Re: [HACKERS] Re: [COMMITTERS] pgsql: Reduce spurious Hot Standby conflicts from never-visible records
On Wed, Jan 5, 2011 at 3:06 PM, Robert Haas wrote: > On Wed, Jan 5, 2011 at 3:00 PM, Simon Riggs wrote: >> On Mon, 2011-01-03 at 23:13 -0500, Robert Haas wrote: >> >>> > Hmmm, my earlier code took xmax only if xmax > xmin. That was wrong; >>> > what I have now is better, but your point is there may be an even better >>> > truth. I'll think on that a little more. >> >> I remember that I thought some more on this and decided that I couldn't >> see a problem. I also see I didn't update the list to say that. >> >>> I guess the problem case here is something like: >>> >>> 1. T1 begins. T1 writes a tuple A (so it gets an XID). >>> 2. T2 begins. T2 writes a tuple B (so it gets a later XID). >>> 3. T1 takes a new snapshot that can see B and deletes B. >>> 4. T2 commits. >>> 5. T1 commits. >> >> How is step (3) possible before step (4)? > > At read committed isolation level, which is the default, we take a new > snapshot after every command. Oh, I'm a dork. You're saying T2 hasn't committed yet. Let me think about this some 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] Re: [COMMITTERS] pgsql: Reduce spurious Hot Standby conflicts from never-visible records
On Wed, Jan 5, 2011 at 3:00 PM, Simon Riggs wrote: > On Mon, 2011-01-03 at 23:13 -0500, Robert Haas wrote: > >> > Hmmm, my earlier code took xmax only if xmax > xmin. That was wrong; >> > what I have now is better, but your point is there may be an even better >> > truth. I'll think on that a little more. > > I remember that I thought some more on this and decided that I couldn't > see a problem. I also see I didn't update the list to say that. > >> I guess the problem case here is something like: >> >> 1. T1 begins. T1 writes a tuple A (so it gets an XID). >> 2. T2 begins. T2 writes a tuple B (so it gets a later XID). >> 3. T1 takes a new snapshot that can see B and deletes B. >> 4. T2 commits. >> 5. T1 commits. > > How is step (3) possible before step (4)? At read committed isolation level, which is the default, we take a new snapshot after every command. > There are later errors in your example also. Well, point them out and let's discuss. -- 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] Re: [COMMITTERS] pgsql: Reduce spurious Hot Standby conflicts from never-visible records
On Mon, 2011-01-03 at 23:13 -0500, Robert Haas wrote: > > Hmmm, my earlier code took xmax only if xmax > xmin. That was wrong; > > what I have now is better, but your point is there may be an even better > > truth. I'll think on that a little more. I remember that I thought some more on this and decided that I couldn't see a problem. I also see I didn't update the list to say that. > I guess the problem case here is something like: > > 1. T1 begins. T1 writes a tuple A (so it gets an XID). > 2. T2 begins. T2 writes a tuple B (so it gets a later XID). > 3. T1 takes a new snapshot that can see B and deletes B. > 4. T2 commits. > 5. T1 commits. How is step (3) possible before step (4)? There are later errors in your example also. It's possible that that's all wrong because of how my head is feeling, if so please explain long hand so I don't have to think too much. -- Simon Riggs http://www.2ndQuadrant.com/books/ PostgreSQL Development, 24x7 Support, Training and 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] Re: [COMMITTERS] pgsql: Reduce spurious Hot Standby conflicts from never-visible records
On Sun, Dec 12, 2010 at 5:15 AM, Simon Riggs wrote: > On Sat, 2010-12-11 at 22:03 +0100, Heikki Linnakangas wrote: >> (Moving to pgsql-hackers) >> >> On 10.12.2010 20:21, Tom Lane wrote: >> > Simon Riggs writes: >> >> Reduce spurious Hot Standby conflicts from never-visible records. >> >> Hot Standby conflicts only with tuples that were visible at >> >> some point. So ignore tuples from aborted transactions or for >> >> tuples updated/deleted during the inserting transaction when >> >> generating the conflict transaction ids. >> > >> >> Following detailed analysis and test case by Noah Misch. >> >> Original report covered btree delete records, correctly observed >> >> by Heikki Linnakangas that this applies to other cases also. >> >> Fix covers all sources of cleanup records via common code. >> >> Includes additional fix compared to commit on HEAD >> > >> > ISTM HeapTupleHeaderAdvanceLatestRemovedXid is still pretty broken, >> > in that it's examining xmax without having checked that xmax is (a) >> > valid or (b) a lock rather than a deletion xmax. >> >> In current use, it's only called for tuples that are known to be dead, >> so either xmax is a valid deletion, or xmin didn't commit in which case >> the function doesn't use xmax for anything. So I think it actually works >> as it is. > > Well, I think you're both right. > > The function shouldn't be called in places where xmax is the wrong > flavour, but there should be specific safeguards in case of mistake. Should we do something about this? An Assert(), maybe? >> Also, I'm not totally convinced it's correct when xmin > xmax, despite >> Simon's follow-up commit to fix that. Shouldn't it advance >> latestRemovedXid to xmin in that case? Or maybe it's ok as it is because >> we know that xmax committed after xmin. The impression I get from the >> comment above the function now is that it advances latestRemovedXid to >> the highest XID present in the tuple, but that's not what it does in the >> xmin > xmax case. That comment needs clarification. > > Hmmm, my earlier code took xmax only if xmax > xmin. That was wrong; > what I have now is better, but your point is there may be an even better > truth. I'll think on that a little more. I guess the problem case here is something like: 1. T1 begins. T1 writes a tuple A (so it gets an XID). 2. T2 begins. T2 writes a tuple B (so it gets a later XID). 3. T1 takes a new snapshot that can see B and deletes B. 4. T2 commits. 5. T1 commits. At this point we have a tuple B with XMAX (T1's XID) < XMIN (T2's XID). Now, on the standby, there can be a transaction TS which takes a snapshot that can see T2's XID but not T1's XID. While that transaction is still running, VACUUM (or a HOT prune) comes along and zaps B, and this record is replayed on the standby, advancing latestRemovedXID to T1's XID, when in fact we also removed T2's later XID. This means we MUST kill TS (at least if it tries to read that block) because otherwise it'll fail to see B and return the wrong answer. Will we actually kill TS? GetConflictingVirtualXIDs() looks for transactions where proc->xmin precedes limitXmin, and if TS has a snapshot that can't see T1's XID then its proc->xmin might be exactly equal to limitXmin = latestRemovedXID = T1s XID, causing it to not get killed. I think. -- 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] Re: [COMMITTERS] pgsql: Reduce spurious Hot Standby conflicts from never-visible records
On Sun, 2010-12-12 at 10:15 +, Simon Riggs wrote: > > Also, I'm not totally convinced it's correct when xmin > xmax, > despite > > Simon's follow-up commit to fix that. Shouldn't it advance > > latestRemovedXid to xmin in that case? Or maybe it's ok as it is > because > > we know that xmax committed after xmin. The impression I get from > the comment above the function now is that it advances > latestRemovedXid to > > the highest XID present in the tuple, but that's not what it does in > the xmin > xmax case. That comment needs clarification. > > Hmmm, my earlier code took xmax only if xmax > xmin. That was wrong; > what I have now is better, but your point is there may be an even > better > truth. I'll think on that a little more. This has a stranger answer than I was expecting. HeapTupleSatisfiesVacuum() shows there is no interaction between the xmin of a tuple and the point at which it is removed, if the xmin transaction commits. So the hot standby conflict depends only upon the xmax, meaning that xmin > xmax is not a problem. So no immediate change to the code is warranted, on that specific point. -- Simon Riggs http://www.2ndQuadrant.com/books/ PostgreSQL Development, 24x7 Support, Training and Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
[HACKERS] Re: [COMMITTERS] pgsql: Reduce spurious Hot Standby conflicts from never-visible records
On Sat, 2010-12-11 at 22:03 +0100, Heikki Linnakangas wrote: > (Moving to pgsql-hackers) > > On 10.12.2010 20:21, Tom Lane wrote: > > Simon Riggs writes: > >> Reduce spurious Hot Standby conflicts from never-visible records. > >> Hot Standby conflicts only with tuples that were visible at > >> some point. So ignore tuples from aborted transactions or for > >> tuples updated/deleted during the inserting transaction when > >> generating the conflict transaction ids. > > > >> Following detailed analysis and test case by Noah Misch. > >> Original report covered btree delete records, correctly observed > >> by Heikki Linnakangas that this applies to other cases also. > >> Fix covers all sources of cleanup records via common code. > >> Includes additional fix compared to commit on HEAD > > > > ISTM HeapTupleHeaderAdvanceLatestRemovedXid is still pretty broken, > > in that it's examining xmax without having checked that xmax is (a) > > valid or (b) a lock rather than a deletion xmax. > > In current use, it's only called for tuples that are known to be dead, > so either xmax is a valid deletion, or xmin didn't commit in which case > the function doesn't use xmax for anything. So I think it actually works > as it is. Well, I think you're both right. The function shouldn't be called in places where xmax is the wrong flavour, but there should be specific safeguards in case of mistake. > I agree it doesn't look right, though. At the very least it needs > comments explaining that, but preferably it should do something sane > when faced with a tuple that's not dead after all. Perhaps throw an > error (though that would be bad during recovery), or an Assert, or just > refrain from advancing latestRemovedXid (or advance it, that would be > the conservative stance given the current use). Yes > Also, I'm not totally convinced it's correct when xmin > xmax, despite > Simon's follow-up commit to fix that. Shouldn't it advance > latestRemovedXid to xmin in that case? Or maybe it's ok as it is because > we know that xmax committed after xmin. The impression I get from the > comment above the function now is that it advances latestRemovedXid to > the highest XID present in the tuple, but that's not what it does in the > xmin > xmax case. That comment needs clarification. Hmmm, my earlier code took xmax only if xmax > xmin. That was wrong; what I have now is better, but your point is there may be an even better truth. I'll think on that a little more. > While we're at it, perhaps it would be better to move this function to > tqual.c. And I feel that a more natural interface would be something like: > > TransactionId > HeapTupleHeaderGetLatestRemovedXid(HeapTupleHeader tuple); > > IOW, instead bumping up the passed-in latestRemovedXid value, return the > highest XID on the tuple (if it was dead). > > PS. it would be good to set hint bits in that function like in > HeapTupleSatisfies* functions. I'm not that happy with refactoring inside a release, plus I'm not even sure if that is the right way. I suspect the best way would be to do this as a side-effect of HeapSatisfiesVacuum(), since this processing should only ever be done in conjunction with that function. Will respond later today on those thoughts. -- Simon Riggs http://www.2ndQuadrant.com/books/ PostgreSQL Development, 24x7 Support, Training and Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
[HACKERS] Re: [COMMITTERS] pgsql: Reduce spurious Hot Standby conflicts from never-visible records
(Moving to pgsql-hackers) On 10.12.2010 20:21, Tom Lane wrote: Simon Riggs writes: Reduce spurious Hot Standby conflicts from never-visible records. Hot Standby conflicts only with tuples that were visible at some point. So ignore tuples from aborted transactions or for tuples updated/deleted during the inserting transaction when generating the conflict transaction ids. Following detailed analysis and test case by Noah Misch. Original report covered btree delete records, correctly observed by Heikki Linnakangas that this applies to other cases also. Fix covers all sources of cleanup records via common code. Includes additional fix compared to commit on HEAD ISTM HeapTupleHeaderAdvanceLatestRemovedXid is still pretty broken, in that it's examining xmax without having checked that xmax is (a) valid or (b) a lock rather than a deletion xmax. In current use, it's only called for tuples that are known to be dead, so either xmax is a valid deletion, or xmin didn't commit in which case the function doesn't use xmax for anything. So I think it actually works as it is. I agree it doesn't look right, though. At the very least it needs comments explaining that, but preferably it should do something sane when faced with a tuple that's not dead after all. Perhaps throw an error (though that would be bad during recovery), or an Assert, or just refrain from advancing latestRemovedXid (or advance it, that would be the conservative stance given the current use). Also, I'm not totally convinced it's correct when xmin > xmax, despite Simon's follow-up commit to fix that. Shouldn't it advance latestRemovedXid to xmin in that case? Or maybe it's ok as it is because we know that xmax committed after xmin. The impression I get from the comment above the function now is that it advances latestRemovedXid to the highest XID present in the tuple, but that's not what it does in the xmin > xmax case. That comment needs clarification. While we're at it, perhaps it would be better to move this function to tqual.c. And I feel that a more natural interface would be something like: TransactionId HeapTupleHeaderGetLatestRemovedXid(HeapTupleHeader tuple); IOW, instead bumping up the passed-in latestRemovedXid value, return the highest XID on the tuple (if it was dead). PS. it would be good to set hint bits in that function like in HeapTupleSatisfies* functions. -- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
