasby.net 512.569.9461 (cell) http://jim.nasby.net -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Hello Dimitri, On 08/23/2011 06:39 PM, Dimitri Fontaine wrote: I'm far from familiar with the detailed concepts here, but allow me to comment. I have two open questions: - is it possible to use a distributed algorithm to produce XIDs, something like Vector Clocks? Then each backend is able to create a snapshot (well, and XID) on its own, and any backend is still able to compare its snapshot to any other snapshot (well, XID) Creation of snapshots and XID assignment are not as related as you imply here. Keep in mind that a read-only transaction have a snapshot, but no XID. (Not sure if it's possible for a transaction to have an XID, but no snapshot. If it only touches system catalogs with SnapshotNow, maybe? Don't think we support that, ATM). - is it possible to cache the production of the next snapshots so that generating an XID only means getting the next in a pre-computed vector? The way I look at it, what Robert proposed can be thought of as cache the production of the next snapshot, with a bit of a stretch of what a cache is, perhaps. I'd rather call it early snapshot creation, maybe look-ahead something. ATM backends all scan ProcArray to generate their snapshot. Instead, what Robert proposes would - sometimes, somewhat - move that work from snapshot creation time to commit time. As Tom points out, the difficulty lies in the question of when it's worth doing that: if you have lots of commits in a row, and no transaction ever uses the (pre generated) snapshots of the point in time in between, then those were wasted. OTOH, if there are just very few COMMITs spread across lots of writes, the read-only backends will re-create the same snapshots, over and over again. Seems wasteful as well (as GetSnapshotData popping up high on profiles confirms somewhat). Hope to have cleared up things a bit. Regards Markus -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Robert, Jim, thanks for thinking out loud about dynamic allocation of shared memory. Very much appreciated. On 08/23/2011 01:22 AM, Robert Haas wrote: With respect to a general-purpose shared memory allocator, I think that there are cases where that would be useful to have, but I don't think there are as many of them as many people seem to think. I wouldn't choose to implement this using a general-purpose allocator even if we had it, both because it's undesirable to allow this or any subsystem to consume an arbitrary amount of memory (nor can it fail... especially in the abort path) and because a ring buffer is almost certainly faster than a general-purpose allocator. I'm in respectful disagreement regarding the ring-buffer approach and think that dynamic allocation can actually be more efficient if done properly, because there doesn't need to be head and tail pointers, which might turn into a point of contention. As a side note: that I've been there with imessages. Those were first organized as a ring-bufffer. The major problem with that approach was the imessages were consumed with varying delay. In case an imessage was left there for a longer amount of time, it blocked creation of new imessages, because the ring-buffer cycled around once and its head arrived back at the unconsumed imessage. IIUC (which might not be the case) the same issue applies for snapshots. Regards Markus Wanner -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Wed, Aug 24, 2011 at 4:30 AM, Markus Wanner mar...@bluegap.ch wrote: I'm in respectful disagreement regarding the ring-buffer approach and think that dynamic allocation can actually be more efficient if done properly, because there doesn't need to be head and tail pointers, which might turn into a point of contention. True; although there are some other complications. With a sufficiently sophisticated allocator you can avoid mutex contention when allocating chunks, but then you have to store a pointer to the chunk somewhere or other, and that then requires some kind of synchronization. As a side note: that I've been there with imessages. Those were first organized as a ring-bufffer. The major problem with that approach was the imessages were consumed with varying delay. In case an imessage was left there for a longer amount of time, it blocked creation of new imessages, because the ring-buffer cycled around once and its head arrived back at the unconsumed imessage. IIUC (which might not be the case) the same issue applies for snapshots. One difference with snapshots is that only the latest snapshot is of any interest. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Robert, On 08/25/2011 04:59 AM, Robert Haas wrote: True; although there are some other complications. With a sufficiently sophisticated allocator you can avoid mutex contention when allocating chunks, but then you have to store a pointer to the chunk somewhere or other, and that then requires some kind of synchronization. Hm.. right. One difference with snapshots is that only the latest snapshot is of any interest. Theoretically, yes. But as far as I understood, you proposed the backends copy that snapshot to local memory. And copying takes some amount of time, possibly being interrupted by other backends which add newer snapshots... Or do you envision the copying to restart whenever a new snapshot arrives? Regards Markus -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Mon, Aug 22, 2011 at 10:25 PM, Robert Haas robertmh...@gmail.com wrote: I've been giving this quite a bit more thought, and have decided to abandon the scheme described above, at least for now. I liked your goal of O(1) snapshots and think you should go for that. I didn't realise you were still working on this, and had some thoughts at the weekend which I recorded just now. Different tack entirely. Heikki has made the suggestion a few times (and a few other people have since made somewhat similar suggestions in different words) of keeping an-up-to-date snapshot in shared memory such that transactions that need a snapshot can simply copy it. I've since noted that in Hot Standby mode, that's more or less what the KnownAssignedXids stuff already does. I objected that, first, the overhead of updating the snapshot for every commit would be too great, and second, it didn't seem to do a whole lot to reduce the size of the critical section, and therefore probably wouldn't improve performance that much. But I'm coming around to the view that these might be solvable problems rather than reasons to give up on the idea altogether. Sounds easy enough to just link up KnownAssignedXids and see... -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Robert Haas robertmh...@gmail.com writes: With respect to the first problem, what I'm imagining is that we not do a complete rewrite of the snapshot in shared memory on every commit. Instead, when a transaction ends, we'll decide whether to (a) write a new snapshot or (b) just record the XIDs that ended. If we do (b), then any backend that wants a snapshot will need to copy from shared memory both the most recently written snapshot and the XIDs that have subsequently ended. From there, it can figure out which XIDs are still running. Of course, if the list of recently-ended XIDs gets too long, then taking a snapshot will start to get expensive, so we'll need to periodically do (a) instead. There are other ways that this could be done as well; for example, the KnownAssignedXids stuff just flags XIDs that should be ignored and then periodically compacts away the ignored entries. I'm a bit concerned that this approach is trying to optimize the heavy contention situation at the cost of actually making things worse anytime that you're not bottlenecked by contention for access to this shared data structure. In particular, given the above design, then every reader of the data structure has to duplicate the work of eliminating subsequently-ended XIDs from the latest stored snapshot. Maybe that's relatively cheap, but if you do it N times it's not going to be so cheap anymore. In fact, it looks to me like that cost would scale about as O(N^2) in the number of transactions you allow to elapse before storing a new snapshot, so you're not going to be able to let very many go by before you do that. I don't say this can't be made to work, but I don't want to blow off performance for single-threaded applications in pursuit of scalability that will only benefit people running massively parallel applications on big iron. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Tue, Aug 23, 2011 at 12:13 PM, Tom Lane t...@sss.pgh.pa.us wrote: I'm a bit concerned that this approach is trying to optimize the heavy contention situation at the cost of actually making things worse anytime that you're not bottlenecked by contention for access to this shared data structure. In particular, given the above design, then every reader of the data structure has to duplicate the work of eliminating subsequently-ended XIDs from the latest stored snapshot. Maybe that's relatively cheap, but if you do it N times it's not going to be so cheap anymore. In fact, it looks to me like that cost would scale about as O(N^2) in the number of transactions you allow to elapse before storing a new snapshot, so you're not going to be able to let very many go by before you do that. That's certainly a fair concern, and it might even be worse than O(n^2). On the other hand, the current approach involves scanning the entire ProcArray for every snapshot, even if nothing has changed and 90% of the backends are sitting around playing tiddlywinks, so I don't think I'm giving up something for nothing except perhaps in the case where there is only one active backend in the entire system. On the other hand, you could be entirely correct that the current implementation wins in the uncontended case. Without testing it, I just don't know... -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Robert Haas robertmh...@gmail.com writes: I think the real trick is figuring out a design that can improve concurrency. I'm far from familiar with the detailed concepts here, but allow me to comment. I have two open questions: - is it possible to use a distributed algorithm to produce XIDs, something like Vector Clocks? Then each backend is able to create a snapshot (well, and XID) on its own, and any backend is still able to compare its snapshot to any other snapshot (well, XID) - is it possible to cache the production of the next snapshots so that generating an XID only means getting the next in a pre-computed vector? My guess by reading the emails here is that we need to add some information at snapshot generation time, it's not just about getting a 32 bit sequence number. I'm not sure I'm being that helpful here, but sometime stating the obviously impossible ideas allows to think about some new design, so I figured I would still try :) Regards, -- Dimitri Fontaine http://2ndQuadrant.fr PostgreSQL : Expertise, Formation et Support -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
Robert Haas robertmh...@gmail.com writes: That's certainly a fair concern, and it might even be worse than O(n^2). On the other hand, the current approach involves scanning the entire ProcArray for every snapshot, even if nothing has changed and 90% of the backends are sitting around playing tiddlywinks, so I don't think I'm giving up something for nothing except perhaps in the case where there is only one active backend in the entire system. On the other hand, you could be entirely correct that the current implementation wins in the uncontended case. Without testing it, I just don't know... Sure. Like I said, I don't know that this can't be made to work. I'm just pointing out that we have to keep an eye on the single-backend case as well as the many-backends case. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
[HACKERS] cheaper snapshots redux
On Wed, Jul 27, 2011 at 10:51 PM, Robert Haas robertmh...@gmail.com wrote: On Wed, Oct 20, 2010 at 10:07 PM, Tom Lane t...@sss.pgh.pa.us wrote: I wonder whether we could do something involving WAL properties --- the current tuple visibility logic was designed before WAL existed, so it's not exploiting that resource at all. I'm imagining that the kernel of a snapshot is just a WAL position, ie the end of WAL as of the time you take the snapshot (easy to get in O(1) time). Visibility tests then reduce to did this transaction commit with a WAL record located before the specified position?. You'd need some index datastructure that made it reasonably cheap to find out the commit locations of recently committed transactions, where recent means back to recentGlobalXmin. That seems possibly do-able, though I don't have a concrete design in mind. [discussion of why I don't think an LSN will work] But having said that an LSN can't work, I don't see why we can't just use a 64-bit counter. In fact, the predicate locking code already does something much like this, using an SLRU, for serializable transactions only. In more detail, what I'm imagining is an array with 4 billion entries, one per XID, probably broken up into files of say 16MB each with 2 million entries per file. Each entry is a 64-bit value. It is 0 if the XID has not yet started, is still running, or has aborted. Otherwise, it is the commit sequence number of the transaction. I've been giving this quite a bit more thought, and have decided to abandon the scheme described above, at least for now. It has the advantage of avoiding virtually all locking, but it's extremely inefficient in its use of memory in the presence of long-running transactions. For example, if there's an open transaction that's been sitting around for 10 million transactions or so and has an XID assigned, any new snapshot is going to need to probe into the big array for any XID in that range. At 8 bytes per entry, that means we're randomly accessing about ~80MB of memory-mapped data. That seems problematic both in terms of blowing out the cache and (on small machines) possibly even blowing out RAM. Nor is that the worst case scenario: a transaction could sit open for 100 million transactions. Heikki has made the suggestion a few times (and a few other people have since made somewhat similar suggestions in different words) of keeping an-up-to-date snapshot in shared memory such that transactions that need a snapshot can simply copy it. I've since noted that in Hot Standby mode, that's more or less what the KnownAssignedXids stuff already does. I objected that, first, the overhead of updating the snapshot for every commit would be too great, and second, it didn't seem to do a whole lot to reduce the size of the critical section, and therefore probably wouldn't improve performance that much. But I'm coming around to the view that these might be solvable problems rather than reasons to give up on the idea altogether. With respect to the first problem, what I'm imagining is that we not do a complete rewrite of the snapshot in shared memory on every commit. Instead, when a transaction ends, we'll decide whether to (a) write a new snapshot or (b) just record the XIDs that ended. If we do (b), then any backend that wants a snapshot will need to copy from shared memory both the most recently written snapshot and the XIDs that have subsequently ended. From there, it can figure out which XIDs are still running. Of course, if the list of recently-ended XIDs gets too long, then taking a snapshot will start to get expensive, so we'll need to periodically do (a) instead. There are other ways that this could be done as well; for example, the KnownAssignedXids stuff just flags XIDs that should be ignored and then periodically compacts away the ignored entries. I think the real trick is figuring out a design that can improve concurrency. If you keep a snapshot in shared memory and periodically overwrite it in place, I don't think you're going to gain much. Everyone who wants a snapshot still needs a share-lock and everyone who wants to commit still needs an exclusive-lock, and while you might be able to make the critical section a bit shorter, I think it's still going to be hard to make big gains that way. What I'm thinking about instead is using a ring buffer with three pointers: a start pointer, a stop pointer, and a write pointer. When a transaction ends, we advance the write pointer, write the XIDs or a whole new snapshot into the buffer, and then advance the stop pointer. If we wrote a whole new snapshot, we advance the start pointer to the beginning of the data we just wrote. Someone who wants to take a snapshot must read the data between the start and stop pointers, and must then check that the write pointer hasn't advanced so far in the meantime that the data they read might have been overwritten before they finished reading it. Obviously,
Re: [HACKERS] cheaper snapshots redux
On Aug 22, 2011, at 4:25 PM, Robert Haas wrote: What I'm thinking about instead is using a ring buffer with three pointers: a start pointer, a stop pointer, and a write pointer. When a transaction ends, we advance the write pointer, write the XIDs or a whole new snapshot into the buffer, and then advance the stop pointer. If we wrote a whole new snapshot, we advance the start pointer to the beginning of the data we just wrote. Someone who wants to take a snapshot must read the data between the start and stop pointers, and must then check that the write pointer hasn't advanced so far in the meantime that the data they read might have been overwritten before they finished reading it. Obviously, that's a little risky, since we'll have to do the whole thing over if a wraparound occurs, but if the ring buffer is large enough it shouldn't happen very often. And a typical snapshot is pretty small unless massive numbers of subxids are in use, so it seems like it might not be too bad. Of course, it's pretty hard to know for sure without coding it up and testing it. Something that would be really nice to fix is our reliance on a fixed size of shared memory, and I'm wondering if this could be an opportunity to start in a new direction. My thought is that we could maintain two distinct shared memory snapshots and alternate between them. That would allow us to actually resize them as needed. We would still need something like what you suggest to allow for adding to the list without locking, but with this scheme we wouldn't need to worry about extra locking when taking a snapshot since we'd be doing that in a new segment that no one is using yet. The downside is such a scheme does add non-trivial complexity on top of what you proposed. I suspect it would be much better if we had a separate mechanism for dealing with shared memory requirements (shalloc?). But if it's just not practical to make a generic shared memory manager it would be good to start thinking about ways we can work around fixed shared memory size issues. -- Jim C. Nasby, Database Architect j...@nasby.net 512.569.9461 (cell) http://jim.nasby.net -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots redux
On Mon, Aug 22, 2011 at 6:45 PM, Jim Nasby j...@nasby.net wrote: Something that would be really nice to fix is our reliance on a fixed size of shared memory, and I'm wondering if this could be an opportunity to start in a new direction. My thought is that we could maintain two distinct shared memory snapshots and alternate between them. That would allow us to actually resize them as needed. We would still need something like what you suggest to allow for adding to the list without locking, but with this scheme we wouldn't need to worry about extra locking when taking a snapshot since we'd be doing that in a new segment that no one is using yet. The downside is such a scheme does add non-trivial complexity on top of what you proposed. I suspect it would be much better if we had a separate mechanism for dealing with shared memory requirements (shalloc?). But if it's just not practical to make a generic shared memory manager it would be good to start thinking about ways we can work around fixed shared memory size issues. Well, the system I'm proposing is actually BETTER than having two distinct shared memory snapshots. For example, right now we cache up to 64 subxids per backend. I'm imagining that going away and using that memory for the ring buffer. Out of the box, that would imply a ring buffer of 64 * 103 = 6592 slots. If the average snapshot lists 100 XIDs, you could rewrite the snapshot dozens of times times before the buffer wraps around, which is obviously a lot more than two. Even if subtransactions are being heavily used and each snapshot lists 1000 XIDs, you still have enough space to rewrite the snapshot several times over before wraparound occurs. Of course, at some point the snapshot gets too big and you have to switch to retaining only the toplevel XIDs, which is more or less the equivalent of what happens under the current implementation when any single transaction's subxid cache overflows. With respect to a general-purpose shared memory allocator, I think that there are cases where that would be useful to have, but I don't think there are as many of them as many people seem to think. I wouldn't choose to implement this using a general-purpose allocator even if we had it, both because it's undesirable to allow this or any subsystem to consume an arbitrary amount of memory (nor can it fail... especially in the abort path) and because a ring buffer is almost certainly faster than a general-purpose allocator. We have enough trouble with palloc overhead already. That having been said, I do think there are cases where it would be nice to have... and it wouldn't surprise me if I end up working on something along those lines in the next year or so. It turns out that memory management is a major issue in lock-free programming; you can't assume that it's safe to recycle an object once the last pointer to it has been removed from shared memory - because someone may have fetched the pointer just before you removed it and still be using it to examine the object. An allocator with some built-in capabilities for handling such problems seems like it might be very useful -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 8:32 PM, Hannu Krosing ha...@2ndquadrant.com wrote: Maybe this is why other databases don't offer per backend async commit ? Oracle has async commit but very few people know about it. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Robert Haas robertmh...@gmail.com wrote: (4) We communicate acceptable snapshots to the replica to make the order of visibility visibility match the master even when that doesn't match the order that transactions returned from commit. I (predictably) like (4) -- even though it's a lot of work I think that (4), beyond being a lot of work, will also have pretty terrible performance. You're basically talking about emitting two WAL records for every commit instead of one. Well, I can think of a great many other ways this could be done, each with its own trade-offs of various types of overhead against how close the replica is to current. At one extreme you could do what you describe, at the other you could generate a new snapshot on the replica once every few minutes. Then there are more clever ways, in discussions a few months ago I suggested that adding two new bit flags to the commit record would suffice, and I don't remember anyone blowing holes in that idea. Of course, that was to achieve serializable behavior on the replica, based on some assumption that the current hot standby already supported repeatable read. We might need another bit or two to solve the problems with that which have surfaced on this thread. -Kevin -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 20:14 -0400, Robert Haas wrote: On Thu, Jul 28, 2011 at 7:54 PM, Ants Aasma ants.aa...@eesti.ee wrote: On Thu, Jul 28, 2011 at 11:54 PM, Kevin Grittner kevin.gritt...@wicourts.gov wrote: (4) We communicate acceptable snapshots to the replica to make the order of visibility visibility match the master even when that doesn't match the order that transactions returned from commit. I wonder if some interpretation of 2 phase commit could make Robert's original suggestion implement this. On the master the commit sequence would look something like: 1. Insert commit record to the WAL 2. Wait for replication 3. Get a commit seq nr and mark XIDs visible 4. WAL log the seq nr 5. Return success to client When replaying: * When replaying commit record, do everything but make the tx visible. * When replaying the commit sequence number if there is a gap between last visible commit and current: insert the commit sequence nr. to list of waiting commits. else: mark current and all directly following waiting tx's visible This would give consistent visibility order on master and slave. Robert is right that this would undesirably increase WAL traffic. Delaying this traffic would undesirably increase replay lag between master and slave. But it seems to me that this could be an optional WAL level on top of hot_standby that would only be enabled if consistent visibility on slaves is desired. I think you nailed it. Agreed, this would keep current semantics on master and same visibility order on master and slave. An additional point to think about: if we were willing to insist on streaming replication, we could send the commit sequence numbers via a side channel rather than writing them to WAL, which would be a lot cheaper. Why do you think that side channel is cheaper than main WAL ? How would you handle synchronising the two ? That might even be a reasonable thing to do, because if you're doing log shipping, this is all going to be super-not-real-time anyway. But perhaps you still may want to preserve visibility order to be able to do PITR to exact transaction commit, no ? OTOH, I know we don't want to make WAL shipping anything less than a first class citizen, so maybe not. At any rate, we may be getting a little sidetracked here from the original point of the thread, which was how to make snapshot-taking cheaper. Maybe there's some tie-in to when transactions become visible, but I think it's pretty weak. The existing system could be hacked up to avoid making transactions visible out of LSN order, and the system I proposed could make them visible either in LSN order or do the same thing we do now. They are basically independent problems, AFAICS. Agreed. -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Fri, Jul 29, 2011 at 10:20 AM, Hannu Krosing ha...@2ndquadrant.com wrote: An additional point to think about: if we were willing to insist on streaming replication, we could send the commit sequence numbers via a side channel rather than writing them to WAL, which would be a lot cheaper. Why do you think that side channel is cheaper than main WAL ? You don't have to flush it to disk, and you can use some other lock that isn't as highly contended as WALInsertLock to synchronize it. That might even be a reasonable thing to do, because if you're doing log shipping, this is all going to be super-not-real-time anyway. But perhaps you still may want to preserve visibility order to be able to do PITR to exact transaction commit, no ? Maybe. In practice, I suspect most people won't be willing to pay the price a feature like this would exact. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Fri, 2011-07-29 at 10:23 -0400, Robert Haas wrote: On Fri, Jul 29, 2011 at 10:20 AM, Hannu Krosing ha...@2ndquadrant.com wrote: An additional point to think about: if we were willing to insist on streaming replication, we could send the commit sequence numbers via a side channel rather than writing them to WAL, which would be a lot cheaper. Why do you think that side channel is cheaper than main WAL ? You don't have to flush it to disk, You can probably write the i became visible WAL record without forcing a flush and still get the same visibility order. and you can use some other lock that isn't as highly contended as WALInsertLock to synchronize it. but you will need to synchronise it with WAL replay on slave anyway. It seems easiest to just insert it in the WAL stream and be done with it. That might even be a reasonable thing to do, because if you're doing log shipping, this is all going to be super-not-real-time anyway. But perhaps you still may want to preserve visibility order to be able to do PITR to exact transaction commit, no ? Maybe. In practice, I suspect most people won't be willing to pay the price a feature like this would exact. Unless we find some really bad problems with different visibility orders on master and slave(s) you are probably right. -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 3:51 AM, Robert Haas robertmh...@gmail.com wrote: All that having been said, even if I haven't made any severe conceptual errors in the above, I'm not sure how well it will work in practice. On the plus side, taking a snapshot becomes O(1) rather than O(MaxBackends) - that's good. On the further plus side, you can check both whether an XID has committed and whether it's visible to your snapshot in a single, atomic action with no lock - that seems really good. On the minus side, checking an xid against your snapshot now has less locality of reference. And, rolling over into a new segment of the array is going to require everyone to map it, and maybe cause some disk I/O as a new file gets created. Sounds like the right set of thoughts to be having. If you do this, you must cover subtransactions and Hot Standby. Work in this area takes longer than you think when you take the complexities into account, as you must. I think you should take the premise of making snapshots O(1) and look at all the ways of doing that. If you grab too early at a solution you may grab the wrong one. For example, another approach would be to use a shared hash table. Snapshots are O(1), committing is O(k), using the snapshot is O(logN). N can be kept small by regularly pruning the hash table. If we crash we lose the hash table - no matter. (I'm not suggesting this is better, just a different approach that should be judged across others). What I'm not sure in any of these ideas is how to derive a snapshot xmin. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Jul28, 2011, at 04:51 , Robert Haas wrote: One fly in the ointment is that 8-byte stores are apparently done as two 4-byte stores on some platforms. But if the counter runs backward, I think even that is OK. If you happen to read an 8 byte value as it's being written, you'll get 4 bytes of the intended value and 4 bytes of zeros. The value will therefore appear to be less than what it should be. However, if the value was in the midst of being written, then it's still in the midst of committing, which means that that XID wasn't going to be visible anyway. Accidentally reading a smaller value doesn't change the answer. That only works if the update of the most-significant word is guaranteed to be visible before the update to the lest-significant one. Which I think you can only enforce if you update the words individually (and use a fence on e.g. PPC32). Otherwise you're at the mercy of the compiler. Otherwise, the following might happen (with a 2-byte value instead of an 8-byte one, and the assumption that 1-byte stores are atomic while 2-bytes ones aren't. Just to keep the numbers smaller. The machine is assumed to be big-endian) The counter is at 0xff00 Backends 1 decrements, i.e. does (1) STORE [counter+1] 0xff (2) STORE [counter], 0x00 Backend 2 reads (1') LOAD [counter+1] (2') LOAD [counter] If the sequence of events is (1), (1'), (2'), (2), backend 2 will read 0x which is higher than it should be. But we could simply use a spin-lock to protect the read on machines where we don't know for sure that 64-bit reads and write are atomic. That'll only really hurt on machines with 16+ cores or so, and the number of architectures which support that isn't that high anyway. If we supported spinlock-less operation on SPARC, x86-64, PPC64 and maybe Itanium, would we miss any important one? best regards, Florian Pflug -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Wed, Oct 20, 2010 at 10:07 PM, Tom Lane t...@sss.pgh.pa.us wrote: I wonder whether we could do something involving WAL properties --- the current tuple visibility logic was designed before WAL existed, so it's not exploiting that resource at all. I'm imagining that the kernel of a snapshot is just a WAL position, ie the end of WAL as of the time you take the snapshot (easy to get in O(1) time). Visibility tests then reduce to did this transaction commit with a WAL record located before the specified position?. Why not just cache a reference snapshots near WAL writer and maybe also save it at some interval in WAL in case you ever need to restore an old snapshot at some val position for things like time travel. It may be cheaper lock-wise not to update ref. snapshot at each commit, but to keep latest saved snapshot and a chain of transactions committed / aborted since. This means that when reading the snapshot you read the current saved snapshot and then apply the list of commits. when moving to a new saved snapshot you really generate a new one and keep the old snapshot + commit chain around for a little while for those who may be still processing it. Seems like this is something that can be done with no locking, You'd need some index datastructure that made it reasonably cheap to find out the commit locations of recently committed transactions, where recent means back to recentGlobalXmin. That seems possibly do-able, though I don't have a concrete design in mind. snapshot + chain of commits is likely as cheap as it gets, unless you additionally cache the commits in a tighter data structure. this is because you will need them all anyway to compute difference from ref snapshot. -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 3:46 AM, Simon Riggs si...@2ndquadrant.com wrote: Sounds like the right set of thoughts to be having. Thanks. If you do this, you must cover subtransactions and Hot Standby. Work in this area takes longer than you think when you take the complexities into account, as you must. Right. This would replace the KnownAssignedXids stuff (a non-trivial project, I am sure). I think you should take the premise of making snapshots O(1) and look at all the ways of doing that. If you grab too early at a solution you may grab the wrong one. Yeah, I'm just brainstorming at this point. This is, I think, the best idea of what I've come up with so far, but it's definitely not the only approach. For example, another approach would be to use a shared hash table. Snapshots are O(1), committing is O(k), using the snapshot is O(logN). N can be kept small by regularly pruning the hash table. If we crash we lose the hash table - no matter. (I'm not suggesting this is better, just a different approach that should be judged across others). Sorry, I'm having a hard time understanding what you are describing here. What would the keys and values in this hash table be, and what do k and N refer to here? What I'm not sure in any of these ideas is how to derive a snapshot xmin. That is a problem. If we have to scan the ProcArray every time we take a snapshot just to derive an xmin, we are kind of hosed. One thought I had is that we might be able to use a sort of sloppy xmin. In other words, we keep a cached xmin, and have some heuristic where we occasionally try to update it. A snapshot with a too-old xmin isn't wrong, just possibly slower. But if xmin is only slightly stale and xids can be tested relatively quickly, it might not matter very much. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 4:16 AM, Florian Pflug f...@phlo.org wrote: On Jul28, 2011, at 04:51 , Robert Haas wrote: One fly in the ointment is that 8-byte stores are apparently done as two 4-byte stores on some platforms. But if the counter runs backward, I think even that is OK. If you happen to read an 8 byte value as it's being written, you'll get 4 bytes of the intended value and 4 bytes of zeros. The value will therefore appear to be less than what it should be. However, if the value was in the midst of being written, then it's still in the midst of committing, which means that that XID wasn't going to be visible anyway. Accidentally reading a smaller value doesn't change the answer. That only works if the update of the most-significant word is guaranteed to be visible before the update to the lest-significant one. Which I think you can only enforce if you update the words individually (and use a fence on e.g. PPC32). Otherwise you're at the mercy of the compiler. Otherwise, the following might happen (with a 2-byte value instead of an 8-byte one, and the assumption that 1-byte stores are atomic while 2-bytes ones aren't. Just to keep the numbers smaller. The machine is assumed to be big-endian) The counter is at 0xff00 Backends 1 decrements, i.e. does (1) STORE [counter+1] 0xff (2) STORE [counter], 0x00 Backend 2 reads (1') LOAD [counter+1] (2') LOAD [counter] If the sequence of events is (1), (1'), (2'), (2), backend 2 will read 0x which is higher than it should be. You're confusing two different things - I agree that you need a spinlock around reading the counter, unless 8-byte loads and stores are atomic. What I'm saying can be done without a lock is reading the commit-order value for a given XID. If that's in the middle of being updated, then the old value was zero, so the scenario you describe can't occur. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 6:50 AM, Hannu Krosing ha...@2ndquadrant.com wrote: On Wed, Oct 20, 2010 at 10:07 PM, Tom Lane t...@sss.pgh.pa.us wrote: I wonder whether we could do something involving WAL properties --- the current tuple visibility logic was designed before WAL existed, so it's not exploiting that resource at all. I'm imagining that the kernel of a snapshot is just a WAL position, ie the end of WAL as of the time you take the snapshot (easy to get in O(1) time). Visibility tests then reduce to did this transaction commit with a WAL record located before the specified position?. Why not just cache a reference snapshots near WAL writer and maybe also save it at some interval in WAL in case you ever need to restore an old snapshot at some val position for things like time travel. It may be cheaper lock-wise not to update ref. snapshot at each commit, but to keep latest saved snapshot and a chain of transactions committed / aborted since. This means that when reading the snapshot you read the current saved snapshot and then apply the list of commits. Yeah, interesting idea. I thought about that. You'd need not only the list of commits but also the list of XIDs that had been published, since the commits have to be removed from the snapshot and the newly-published XIDs have to be added to it (in case they commit later while the snapshot is still in use). You can imagine doing this with a pair of buffers. You write a snapshot into the beginning of the first buffer and then write each XID that is published or commits into the next slot in the array. When the buffer is filled up, the next process that wants to publish an XID or commit scans through the array and constructs a new snapshot that compacts away all the begin/commit pairs and writes it into the second buffer, and all new snapshots are taken there. When that buffer fills up you flip back to the first one. Of course, you need some kind of synchronization to make sure that you don't flip back to the first buffer while some laggard is still using it to construct a snapshot that he started taking before you flipped to the second one, but maybe that could be made light-weight enough not to matter. I am somewhat concerned that this approach might lead to a lot of contention over the snapshot buffers. In particular, the fact that you have to touch shared cache lines both to advertise a new XID and when it gets committed seems less than ideal. One thing that's kind of interesting about the commit sequence number approach is that - as far as I can tell - it doesn't require new XIDs to be advertised anywhere at all. You don't have to worry about overflowing the subxids[] array because it goes away altogether. The commit sequence number itself is going to be a contention hotspot, but at least it's small and fixed-size. Another concern I have with this approach is - how large do you make the buffers? If you make them too small, then you're going to have to regenerate the snapshot frequently, which will lead to the same sort of lock contention we have today - no one can commit while the snapshot is being regenerated. On the other hand, if you make them too big, then deriving a snapshot gets slow. Maybe there's some way to make it work, but I'm afraid it might end up being yet another arcane thing the tuning of which will become a black art among hackers... -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 09:38 -0400, Robert Haas wrote: On Thu, Jul 28, 2011 at 6:50 AM, Hannu Krosing ha...@2ndquadrant.com wrote: On Wed, Oct 20, 2010 at 10:07 PM, Tom Lane t...@sss.pgh.pa.us wrote: I wonder whether we could do something involving WAL properties --- the current tuple visibility logic was designed before WAL existed, so it's not exploiting that resource at all. I'm imagining that the kernel of a snapshot is just a WAL position, ie the end of WAL as of the time you take the snapshot (easy to get in O(1) time). Visibility tests then reduce to did this transaction commit with a WAL record located before the specified position?. Why not just cache a reference snapshots near WAL writer and maybe also save it at some interval in WAL in case you ever need to restore an old snapshot at some val position for things like time travel. It may be cheaper lock-wise not to update ref. snapshot at each commit, but to keep latest saved snapshot and a chain of transactions committed / aborted since. This means that when reading the snapshot you read the current saved snapshot and then apply the list of commits. Yeah, interesting idea. I thought about that. You'd need not only the list of commits but also the list of XIDs that had been published, since the commits have to be removed from the snapshot and the newly-published XIDs have to be added to it (in case they commit later while the snapshot is still in use). You can imagine doing this with a pair of buffers. You write a snapshot into the beginning of the first buffer and then write each XID that is published or commits into the next slot in the array. When the buffer is filled up, the next process that wants to publish an XID or commit scans through the array and constructs a new snapshot that compacts away all the begin/commit pairs and writes it into the second buffer, and all new snapshots are taken there. When that buffer fills up you flip back to the first one. Of course, you need some kind of synchronization to make sure that you don't flip back to the first buffer while some laggard is still using it to construct a snapshot that he started taking before you flipped to the second one, but maybe that could be made light-weight enough not to matter. I am somewhat concerned that this approach might lead to a lot of contention over the snapshot buffers. My hope was, that this contention would be the same than simply writing the WAL buffers currently, and thus largely hidden by the current WAL writing sync mechanisma. It really covers just the part which writes commit records to WAL, as non-commit WAL records dont participate in snapshot updates. Writing WAL is already a single point which needs locks or other kind of synchronization. This will stay with us at least until we start supporting multiple WAL streams, and even then we will need some synchronisation between those. In particular, the fact that you have to touch shared cache lines both to advertise a new XID and when it gets committed seems less than ideal. Every commit record writer should do this as part of writing the commit record. And as mostly you still want the latest snapshot, why not just update the snapshot as part of the commit/abort ? Do we need the ability for fast recent snapshots at all ? One thing that's kind of interesting about the commit sequence number approach is that - as far as I can tell - it doesn't require new XIDs to be advertised anywhere at all. You don't have to worry about overflowing the subxids[] array because it goes away altogether. The commit sequence number itself is going to be a contention hotspot, but at least it's small and fixed-size. Another concern I have with this approach is - how large do you make the buffers? If you make them too small, then you're going to have to regenerate the snapshot frequently, which will lead to the same sort of lock contention we have today - no one can commit while the snapshot is being regenerated. On the other hand, if you make them too big, then deriving a snapshot gets slow. Maybe there's some way to make it work, but I'm afraid it might end up being yet another arcane thing the tuning of which will become a black art among hackers... -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 10:17 AM, Hannu Krosing ha...@2ndquadrant.com wrote: My hope was, that this contention would be the same than simply writing the WAL buffers currently, and thus largely hidden by the current WAL writing sync mechanisma. It really covers just the part which writes commit records to WAL, as non-commit WAL records dont participate in snapshot updates. I'm confused by this, because I don't think any of this can be done when we insert the commit record into the WAL stream. It has to be done later, at the time we currently remove ourselves from the ProcArray. Those things need not happen in the same order, as I noted in my original post. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Robert Haas robertmh...@gmail.com writes: On Thu, Jul 28, 2011 at 10:17 AM, Hannu Krosing ha...@2ndquadrant.com wrote: My hope was, that this contention would be the same than simply writing the WAL buffers currently, and thus largely hidden by the current WAL writing sync mechanisma. It really covers just the part which writes commit records to WAL, as non-commit WAL records dont participate in snapshot updates. I'm confused by this, because I don't think any of this can be done when we insert the commit record into the WAL stream. It has to be done later, at the time we currently remove ourselves from the ProcArray. Those things need not happen in the same order, as I noted in my original post. But should we rethink that? Your point that hot standby transactions on a slave could see snapshots that were impossible on the parent was disturbing. Should we look for a way to tie transaction becomes visible to its creation of a commit WAL record? I think the fact that they are not an indivisible operation is an implementation artifact, and not a particularly nice one. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 10:23 -0400, Robert Haas wrote: On Thu, Jul 28, 2011 at 10:17 AM, Hannu Krosing ha...@2ndquadrant.com wrote: My hope was, that this contention would be the same than simply writing the WAL buffers currently, and thus largely hidden by the current WAL writing sync mechanisma. It really covers just the part which writes commit records to WAL, as non-commit WAL records dont participate in snapshot updates. I'm confused by this, because I don't think any of this can be done when we insert the commit record into the WAL stream. It has to be done later, at the time we currently remove ourselves from the ProcArray. Those things need not happen in the same order, as I noted in my original post. The update to stored snapshot needs to happen at the moment when the WAL record is considered to be on stable storage, so the current snapshot update presumably can be done by the same process which forces it to stable storage, with the same contention pattern that applies to writing WAL records, no ? If the problem is with a backend which requested an async commit, then it is free to apply it's additional local commit changes from its own memory if the global latest snapshot disgrees with it. -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Hannu Krosing ha...@2ndquadrant.com writes: On Thu, 2011-07-28 at 10:23 -0400, Robert Haas wrote: I'm confused by this, because I don't think any of this can be done when we insert the commit record into the WAL stream. The update to stored snapshot needs to happen at the moment when the WAL record is considered to be on stable storage, so the current snapshot update presumably can be done by the same process which forces it to stable storage, with the same contention pattern that applies to writing WAL records, no ? No. There is no reason to tie this to fsyncing WAL. For purposes of other currently-running transactions, the commit can be considered to occur at the instant the commit record is inserted into WAL buffers. If we crash before that makes it to disk, no problem, because nothing those other transactions did will have made it to disk either. The advantage of defining it that way is you don't have weirdly different behaviors for sync and async transactions. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 10:33 AM, Tom Lane t...@sss.pgh.pa.us wrote: Robert Haas robertmh...@gmail.com writes: On Thu, Jul 28, 2011 at 10:17 AM, Hannu Krosing ha...@2ndquadrant.com wrote: My hope was, that this contention would be the same than simply writing the WAL buffers currently, and thus largely hidden by the current WAL writing sync mechanisma. It really covers just the part which writes commit records to WAL, as non-commit WAL records dont participate in snapshot updates. I'm confused by this, because I don't think any of this can be done when we insert the commit record into the WAL stream. It has to be done later, at the time we currently remove ourselves from the ProcArray. Those things need not happen in the same order, as I noted in my original post. But should we rethink that? Your point that hot standby transactions on a slave could see snapshots that were impossible on the parent was disturbing. Should we look for a way to tie transaction becomes visible to its creation of a commit WAL record? I think the fact that they are not an indivisible operation is an implementation artifact, and not a particularly nice one. Well, I agree with you that it isn't especially nice, but it seems like a fairly intractable problem. Currently, the standby has no way of knowing in what order the transactions became visible on the master. Unless we want to allow only SR and not log shipping, the only way to communicate that information would be to WAL log it. Aside from the expense, what do we do if XLogInsert() fails, given that we've already committed? -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 10:45 -0400, Tom Lane wrote: Hannu Krosing ha...@2ndquadrant.com writes: On Thu, 2011-07-28 at 10:23 -0400, Robert Haas wrote: I'm confused by this, because I don't think any of this can be done when we insert the commit record into the WAL stream. The update to stored snapshot needs to happen at the moment when the WAL record is considered to be on stable storage, so the current snapshot update presumably can be done by the same process which forces it to stable storage, with the same contention pattern that applies to writing WAL records, no ? No. There is no reason to tie this to fsyncing WAL. For purposes of other currently-running transactions, the commit can be considered to occur at the instant the commit record is inserted into WAL buffers. If we crash before that makes it to disk, no problem, because nothing those other transactions did will have made it to disk either. Agreed. Actually figured it out right after pushing send :) The advantage of defining it that way is you don't have weirdly different behaviors for sync and async transactions. My main point was, that we already do synchronization when writing wal, why not piggyback on this to also update latest snapshot . -- --- Hannu Krosing PostgreSQL (Infinite) Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 11:10 AM, Hannu Krosing ha...@2ndquadrant.com wrote: My main point was, that we already do synchronization when writing wal, why not piggyback on this to also update latest snapshot . Well, one problem is that it would break sync rep. Another problem is that pretty much the last thing I want to do is push more work under WALInsertLock. Based on the testing I've done so far, it seems like WALInsertLock, ProcArrayLock, and CLogControlLock are the main bottlenecks here. I'm focusing on ProcArrayLock and CLogControlLock right now, but I am pretty well convinced that WALInsertLock is going to be the hardest nut to crack, so putting anything more under there seems like it's going in the wrong direction. IMHO, anyway. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 17:10 +0200, Hannu Krosing wrote: On Thu, 2011-07-28 at 10:45 -0400, Tom Lane wrote: Hannu Krosing ha...@2ndquadrant.com writes: On Thu, 2011-07-28 at 10:23 -0400, Robert Haas wrote: I'm confused by this, because I don't think any of this can be done when we insert the commit record into the WAL stream. The update to stored snapshot needs to happen at the moment when the WAL record is considered to be on stable storage, so the current snapshot update presumably can be done by the same process which forces it to stable storage, with the same contention pattern that applies to writing WAL records, no ? No. There is no reason to tie this to fsyncing WAL. For purposes of other currently-running transactions, the commit can be considered to occur at the instant the commit record is inserted into WAL buffers. If we crash before that makes it to disk, no problem, because nothing those other transactions did will have made it to disk either. Agreed. Actually figured it out right after pushing send :) The advantage of defining it that way is you don't have weirdly different behaviors for sync and async transactions. My main point was, that we already do synchronization when writing wal, why not piggyback on this to also update latest snapshot . So the basic design could be a sparse snapshot, consisting of 'xmin, xmax, running_txids[numbackends] where each backend manages its own slot in running_txids - sets a txid when aquiring one and nulls it at commit, possibly advancing xmin if xmin==mytxid. as xmin update requires full scan of running_txids, it is also a good time to update xmax - no need to advance xmax when inserting your next txid, so you don't need to locak anything at insert time. the valid xmax is still computed when getting the snapshot. hmm, probably no need to store xmin and xmax at all. it needs some further analysis to figure out, if doing it this way without any locks can produce any relevantly bad snapshots. maybe you still need one spinlock + memcpy of running_txids to local memory to get snapshot. also, as the running_txids array is global, it may need to be made even sparser to minimise cache-line collisions. needs to be a tuning decision between cache conflicts and speed of memcpy. -- --- Hannu Krosing PostgreSQL (Infinite) Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 11:15 -0400, Robert Haas wrote: On Thu, Jul 28, 2011 at 11:10 AM, Hannu Krosing ha...@2ndquadrant.com wrote: My main point was, that we already do synchronization when writing wal, why not piggyback on this to also update latest snapshot . Well, one problem is that it would break sync rep. Can you elaborate, in what way it breaks sync rep ? Another problem is that pretty much the last thing I want to do is push more work under WALInsertLock. Based on the testing I've done so far, it seems like WALInsertLock, ProcArrayLock, and CLogControlLock are the main bottlenecks here. I'm focusing on ProcArrayLock and CLogControlLock right now, but I am pretty well convinced that WALInsertLock is going to be the hardest nut to crack, so putting anything more under there seems like it's going in the wrong direction. probably it is not just the WALInsertLock, but the fact that we have just one WAL. It can become a bottleneck once we have significant number of processors fighting to write in single WAL. IMHO, anyway. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Robert Haas robertmh...@gmail.com writes: On Thu, Jul 28, 2011 at 10:33 AM, Tom Lane t...@sss.pgh.pa.us wrote: But should we rethink that? Your point that hot standby transactions on a slave could see snapshots that were impossible on the parent was disturbing. Should we look for a way to tie transaction becomes visible to its creation of a commit WAL record? I think the fact that they are not an indivisible operation is an implementation artifact, and not a particularly nice one. Well, I agree with you that it isn't especially nice, but it seems like a fairly intractable problem. Currently, the standby has no way of knowing in what order the transactions became visible on the master. Right, but if the visibility order were *defined* as the order in which commit records appear in WAL, that problem neatly goes away. It's only because we have the implementation artifact that set my xid to 0 in the ProcArray is decoupled from inserting the commit record that there's any difference. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Wed, 2011-07-27 at 22:51 -0400, Robert Haas wrote: On Wed, Oct 20, 2010 at 10:07 PM, Tom Lane t...@sss.pgh.pa.us wrote: I wonder whether we could do something involving WAL properties --- the current tuple visibility logic was designed before WAL existed, so it's not exploiting that resource at all. I'm imagining that the kernel of a snapshot is just a WAL position, ie the end of WAL as of the time you take the snapshot (easy to get in O(1) time). Visibility tests then reduce to did this transaction commit with a WAL record located before the specified position?. You'd need some index datastructure that made it reasonably cheap to find out the commit locations of recently committed transactions, where recent means back to recentGlobalXmin. That seems possibly do-able, though I don't have a concrete design in mind. I was mulling this idea over some more (the same ideas keep floating back to the top...). I don't think an LSN can actually work, because there's no guarantee that the order in which the WAL records are emitted is the same order in which the effects of the transactions become visible to new snapshots. For example: 1. Transaction A inserts its commit record, flushes WAL, and begins waiting for sync rep. 2. A moment later, transaction B sets synchronous_commit=off, inserts its commit record, requests a background WAL flush, and removes itself from the ProcArray. 3. Transaction C takes a snapshot. It is Transaction A here which is acting badly - it should also remove itself from procArray right after it inserts its commit record, as for everybody else except the client app of transaction A it is committed at this point. It just cant report back to client before getting confirmation that it is actually syncrepped (or locally written to stable storage). At least at the point of consistent snapshots the right sequence should be: 1) inert commit record into wal 2) remove yourself from ProcArray (or use some other means to declare that your transaction is no longer running) 3) if so configured, wait for WAL flus to stable storage and/or SYnc Rep confirmation Based on this let me suggest a simple snapshot cache mechanism A simple snapshot cache mechanism = have an array of running transactions, with one slot per backend txid running_transactions[max_connections]; there are exactly 3 operations on this array 1. insert backends running transaction id - this is done at the moment of acquiring your transaction id from system, and synchronized by the same mechanism as getting the transaction id running_transactions[my_backend] = current_transaction_id 2. remove backends running transaction id - this is done at the moment of committing or aborting the transaction, again synchronized by the write commit record mechanism. running_transactions[my_backend] = NULL should be first thing after insertin WAcommit record 3. getting a snapshot - memcpy() running_transactions to local memory, then construct a snapshot it may be that you need to protect all3 operations with a single spinlock, if so then I'd propose the same spinlock used when getting your transaction id (and placing the array near where latest transaction id is stored so they share cache line). But it is also possible, that you can get logically consistent snapshots by protecting only some ops. for example, if you protect only insert and get snapshot, then the worst that can happen is that you get a snapshot that is a few commits older than what youd get with full locking and it may well be ok for all real uses. -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 11:57 -0400, Tom Lane wrote: Robert Haas robertmh...@gmail.com writes: On Thu, Jul 28, 2011 at 10:33 AM, Tom Lane t...@sss.pgh.pa.us wrote: But should we rethink that? Your point that hot standby transactions on a slave could see snapshots that were impossible on the parent was disturbing. Should we look for a way to tie transaction becomes visible to its creation of a commit WAL record? I think the fact that they are not an indivisible operation is an implementation artifact, and not a particularly nice one. Well, I agree with you that it isn't especially nice, but it seems like a fairly intractable problem. Currently, the standby has no way of knowing in what order the transactions became visible on the master. Right, but if the visibility order were *defined* as the order in which commit records appear in WAL, that problem neatly goes away. It's only because we have the implementation artifact that set my xid to 0 in the ProcArray is decoupled from inserting the commit record that there's any difference. Yes, as I explain in another e-mail, the _only_ one for whom the transaction is not yet committed is the waiting backend itself. for all others it should show as committed the moment after the wal record is written. It's kind of local 2 phase commit thing :) regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 18:05 +0200, Hannu Krosing wrote: But it is also possible, that you can get logically consistent snapshots by protecting only some ops. for example, if you protect only insert and get snapshot, then the worst that can happen is that you get a snapshot that is a few commits older than what youd get with full locking and it may well be ok for all real uses. Thinking more of it, we should lock commit/remove_txid and get_snapshot having a few more running backends does not make a difference, but seeing commits in wrong order may. this will cause contention between commit and get_snapshot, but hopefully less than current ProcArray manipulation, as there is just one simple C array to lock and copy. -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 18:48 +0200, Hannu Krosing wrote: On Thu, 2011-07-28 at 18:05 +0200, Hannu Krosing wrote: But it is also possible, that you can get logically consistent snapshots by protecting only some ops. for example, if you protect only insert and get snapshot, then the worst that can happen is that you get a snapshot that is a few commits older than what youd get with full locking and it may well be ok for all real uses. Thinking more of it, we should lock commit/remove_txid and get_snapshot having a few more running backends does not make a difference, but seeing commits in wrong order may. Sorry, not true as this may advanxe xmax to include some running transactions which were missed during memcpy. so we still need some mechanism to either synchronize the the copy with both inserts and removes, or make it atomic even in presence of multiple CPUs. this will cause contention between commit and get_snapshot, but hopefully less than current ProcArray manipulation, as there is just one simple C array to lock and copy. -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 11:57 AM, Tom Lane t...@sss.pgh.pa.us wrote: Robert Haas robertmh...@gmail.com writes: On Thu, Jul 28, 2011 at 10:33 AM, Tom Lane t...@sss.pgh.pa.us wrote: But should we rethink that? Your point that hot standby transactions on a slave could see snapshots that were impossible on the parent was disturbing. Should we look for a way to tie transaction becomes visible to its creation of a commit WAL record? I think the fact that they are not an indivisible operation is an implementation artifact, and not a particularly nice one. Well, I agree with you that it isn't especially nice, but it seems like a fairly intractable problem. Currently, the standby has no way of knowing in what order the transactions became visible on the master. Right, but if the visibility order were *defined* as the order in which commit records appear in WAL, that problem neatly goes away. It's only because we have the implementation artifact that set my xid to 0 in the ProcArray is decoupled from inserting the commit record that there's any difference. Hmm, interesting idea. However, consider the scenario where some transactions are using synchronous_commit or synchronous replication, and others are not. If a transaction that needs to wait (either just for WAL flush, or for WAL flush and synchronous replication) inserts its commit record, and then another transaction with synchronous_commit=off comes along and inserts its commit record, the second transaction will have to block until the first transaction is done waiting. We can't make either transaction visible without making both visible, and we certainly can't acknowledge the second transaction to the client until we've made it visible. I'm not going to say that's so horrible we shouldn't even consider it, but it doesn't seem great, either. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 11:36 AM, Hannu Krosing ha...@krosing.net wrote: On Thu, 2011-07-28 at 11:15 -0400, Robert Haas wrote: On Thu, Jul 28, 2011 at 11:10 AM, Hannu Krosing ha...@2ndquadrant.com wrote: My main point was, that we already do synchronization when writing wal, why not piggyback on this to also update latest snapshot . Well, one problem is that it would break sync rep. Can you elaborate, in what way it breaks sync rep ? Well, the point of synchronous replication is that the local machine doesn't see the effects of the transaction until it's been replicated. Therefore, no one can be relying on data that might disappear in the event the system is crushed by a falling meteor. It would be easy, technically speaking, to remove the transaction from the ProcArray and *then* wait for synchronous replication, but that would offer a much weaker guarantee than what the current version provides. We would still guarantee that the commit wouldn't be acknowledged to the client which submitted it until it was replicated, but we would no longer be able to guarantee that no one else relied on data written by the transaction prior to successful replication. For example, consider this series of events: 1. User asks ATM what is my balance?. ATM inquires of database, which says $500. 2. User deposits a check for $100. ATM does an UPDATE to add $100 to balance and issues a COMMIT. But the master has become disconnected from the synchronous standby, so the sync rep wait hangs. 3. ATM eventually times out and tells user sorry, i can't complete your transaction right now. 4. User wants to know whether their check got deposited, so they walk into the bank and ask a teller to check their balance. Teller's computer connects to the database and gets $600. User is happy and leaves. 5. Master dies. Failover. 6. User's balance is now back to $500. When the user finds out much later, they say wtf? you told me before it was $600!. Right now, when using synchronous replication, this series of events CANNOT HAPPEN. If some other transaction interrogates the state of the database and sees the results of some transaction, it is an ironclad guarantee that the transaction has been replicated. If we start making transactions visible when their WAL record is flushed or - worse - when it's inserted, then those guarantees go away. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 14:27 -0400, Robert Haas wrote: On Thu, Jul 28, 2011 at 11:57 AM, Tom Lane t...@sss.pgh.pa.us wrote: Robert Haas robertmh...@gmail.com writes: On Thu, Jul 28, 2011 at 10:33 AM, Tom Lane t...@sss.pgh.pa.us wrote: But should we rethink that? Your point that hot standby transactions on a slave could see snapshots that were impossible on the parent was disturbing. Should we look for a way to tie transaction becomes visible to its creation of a commit WAL record? I think the fact that they are not an indivisible operation is an implementation artifact, and not a particularly nice one. Well, I agree with you that it isn't especially nice, but it seems like a fairly intractable problem. Currently, the standby has no way of knowing in what order the transactions became visible on the master. Right, but if the visibility order were *defined* as the order in which commit records appear in WAL, that problem neatly goes away. It's only because we have the implementation artifact that set my xid to 0 in the ProcArray is decoupled from inserting the commit record that there's any difference. Hmm, interesting idea. However, consider the scenario where some transactions are using synchronous_commit or synchronous replication, and others are not. If a transaction that needs to wait (either just for WAL flush, or for WAL flush and synchronous replication) inserts its commit record, and then another transaction with synchronous_commit=off comes along and inserts its commit record, the second transaction will have to block until the first transaction is done waiting. What is the current behavior when the synchronous replication fails (say the slave breaks down) - will the transaction be rolled back at some point or will it wait indefinitely , that is until a new slave is installed ? Or will the sync rep transaction commit when archive_command returns true after copying the WAL segment containing this commit ? We can't make either transaction visible without making both visible, and we certainly can't acknowledge the second transaction to the client until we've made it visible. I'm not going to say that's so horrible we shouldn't even consider it, but it doesn't seem great, either. Maybe this is why other databases don't offer per backend async commit ? -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Hannu Krosing ha...@krosing.net writes: So the basic design could be a sparse snapshot, consisting of 'xmin, xmax, running_txids[numbackends] where each backend manages its own slot in running_txids - sets a txid when aquiring one and nulls it at commit, possibly advancing xmin if xmin==mytxid. How is that different from what we're doing now? Basically, what you're describing is pulling the xids out of the ProcArray and moving them into a separate data structure. That could be a win I guess if non-snapshot- related reasons to take ProcArrayLock represent enough of the contention to be worth separating out, but I suspect they don't. In particular, the data structure you describe above *cannot* be run lock-free, because it doesn't provide any consistency guarantees without a lock. You need everyone to have the same ideas about commit order, and random backends independently changing array elements without locks won't guarantee that. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 21:32 +0200, Hannu Krosing wrote: On Thu, 2011-07-28 at 14:27 -0400, Robert Haas wrote: Hmm, interesting idea. However, consider the scenario where some transactions are using synchronous_commit or synchronous replication, and others are not. If a transaction that needs to wait (either just for WAL flush, or for WAL flush and synchronous replication) inserts its commit record, and then another transaction with synchronous_commit=off comes along and inserts its commit record, the second transaction will have to block until the first transaction is done waiting. What is the current behavior when the synchronous replication fails (say the slave breaks down) - will the transaction be rolled back at some point or will it wait indefinitely , that is until a new slave is installed ? More importantly, if the master crashes after the commit is written to WAL, will the transaction be rolled back after recovery based on the fact that no confirmation from synchronous slave is received ? Or will the sync rep transaction commit when archive_command returns true after copying the WAL segment containing this commit ? We can't make either transaction visible without making both visible, and we certainly can't acknowledge the second transaction to the client until we've made it visible. I'm not going to say that's so horrible we shouldn't even consider it, but it doesn't seem great, either. Maybe this is why other databases don't offer per backend async commit ? -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Hannu Krosing ha...@2ndquadrant.com writes: On Thu, 2011-07-28 at 14:27 -0400, Robert Haas wrote: We can't make either transaction visible without making both visible, and we certainly can't acknowledge the second transaction to the client until we've made it visible. I'm not going to say that's so horrible we shouldn't even consider it, but it doesn't seem great, either. Maybe this is why other databases don't offer per backend async commit ? Yeah, I've always thought that feature wasn't as simple as it appeared. It got in only because it was claimed to be cost-free, and it's now obvious that it isn't. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 15:42 -0400, Tom Lane wrote: Hannu Krosing ha...@2ndquadrant.com writes: On Thu, 2011-07-28 at 14:27 -0400, Robert Haas wrote: We can't make either transaction visible without making both visible, and we certainly can't acknowledge the second transaction to the client until we've made it visible. I'm not going to say that's so horrible we shouldn't even consider it, but it doesn't seem great, either. Maybe this is why other databases don't offer per backend async commit ? Yeah, I've always thought that feature wasn't as simple as it appeared. It got in only because it was claimed to be cost-free, and it's now obvious that it isn't. I still think it is cost-free if you get the semantics of the COMMIT contract right. (Of course it is not cost free as in not wasting developers time in discussions ;) ) I'm still with you in claiming that transaction should be visible to other backends as committed as soon as the WAL record is inserted. the main thing to keep in mind is that getting back positive commit confirmation really means (depending on various sync settings) that your transaction is on stable storage. BUT, _not_ getting back confirmation on commit does not quaranee that it is not committed, just that you need to check. It may well be that it was committed, written to stable storage _and_ also syncrepped but then the confirnation did not come bac to you due to some network outage. or your client computer crashed. or your child spilled black paint over the monitor. or thousand other reasons. async commit has the contract that you are ready to check a few latest commits after crash. but I still think that it is right semantics to make your commit visible to others, even before you have gotten back the confirmation yourself. --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 15:38 -0400, Tom Lane wrote: Hannu Krosing ha...@krosing.net writes: So the basic design could be a sparse snapshot, consisting of 'xmin, xmax, running_txids[numbackends] where each backend manages its own slot in running_txids - sets a txid when aquiring one and nulls it at commit, possibly advancing xmin if xmin==mytxid. How is that different from what we're doing now? Basically, what you're describing is pulling the xids out of the ProcArray and moving them into a separate data structure. That could be a win I guess if non-snapshot- related reasons to take ProcArrayLock represent enough of the contention to be worth separating out, but I suspect they don't. the idea was to make the thid array small enough to be able to memcpy it to backend local memory fast. But I agree it takes testing to see if it is an overall win In particular, the data structure you describe above *cannot* be run lock-free, because it doesn't provide any consistency guarantees without a lock. You need everyone to have the same ideas about commit order, and random backends independently changing array elements without locks won't guarantee that. regards, tom lane -- --- Hannu Krosing PostgreSQL Infinite Scalability and Performance Consultant PG Admin Book: http://www.2ndQuadrant.com/books/ -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Hannu Krosing ha...@2ndquadrant.com wrote: but I still think that it is right semantics to make your commit visible to others, even before you have gotten back the confirmation yourself. Possibly. That combined with building snapshots based on the order of WAL entries of commit records certainly has several appealing aspects. It is hard to get over the fact that you lose an existing guarantee, though: right now, if you have one synchronous replica, you can never see a transaction's work on the master and then *not* see it on the slave -- the slave always has first visibility. I don't see how such a guarantee can exist in *either* direction with the semantics you describe. After seeing a transaction's work on one system it would always be unknown whether it was visible on the other. There are situations where that is OK as long as each copy has a sane order of visibility, but there are situations where losing that guarantee might matter. On the bright side, it means that transactions would become visible on the replica in the same order as on the master, and that blocking would be reduced. -Kevin -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 3:32 PM, Hannu Krosing ha...@2ndquadrant.com wrote: Hmm, interesting idea. However, consider the scenario where some transactions are using synchronous_commit or synchronous replication, and others are not. If a transaction that needs to wait (either just for WAL flush, or for WAL flush and synchronous replication) inserts its commit record, and then another transaction with synchronous_commit=off comes along and inserts its commit record, the second transaction will have to block until the first transaction is done waiting. What is the current behavior when the synchronous replication fails (say the slave breaks down) - will the transaction be rolled back at some point or will it wait indefinitely , that is until a new slave is installed ? It will wait forever, unless you shut down the database or hit ^C. We can't make either transaction visible without making both visible, and we certainly can't acknowledge the second transaction to the client until we've made it visible. I'm not going to say that's so horrible we shouldn't even consider it, but it doesn't seem great, either. Maybe this is why other databases don't offer per backend async commit ? Yeah, possibly. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 3:40 PM, Hannu Krosing ha...@2ndquadrant.com wrote: On Thu, 2011-07-28 at 21:32 +0200, Hannu Krosing wrote: On Thu, 2011-07-28 at 14:27 -0400, Robert Haas wrote: Hmm, interesting idea. However, consider the scenario where some transactions are using synchronous_commit or synchronous replication, and others are not. If a transaction that needs to wait (either just for WAL flush, or for WAL flush and synchronous replication) inserts its commit record, and then another transaction with synchronous_commit=off comes along and inserts its commit record, the second transaction will have to block until the first transaction is done waiting. What is the current behavior when the synchronous replication fails (say the slave breaks down) - will the transaction be rolled back at some point or will it wait indefinitely , that is until a new slave is installed ? More importantly, if the master crashes after the commit is written to WAL, will the transaction be rolled back after recovery based on the fact that no confirmation from synchronous slave is received ? No. You can't roll back a transaction once it's committed - ever. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 4:12 PM, Kevin Grittner kevin.gritt...@wicourts.gov wrote: Hannu Krosing ha...@2ndquadrant.com wrote: but I still think that it is right semantics to make your commit visible to others, even before you have gotten back the confirmation yourself. Possibly. That combined with building snapshots based on the order of WAL entries of commit records certainly has several appealing aspects. It is hard to get over the fact that you lose an existing guarantee, though: right now, if you have one synchronous replica, you can never see a transaction's work on the master and then *not* see it on the slave -- the slave always has first visibility. I don't see how such a guarantee can exist in *either* direction with the semantics you describe. After seeing a transaction's work on one system it would always be unknown whether it was visible on the other. There are situations where that is OK as long as each copy has a sane order of visibility, but there are situations where losing that guarantee might matter. On the bright side, it means that transactions would become visible on the replica in the same order as on the master, and that blocking would be reduced. Having transactions become visible in the same order on the master and the standby is very appealing, but I'm pretty well convinced that allowing commits to become visible before they've been durably committed is throwing the D an ACID out the window. If synchronous_commit is off, sure, but otherwise... ...Robert -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 16:20 -0400, Robert Haas wrote: On Thu, Jul 28, 2011 at 3:40 PM, Hannu Krosing ha...@2ndquadrant.com wrote: On Thu, 2011-07-28 at 21:32 +0200, Hannu Krosing wrote: On Thu, 2011-07-28 at 14:27 -0400, Robert Haas wrote: Hmm, interesting idea. However, consider the scenario where some transactions are using synchronous_commit or synchronous replication, and others are not. If a transaction that needs to wait (either just for WAL flush, or for WAL flush and synchronous replication) inserts its commit record, and then another transaction with synchronous_commit=off comes along and inserts its commit record, the second transaction will have to block until the first transaction is done waiting. What is the current behavior when the synchronous replication fails (say the slave breaks down) - will the transaction be rolled back at some point or will it wait indefinitely , that is until a new slave is installed ? More importantly, if the master crashes after the commit is written to WAL, will the transaction be rolled back after recovery based on the fact that no confirmation from synchronous slave is received ? No. You can't roll back a transaction once it's committed - ever. so in case of stuck slave the syncrep transcation is committed after crash, but is not committed before the crash happens ? ow will the replay process get stuc gaian during recovery ? -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 4:36 PM, Hannu Krosing ha...@krosing.net wrote: so in case of stuck slave the syncrep transcation is committed after crash, but is not committed before the crash happens ? Yep. ow will the replay process get stuc gaian during recovery ? Nope. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Robert Haas robertmh...@gmail.com wrote: Having transactions become visible in the same order on the master and the standby is very appealing, but I'm pretty well convinced that allowing commits to become visible before they've been durably committed is throwing the D an ACID out the window. If synchronous_commit is off, sure, but otherwise... It has been durably committed on the master, but not on the supposedly synchronous copy; so it's not so much through out the D in ACID as throwing out the synchronous in synchronous replication. :-( Unless I'm missing something we have a choice to make -- I see four possibilities (already mentioned on this thread, I think): (1) Transactions are visible on the master which won't necessarily be there if a meteor takes out the master and you need to resume operations on the replica. (2) An asynchronous commit must block behind any pending synchronous commits if synchronous replication is in use. (3) Transactions become visible on the replica in a different order than they became visible on the master. (4) We communicate acceptable snapshots to the replica to make the order of visibility visibility match the master even when that doesn't match the order that transactions returned from commit. I don't see how we can accept (1) and call it synchronous replication. I'm pretty dubious about (3), because we don't even have Snapshot Isolation on the replica, really. Is (3) where we're currently at? An advantage of (4) is that on the replica we would get the same SI behavior at Repeatable Read that exists on the master, and we could even use the same mechanism for SSI to provide Serializable isolation on the replica. I (predictably) like (4) -- even though it's a lot of work -Kevin -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 14:27 -0400, Robert Haas wrote: Right, but if the visibility order were *defined* as the order in which commit records appear in WAL, that problem neatly goes away. It's only because we have the implementation artifact that set my xid to 0 in the ProcArray is decoupled from inserting the commit record that there's any difference. Hmm, interesting idea. However, consider the scenario where some transactions are using synchronous_commit or synchronous replication, and others are not. If a transaction that needs to wait (either just for WAL flush, or for WAL flush and synchronous replication) inserts its commit record, and then another transaction with synchronous_commit=off comes along and inserts its commit record, the second transaction will have to block until the first transaction is done waiting. We can't make either transaction visible without making both visible, and we certainly can't acknowledge the second transaction to the client until we've made it visible. I'm not going to say that's so horrible we shouldn't even consider it, but it doesn't seem great, either. I'm trying to follow along here. Wouldn't the same issue exist if one transaction is waiting for sync rep (synchronous_commit=on), and another is waiting for just a WAL flush (synchronous_commit=local)? I don't think that a synchronous_commit=off is required. Regards, Jeff Davis -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Jeff Davis pg...@j-davis.com wrote: Wouldn't the same issue exist if one transaction is waiting for sync rep (synchronous_commit=on), and another is waiting for just a WAL flush (synchronous_commit=local)? I don't think that a synchronous_commit=off is required. I think you're right -- basically, to make visibility atomic with commit and allow a fast snapshot build based on that order, any new commit request would need to block behind any pending request, regardless of that setting. At least, no way around that is apparent to me. -Kevin -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
Kevin Grittner kevin.gritt...@wicourts.gov wrote: to make visibility atomic with commit I meant: to make visibility atomic with WAL-write of the commit record -Kevin -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
- Цитат от Hannu Krosing (ha...@2ndquadrant.com), на 28.07.2011 в 22:40 - Maybe this is why other databases don't offer per backend async commit ? Isn't Oracle's COMMIT WRITE NOWAIT; basically the same - ad hoc async commit? Though their idea of backend do not maps exactly to postgrsql's idea. The closest thing is per session async commit: ALTER SESSION SET COMMIT_WRITE='NOWAIT'; Best regards -- Luben Karavelov
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 4:54 PM, Kevin Grittner
kevin.gritt...@wicourts.gov wrote:
Robert Haas robertmh...@gmail.com wrote:
Having transactions become visible in the same order on the master
and the standby is very appealing, but I'm pretty well convinced
that allowing commits to become visible before they've been
durably committed is throwing the D an ACID out the window. If
synchronous_commit is off, sure, but otherwise...
It has been durably committed on the master, but not on the
supposedly synchronous copy; so it's not so much through out the D
in ACID as throwing out the synchronous in synchronous
replication. :-(
Well, depends. Currently, the sequence of events is:
1. Insert commit record.
2. Flush commit record, if synchronous_commit in {local, on}.
3. Wait for synchronous replication, if synchronous_commit = on and
synchronous_standby_names is non-empty.
4. Make transaction visible.
If you move (4) before (3), you're throwing out the synchronous in
synchronous replication. If you move (4) before (2), you're throwing
out the D in ACID.
Unless I'm missing something we have a choice to make -- I see four
possibilities (already mentioned on this thread, I think):
(1) Transactions are visible on the master which won't necessarily
be there if a meteor takes out the master and you need to resume
operations on the replica.
(2) An asynchronous commit must block behind any pending
synchronous commits if synchronous replication is in use.
Well, again, there are three levels:
(A) synchronous_commit=off. No waiting!
(B) synchronous_commit=local transactions, and synchronous_commit=on
transactions when sync rep is not in use. Wait for xlog flush.
(C) synchronous_commit=on transactions when sync rep IS in use. Wait
for xlog flush and replication.
Under your option #2, if a type-A transaction commits after a type-B
transaction, it will need to wait for the type-B transaction's xlog
flush. If a type-A transaction commits after a type-C transaction, it
will need to wait for the type-C transaction to flush xlog and
replicate. And if a type-B transaction commits after a type-C
transaction, there's no additional waiting for xlog flush, because the
type-B transaction would have to wait for that anyway. But it will
also have to wait for the preceding type-C transaction to replicate.
So basically, you can't be more asynchronous than the guy in front of
you.
Aside from the fact that this behavior isn't too hot from a user
perspective, it might lead to some pretty complicated locking. Every
time a transaction finishes xlog flush or sync rep, it's got to go
release the transactions that piled up behind it - but not too many,
just up to the next one that still needs to wait on some higher LSN.
(3) Transactions become visible on the replica in a different order
than they became visible on the master.
(4) We communicate acceptable snapshots to the replica to make the
order of visibility visibility match the master even when that
doesn't match the order that transactions returned from commit.
I don't see how we can accept (1) and call it synchronous
replication. I'm pretty dubious about (3), because we don't even
have Snapshot Isolation on the replica, really. Is (3) where we're
currently at? An advantage of (4) is that on the replica we would
get the same SI behavior at Repeatable Read that exists on the
master, and we could even use the same mechanism for SSI to provide
Serializable isolation on the replica.
I (predictably) like (4) -- even though it's a lot of work
I think that (4), beyond being a lot of work, will also have pretty
terrible performance. You're basically talking about emitting two WAL
records for every commit instead of one. That's not going to be
awesome. It might be OK for small or relatively lightly loaded
systems, or those with big transactions. But for something like
pgbench or DBT-2, I think it's going to be a big problem. WAL is
already a major bottleneck for us; we need to find a way to make it
less of one, not more.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, 2011-07-28 at 16:42 -0400, Robert Haas wrote: On Thu, Jul 28, 2011 at 4:36 PM, Hannu Krosing ha...@krosing.net wrote: so in case of stuck slave the syncrep transcation is committed after crash, but is not committed before the crash happens ? Yep. ow will the replay process get stuc gaian during recovery ? Nope. Are you sure ? I mean the case when a stuck master comes up but slave is still not functional. How does this behavior currently fit in with ACID and sync guarantees ? -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 11:54 PM, Kevin Grittner kevin.gritt...@wicourts.gov wrote: (4) We communicate acceptable snapshots to the replica to make the order of visibility visibility match the master even when that doesn't match the order that transactions returned from commit. I wonder if some interpretation of 2 phase commit could make Robert's original suggestion implement this. On the master the commit sequence would look something like: 1. Insert commit record to the WAL 2. Wait for replication 3. Get a commit seq nr and mark XIDs visible 4. WAL log the seq nr 5. Return success to client When replaying: * When replaying commit record, do everything but make the tx visible. * When replaying the commit sequence number if there is a gap between last visible commit and current: insert the commit sequence nr. to list of waiting commits. else: mark current and all directly following waiting tx's visible This would give consistent visibility order on master and slave. Robert is right that this would undesirably increase WAL traffic. Delaying this traffic would undesirably increase replay lag between master and slave. But it seems to me that this could be an optional WAL level on top of hot_standby that would only be enabled if consistent visibility on slaves is desired. -- Ants Aasma -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Fri, Jul 29, 2011 at 2:20 AM, Robert Haas robertmh...@gmail.com wrote: Well, again, there are three levels: (A) synchronous_commit=off. No waiting! (B) synchronous_commit=local transactions, and synchronous_commit=on transactions when sync rep is not in use. Wait for xlog flush. (C) synchronous_commit=on transactions when sync rep IS in use. Wait for xlog flush and replication. ... So basically, you can't be more asynchronous than the guy in front of you. (A) still gives a guarantee - transactions that begin after the commit returns see the commited transaction. A weaker variant would say that if the commit returns, and the server doesn't crash in the meantime, the commit would at some point become visible. Maybe even that transactions that begin after the commit returns become visible after that commit. -- Ants Aasma -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 7:54 PM, Ants Aasma ants.aa...@eesti.ee wrote: On Thu, Jul 28, 2011 at 11:54 PM, Kevin Grittner kevin.gritt...@wicourts.gov wrote: (4) We communicate acceptable snapshots to the replica to make the order of visibility visibility match the master even when that doesn't match the order that transactions returned from commit. I wonder if some interpretation of 2 phase commit could make Robert's original suggestion implement this. On the master the commit sequence would look something like: 1. Insert commit record to the WAL 2. Wait for replication 3. Get a commit seq nr and mark XIDs visible 4. WAL log the seq nr 5. Return success to client When replaying: * When replaying commit record, do everything but make the tx visible. * When replaying the commit sequence number if there is a gap between last visible commit and current: insert the commit sequence nr. to list of waiting commits. else: mark current and all directly following waiting tx's visible This would give consistent visibility order on master and slave. Robert is right that this would undesirably increase WAL traffic. Delaying this traffic would undesirably increase replay lag between master and slave. But it seems to me that this could be an optional WAL level on top of hot_standby that would only be enabled if consistent visibility on slaves is desired. I think you nailed it. An additional point to think about: if we were willing to insist on streaming replication, we could send the commit sequence numbers via a side channel rather than writing them to WAL, which would be a lot cheaper. That might even be a reasonable thing to do, because if you're doing log shipping, this is all going to be super-not-real-time anyway. OTOH, I know we don't want to make WAL shipping anything less than a first class citizen, so maybe not. At any rate, we may be getting a little sidetracked here from the original point of the thread, which was how to make snapshot-taking cheaper. Maybe there's some tie-in to when transactions become visible, but I think it's pretty weak. The existing system could be hacked up to avoid making transactions visible out of LSN order, and the system I proposed could make them visible either in LSN order or do the same thing we do now. They are basically independent problems, AFAICS. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] cheaper snapshots
On Thu, Jul 28, 2011 at 8:12 PM, Ants Aasma ants.aa...@eesti.ee wrote: On Fri, Jul 29, 2011 at 2:20 AM, Robert Haas robertmh...@gmail.com wrote: Well, again, there are three levels: (A) synchronous_commit=off. No waiting! (B) synchronous_commit=local transactions, and synchronous_commit=on transactions when sync rep is not in use. Wait for xlog flush. (C) synchronous_commit=on transactions when sync rep IS in use. Wait for xlog flush and replication. ... So basically, you can't be more asynchronous than the guy in front of you. (A) still gives a guarantee - transactions that begin after the commit returns see the commited transaction. A weaker variant would say that if the commit returns, and the server doesn't crash in the meantime, the commit would at some point become visible. Maybe even that transactions that begin after the commit returns become visible after that commit. Yeah, you could do that. But that's such a weak guarantee that I'm not sure it has much practical utility. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
[HACKERS] cheaper snapshots
On Wed, Oct 20, 2010 at 10:07 PM, Tom Lane t...@sss.pgh.pa.us wrote: I wonder whether we could do something involving WAL properties --- the current tuple visibility logic was designed before WAL existed, so it's not exploiting that resource at all. I'm imagining that the kernel of a snapshot is just a WAL position, ie the end of WAL as of the time you take the snapshot (easy to get in O(1) time). Visibility tests then reduce to did this transaction commit with a WAL record located before the specified position?. You'd need some index datastructure that made it reasonably cheap to find out the commit locations of recently committed transactions, where recent means back to recentGlobalXmin. That seems possibly do-able, though I don't have a concrete design in mind. I was mulling this idea over some more (the same ideas keep floating back to the top...). I don't think an LSN can actually work, because there's no guarantee that the order in which the WAL records are emitted is the same order in which the effects of the transactions become visible to new snapshots. For example: 1. Transaction A inserts its commit record, flushes WAL, and begins waiting for sync rep. 2. A moment later, transaction B sets synchronous_commit=off, inserts its commit record, requests a background WAL flush, and removes itself from the ProcArray. 3. Transaction C takes a snapshot. Sync rep doesn't create this problem; there's a race anyway. The order of acquisition for WALInsertLock needn't match that for ProcArrayLock. This has the more-than-slightly-odd characteristic that you could end up with a snapshot on the master that can see A but not B and a snapshot on the slave that can see B but not A. But having said that an LSN can't work, I don't see why we can't just use a 64-bit counter. In fact, the predicate locking code already does something much like this, using an SLRU, for serializable transactions only. In more detail, what I'm imagining is an array with 4 billion entries, one per XID, probably broken up into files of say 16MB each with 2 million entries per file. Each entry is a 64-bit value. It is 0 if the XID has not yet started, is still running, or has aborted. Otherwise, it is the commit sequence number of the transaction. For reasons I'll explain below, I'm imagining starting the commit sequence number counter at some very large value and having it count down from there. So the basic operations are: - To take a snapshot, you just read the counter. - To commit a transaction which has an XID, you read the counter, stamp all your XIDs with that value, and decrement the counter. - To find out whether an XID is visible to your snapshot, you look up the XID in the array and get the counter value. If the value you read is greater than your snapshot value, it's visible. If it's less, it's not. Now, is this algorithm any good, and how little locking can we get away with? It seems to me that if we used an SLRU to store the array, the lock contention would be even worse than it is under our current system, wherein everybody fights over ProcArrayLock. A system like this is going to involve lots and lots of probes into the array (even if we build a per-backend cache of some kind) and an SLRU will require at least one LWLock acquire and release per probe. Some kind of locking is pretty much unavoidable, because you have to worry about pages getting evicted from shared memory. However, what if we used a set of files (like SLRU) but mapped them separately into each backend's address space? I think this would allow both loads and stores from the array to be done unlocked. One fly in the ointment is that 8-byte stores are apparently done as two 4-byte stores on some platforms. But if the counter runs backward, I think even that is OK. If you happen to read an 8 byte value as it's being written, you'll get 4 bytes of the intended value and 4 bytes of zeros. The value will therefore appear to be less than what it should be. However, if the value was in the midst of being written, then it's still in the midst of committing, which means that that XID wasn't going to be visible anyway. Accidentally reading a smaller value doesn't change the answer. Committing will require a lock on the counter. Taking a snapshot can be done unlocked if (1) 8-byte reads are atomic and either (2a) the architecture has strong memory ordering (no store/store reordering) or (2b) you insert a memory fence between stamping the XIDs and decrementing the counter. Otherwise, taking a snapshot will also require a lock on the counter. Once a particular XID precedes RecentGlobalXmin, you no longer care about the associated counter value. You just need to know that it committed; the order no longer matters. So after a crash, assuming that you have the CLOG bits available, you can just throw away all the array contents and start the counter over at the highest possible value. And, as RecentGlobalXmin advances, you can
