Re: [HACKERS] [BUGS] BUG #14825: enum type: unsafe use?

2017-10-05 Thread Robert Haas 
On Tue, Oct 3, 2017 at 9:38 PM, Andres Freund  wrote:
>> Do you have any suggestion as to how we should transmit the blacklist to
>> parallel workers?
>
> How about storing them in the a dshash table instead of dynahash?
> Similar to how we're now dealing with the shared typmod registry stuff?
> It should be fairly simple to now simply add a new struct Session member
> shared_enum_whatevs_table.

Yeah, that approach seems worth exploring.

-- 
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] [BUGS] BUG #14825: enum type: unsafe use?

2017-10-03 Thread Andres Freund 
On 2017-10-03 19:53:41 -0400, Andrew Dunstan wrote:
> On 09/27/2017 02:52 PM, Tom Lane wrote:
> > Andrew Dunstan  writes:
> >> At this stage on reflection I agree it should be pulled :-(
> > That seems to be the consensus, so I'll go make it happen.
> >
> >> I'm not happy about the idea of marking an input function as not
> >> parallel safe, certainly not without a good deal of thought and
> >> discussion that we don't have time for this cycle.
> > I think the way forward is to do what we had as of HEAD (984c92074),
> > but add the ability to transmit the blacklist table to parallel
> > workers.  Since we expect the blacklist table would be empty most of
> > the time, this should be close to no overhead in practice.  I concur
> > that the idea of marking the relevant functions parallel-restricted is
> > probably not as safe a fix as I originally thought, and it's not a
> > very desirable restriction even if it did fix the problem.

> Do you have any suggestion as to how we should transmit the blacklist to
> parallel workers?

How about storing them in the a dshash table instead of dynahash?
Similar to how we're now dealing with the shared typmod registry stuff?
It should be fairly simple to now simply add a new struct Session member
shared_enum_whatevs_table.

Greetings,

Andres Freund


-- 
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] [BUGS] BUG #14825: enum type: unsafe use?

2017-10-03 Thread Tom Lane 
Andrew Dunstan  writes:
> Do you have any suggestion as to how we should transmit the blacklist to
> parallel workers?

Perhaps serialize the contents into an array in DSM, then rebuild a hash
table from that in the worker.  Robert might have a better idea though.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-10-03 Thread Andrew Dunstan 


On 09/27/2017 02:52 PM, Tom Lane wrote:
> Andrew Dunstan  writes:
>> At this stage on reflection I agree it should be pulled :-(
> That seems to be the consensus, so I'll go make it happen.
>
>> I'm not happy about the idea of marking an input function as not
>> parallel safe, certainly not without a good deal of thought and
>> discussion that we don't have time for this cycle.
> I think the way forward is to do what we had as of HEAD (984c92074),
> but add the ability to transmit the blacklist table to parallel
> workers.  Since we expect the blacklist table would be empty most of
> the time, this should be close to no overhead in practice.  I concur
> that the idea of marking the relevant functions parallel-restricted is
> probably not as safe a fix as I originally thought, and it's not a
> very desirable restriction even if it did fix the problem.
>
>   


Do you have any suggestion as to how we should transmit the blacklist to
parallel workers?

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-27 Thread Tom Lane 
Andrew Dunstan  writes:
> At this stage on reflection I agree it should be pulled :-(

That seems to be the consensus, so I'll go make it happen.

> I'm not happy about the idea of marking an input function as not
> parallel safe, certainly not without a good deal of thought and
> discussion that we don't have time for this cycle.

I think the way forward is to do what we had as of HEAD (984c92074),
but add the ability to transmit the blacklist table to parallel
workers.  Since we expect the blacklist table would be empty most of
the time, this should be close to no overhead in practice.  I concur
that the idea of marking the relevant functions parallel-restricted is
probably not as safe a fix as I originally thought, and it's not a
very desirable restriction even if it did fix the problem.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Tom Lane 
Andrew Dunstan  writes:
> I'm not happy about the idea of marking an input function as not
> parallel safe, certainly not without a good deal of thought and
> discussion that we don't have time for this cycle.

Yeah, that aspect of it was bothering me too: it's easy to say
"mark the function unsafe", but that only helps to the extent that
the function is used in queries where the planner has control of
whether to parallelize or not.  There's an awful lot of hard-wired
calls to I/O functions in our code, and I would not want to promise
that none of those are reachable in a parallel worker.

As for Stephen's concern, I had already looked at reverting 15bc038f9
earlier, and concluded that none of that code had changed significantly
since then.  There's some conflicts due to pgindent activity but I think
pulling it out will be a straightforward thing to do.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Andrew Dunstan 


On 09/26/2017 06:04 PM, Andrew Dunstan wrote:
>
> On 09/26/2017 05:45 PM, Stephen Frost wrote:
>> I've not been following along very closely- are we sure that ripping
>> this out won't be worse than dealing with it in-place?  Will pulling it
>> out also require a post-RC1 catversion bump?
>>
>>
>
> It shouldn't do AFAIK - the function signatures weren't changed.
>


At this stage on reflection I agree it should be pulled :-(

I'm not happy about the idea of marking an input function as not
parallel safe, certainly not without a good deal of thought and
discussion that we don't have time for this cycle.

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Andrew Dunstan 


On 09/26/2017 05:45 PM, Stephen Frost wrote:
>
> I've not been following along very closely- are we sure that ripping
> this out won't be worse than dealing with it in-place?  Will pulling it
> out also require a post-RC1 catversion bump?
>
>


It shouldn't do AFAIK - the function signatures weren't changed.

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Stephen Frost 
Tom, all,

* Tom Lane (t...@sss.pgh.pa.us) wrote:
> Bruce Momjian  writes:
> > On Tue, Sep 26, 2017 at 04:07:02PM -0400, Tom Lane wrote:
> >> Any other votes out there?
> 
> > Well, I was concerned yesterday that we had a broken build farm so close
> > to release. (I got consistent regression failures.)  I think PG 11 would
> > be better for this feature change, so I support reverting this.
> 
> I'll take the blame for (most of) yesterday's failures in the v10
> branch, but they were unrelated to this patch --- they were because
> of that SIGBUS patch I messed up.  So that doesn't seem like a very
> applicable argument.  Still, it's true that this seems like the most
> consequential patch that's gone into v10 post-RC1, certainly so if
> you discount stuff that was back-patched further than v10.

I've not been following along very closely- are we sure that ripping
this out won't be worse than dealing with it in-place?  Will pulling it
out also require a post-RC1 catversion bump?

If we can pull it out without bumping catversion and with confidence
that it won't cause more problems then, as much as I hate it, I'm
inclined to say we pull it out and come back to it in v11.  I really
don't like the idea of a post-rc1 catversion bump and it doesn't seem
like there's a good solution here that doesn't involve more changes and
most likely a catversion bump.  If it was reasonably fixable with only
small/local changes and without a catversion bump then I'd be more
inclined to keep it, but I gather from the discussion that's not the
case.

Thanks!

Stephen


signature.asc
Description: Digital signature

Re: [HACKERS] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Bruce Momjian 
On Tue, Sep 26, 2017 at 05:32:15PM -0400, Tom Lane wrote:
> Bruce Momjian  writes:
> > On Tue, Sep 26, 2017 at 04:07:02PM -0400, Tom Lane wrote:
> >> Any other votes out there?
> 
> > Well, I was concerned yesterday that we had a broken build farm so close
> > to release. (I got consistent regression failures.)  I think PG 11 would
> > be better for this feature change, so I support reverting this.
> 
> I'll take the blame for (most of) yesterday's failures in the v10
> branch, but they were unrelated to this patch --- they were because
> of that SIGBUS patch I messed up.  So that doesn't seem like a very
> applicable argument.  Still, it's true that this seems like the most
> consequential patch that's gone into v10 post-RC1, certainly so if
> you discount stuff that was back-patched further than v10.

Oh, I couldn't untangle that the regression failures were unrelated to
enums, so please ignore my opinion.

-- 
  Bruce Momjian  http://momjian.us
  EnterpriseDB http://enterprisedb.com

+ As you are, so once was I.  As I am, so you will be. +
+  Ancient Roman grave inscription +


-- 
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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Tom Lane 
Bruce Momjian  writes:
> On Tue, Sep 26, 2017 at 04:07:02PM -0400, Tom Lane wrote:
>> Any other votes out there?

> Well, I was concerned yesterday that we had a broken build farm so close
> to release. (I got consistent regression failures.)  I think PG 11 would
> be better for this feature change, so I support reverting this.

I'll take the blame for (most of) yesterday's failures in the v10
branch, but they were unrelated to this patch --- they were because
of that SIGBUS patch I messed up.  So that doesn't seem like a very
applicable argument.  Still, it's true that this seems like the most
consequential patch that's gone into v10 post-RC1, certainly so if
you discount stuff that was back-patched further than v10.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Bruce Momjian 
On Tue, Sep 26, 2017 at 04:07:02PM -0400, Tom Lane wrote:
> Andrew Dunstan  writes:
> > On 09/26/2017 02:37 PM, Tom Lane wrote:
> >> ... and the buildfarm's not too happy.  It looks like force_parallel_mode
> >> breaks all the regression test cases around unsafe enums; which on
> >> reflection is unsurprising, because parallel workers will not have access
> >> to the parent's blacklist hash, so they will think unsafe values are safe.
> 
> > I think I would mark enum_in and friends as parallel-restricted. Yes I
> > know it would involve a cat version bump, so I'll understand if that's
> > not acceptable, but it seems to me the best of a bad bunch of choices.
> > Second choice might be turning off parallel mode if the hash exists, but
> > I'm unclear how that would work.
> 
> Meh.  I'm starting to slide back to my original opinion that we should
> revert back to 9.6 behavior.  Even if a post-RC1 catversion bump is OK,
> making these sorts of changes a week before GA is not comfort inducing.
> I'm losing faith that we've thought through the issue thoroughly, and
> there's no longer time to catch any remaining oversights through testing.
> 
> Any other votes out there?

Well, I was concerned yesterday that we had a broken build farm so close
to release. (I got consistent regression failures.)  I think PG 11 would
be better for this feature change, so I support reverting this.

-- 
  Bruce Momjian  http://momjian.us
  EnterpriseDB http://enterprisedb.com

+ As you are, so once was I.  As I am, so you will be. +
+  Ancient Roman grave inscription +


-- 
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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Tom Lane 
Andrew Dunstan  writes:
> On 09/26/2017 02:37 PM, Tom Lane wrote:
>> ... and the buildfarm's not too happy.  It looks like force_parallel_mode
>> breaks all the regression test cases around unsafe enums; which on
>> reflection is unsurprising, because parallel workers will not have access
>> to the parent's blacklist hash, so they will think unsafe values are safe.

> I think I would mark enum_in and friends as parallel-restricted. Yes I
> know it would involve a cat version bump, so I'll understand if that's
> not acceptable, but it seems to me the best of a bad bunch of choices.
> Second choice might be turning off parallel mode if the hash exists, but
> I'm unclear how that would work.

Meh.  I'm starting to slide back to my original opinion that we should
revert back to 9.6 behavior.  Even if a post-RC1 catversion bump is OK,
making these sorts of changes a week before GA is not comfort inducing.
I'm losing faith that we've thought through the issue thoroughly, and
there's no longer time to catch any remaining oversights through testing.

Any other votes out there?

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Andrew Dunstan 


On 09/26/2017 02:37 PM, Tom Lane wrote:
> I wrote:
>> Pushed; sorry for the delay.
> ... and the buildfarm's not too happy.  It looks like force_parallel_mode
> breaks all the regression test cases around unsafe enums; which on
> reflection is unsurprising, because parallel workers will not have access
> to the parent's blacklist hash, so they will think unsafe values are safe.
>
> Now, as long as parallel workers are read-only, perhaps this matters
> little; they would not be allowed to write unsafe values into tables
> anyway.  I'm concerned though about whether it might be possible for a
> parallel worker to return an unsafe value to the parent (in OID form)
> and then the parent writes it into a table.  If we can convince ourselves
> that's not possible, it might be okay to just turn off force_parallel_mode
> for these test cases.
>
> A safer answer would be to mark enum_in() and other callers of
> check_safe_enum_use() as parallel-restricted.  That'd require a
> post-RC1 catversion bump, which seems pretty unpleasant, but
> none of the other answers are nice either.
>
> Transmitting the blacklist hash to workers would be a good long-term
> answer, but I don't want to try to shoehorn it in for v10.
>
> Another idea is that maybe the existence of a blacklist hash should
> be enough to turn off parallel mode altogether ... but ugh.
>
> Or maybe we're back to "revert the whole feature, go back to 9.6
> behavior".
>
> Thoughts?


I think I would mark enum_in and friends as parallel-restricted. Yes I
know it would involve a cat version bump, so I'll understand if that's
not acceptable, but it seems to me the best of a bad bunch of choices.
Second choice might be turning off parallel mode if the hash exists, but
I'm unclear how that would work.

cheers

andrew

-- 

Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Tom Lane 
I wrote:
> Pushed; sorry for the delay.

... and the buildfarm's not too happy.  It looks like force_parallel_mode
breaks all the regression test cases around unsafe enums; which on
reflection is unsurprising, because parallel workers will not have access
to the parent's blacklist hash, so they will think unsafe values are safe.

Now, as long as parallel workers are read-only, perhaps this matters
little; they would not be allowed to write unsafe values into tables
anyway.  I'm concerned though about whether it might be possible for a
parallel worker to return an unsafe value to the parent (in OID form)
and then the parent writes it into a table.  If we can convince ourselves
that's not possible, it might be okay to just turn off force_parallel_mode
for these test cases.

A safer answer would be to mark enum_in() and other callers of
check_safe_enum_use() as parallel-restricted.  That'd require a
post-RC1 catversion bump, which seems pretty unpleasant, but
none of the other answers are nice either.

Transmitting the blacklist hash to workers would be a good long-term
answer, but I don't want to try to shoehorn it in for v10.

Another idea is that maybe the existence of a blacklist hash should
be enough to turn off parallel mode altogether ... but ugh.

Or maybe we're back to "revert the whole feature, go back to 9.6
behavior".

Thoughts?

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-26 Thread Tom Lane 
I wrote:
> Andrew Dunstan  writes:
>> OK, that seems to be the consensus. So let's apply the blacklist patch
>> and then separately remove the 'created in the same transaction' test.
>> We'll need to adjust the regression tests and docs accordingly.

> Agreed.  I'll work on that in a little bit.

Pushed; sorry for the delay.

I noticed that the blacklist mechanism effectively removed the prohibition
against using a renamed enum value later in the same transaction, so I
added a regression test for that.  Also, as committed, I used RENAME TYPE
rather than ALTER OWNER in the test cases requiring an updated pg_type
row.  That way we don't need to create a role, even a transient one, which
is a good thing in terms of not risking collisions with other sessions.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-25 Thread Tom Lane 
Andrew Dunstan  writes:
> OK, that seems to be the consensus. So let's apply the blacklist patch
> and then separately remove the 'created in the same transaction' test.
> We'll need to adjust the regression tests and docs accordingly.

Agreed.  I'll work on that in a little bit.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-25 Thread Andrew Dunstan 


On 09/25/2017 01:34 PM, David E. Wheeler wrote:
> On Sep 25, 2017, at 10:55, Andrew Dunstan  
> wrote:
>
>> Let's ask a couple of users who I think are or have been actually
>> hurting on this point. Christophe and David, any opinions?
> If I understand the issue correctly, I think I’d be fine with requiring ALTER 
> TYPE ADD LABEL to be disallowed in a transaction that also CREATEs the type 
> if it’s not currently possible to reliably tell when an enum was created in a 
> transaction. Once you can do that, then by all means allow it!
>


OK, that seems to be the consensus. So let's apply the blacklist patch
and then separately remove the 'created in the same transaction' test.
We'll need to adjust the regression tests and docs accordingly.

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-25 Thread David E. Wheeler 
On Sep 25, 2017, at 10:55, Andrew Dunstan  
wrote:

> Let's ask a couple of users who I think are or have been actually
> hurting on this point. Christophe and David, any opinions?

If I understand the issue correctly, I think I’d be fine with requiring ALTER 
TYPE ADD LABEL to be disallowed in a transaction that also CREATEs the type if 
it’s not currently possible to reliably tell when an enum was created in a 
transaction. Once you can do that, then by all means allow it!

My $2.

Best,

David



signature.asc
Description: Message signed with OpenPGP

Re: [HACKERS] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-25 Thread Christophe Pettus 

> On Sep 25, 2017, at 07:55, Andrew Dunstan  
> wrote:
> Let's ask a couple of users who I think are or have been actually
> hurting on this point. Christophe and David, any opinions?

Since about 90% of what I encounter in this area are automatically-generated 
migrations, having a clear set of (perhaps restrictive) rules which never fail 
is the most important.  It's easy to split the CREATE or ALTERs out into their 
own transaction, and leave usage (such as populating a table from a migration) 
to a second transaction.

It's not clear to me that this is a vote either way, but I think the easiest 
thing to explain ("you cannot use a new enum value in the same transaction that 
created it") is the best in this situation.

-- 
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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-25 Thread Andrew Dunstan 


On 09/25/2017 10:42 AM, Tom Lane wrote:
> Andrew Dunstan  writes:
>> On 09/25/2017 10:14 AM, Tom Lane wrote:
>>> Oh ... I did not think we were on the same page, because your patch
>>> didn't include removal of the same-transaction heuristic.  It'd be
>>> sensible to do that as a separate patch, though, to make it easier
>>> to put back if we decide we do want it.
>> I understood you to say that the blacklist patch was all we needed to do
>> for v10. That's my position, i.e. I think we can live with the heuristic
>> test for now if the blacklist patch is applied. Maybe we need to
>> document that the heuristic test can generate some false negatives when
>> testing for a type that is created in the current transaction.
> No, as I said upthread, I want the heuristic out of there.  I think the
> blacklist idea covers enough use-cases that we possibly don't need the
> same-transaction test at all.  Furthermore I'm doubtful that the heuristic
> form of the same-transaction test is adequate to satisfy the use-cases
> that the blacklist test doesn't cover.  So I think we should remove that
> test and see whether we get any complaints, and if so what the details of
> the real-world use-cases look like.
>
>   



Let's ask a couple of users who I think are or have been actually
hurting on this point. Christophe and David, any opinions?

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-25 Thread Tom Lane 
Andrew Dunstan  writes:
> On 09/25/2017 10:14 AM, Tom Lane wrote:
>> Oh ... I did not think we were on the same page, because your patch
>> didn't include removal of the same-transaction heuristic.  It'd be
>> sensible to do that as a separate patch, though, to make it easier
>> to put back if we decide we do want it.

> I understood you to say that the blacklist patch was all we needed to do
> for v10. That's my position, i.e. I think we can live with the heuristic
> test for now if the blacklist patch is applied. Maybe we need to
> document that the heuristic test can generate some false negatives when
> testing for a type that is created in the current transaction.

No, as I said upthread, I want the heuristic out of there.  I think the
blacklist idea covers enough use-cases that we possibly don't need the
same-transaction test at all.  Furthermore I'm doubtful that the heuristic
form of the same-transaction test is adequate to satisfy the use-cases
that the blacklist test doesn't cover.  So I think we should remove that
test and see whether we get any complaints, and if so what the details of
the real-world use-cases look like.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-25 Thread Andrew Dunstan 


On 09/25/2017 10:14 AM, Tom Lane wrote:
> Andrew Dunstan  writes:
>> On 09/24/2017 07:06 PM, Tom Lane wrote:
>>> So I think we should just stop with the blacklist test for v10,
>>> and then see if we still get complaints (and exactly what they're
>>> about) so that we can judge how much more work the problem deserves.
>>> It's still ahead of where we were in previous releases, and ahead of
>>> where we'd be if we end up reverting the patch altogether.
>> That's pretty much what I was saying.
> Oh ... I did not think we were on the same page, because your patch
> didn't include removal of the same-transaction heuristic.  It'd be
> sensible to do that as a separate patch, though, to make it easier
> to put back if we decide we do want it.
>
>   


I understood you to say that the blacklist patch was all we needed to do
for v10. That's my position, i.e. I think we can live with the heuristic
test for now if the blacklist patch is applied. Maybe we need to
document that the heuristic test can generate some false negatives when
testing for a type that is created in the current transaction.

cheers

andrew

-- 

Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-25 Thread Tom Lane 
Andrew Dunstan  writes:
> On 09/24/2017 07:06 PM, Tom Lane wrote:
>> So I think we should just stop with the blacklist test for v10,
>> and then see if we still get complaints (and exactly what they're
>> about) so that we can judge how much more work the problem deserves.
>> It's still ahead of where we were in previous releases, and ahead of
>> where we'd be if we end up reverting the patch altogether.

> That's pretty much what I was saying.

Oh ... I did not think we were on the same page, because your patch
didn't include removal of the same-transaction heuristic.  It'd be
sensible to do that as a separate patch, though, to make it easier
to put back if we decide we do want it.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-24 Thread Andrew Dunstan 


On 09/24/2017 07:06 PM, Tom Lane wrote:
>
> So I think we should just stop with the blacklist test for v10,
> and then see if we still get complaints (and exactly what they're
> about) so that we can judge how much more work the problem deserves.
> It's still ahead of where we were in previous releases, and ahead of
> where we'd be if we end up reverting the patch altogether.
>
>


That's pretty much what I was saying.

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-24 Thread Tom Lane 
Andrew Dunstan  writes:
> On 09/24/2017 04:37 PM, Tom Lane wrote:
>> What we still need to debate is whether to remove the heuristic
>> type-is-from-same-transaction test, making the user-visible behavior
>> simply "you must commit an ALTER TYPE ADD VALUE before you can use the
>> new value".  I'm kind of inclined to do so; the fuzzy (and inadequately
>> documented) behavior we'll have if we keep it doesn't seem very nice to
>> me.

> I'd rather not. The failure cases are going to be vanishingly small, I
> suspect, and we've already discussed how we might improve that test. If
> you want to put some weasel words in the docs that might be ok.

I'm unconvinced.  We get enough complaints about heuristic behaviors
we have elsewhere.  Also, if we ship it like this, we're going to
have backward compatibility concerns if we try to change the behavior
later.  Now admittedly, the next step forward might well be an exact
solution which would necessarily take every case the heuristic allows
--- but I don't want to box us into having to support exactly the
cases the heuristic would allow.  And I don't want to have to
document which those are, either.

Basically, I don't think anyone's shown an important use case that
wouldn't be covered by "committed or not blacklisted".  That fixes
the original complaint that you couldn't do ALTER ADD VALUE in a
transaction block at all, and with or without the heuristic test,
you can't use the added value without committing.  The case not
covered is where an enum type is built with multiple commands in a
single transaction --- which might be of value, but since it doesn't
work for every such case, we don't know if the heuristic is really
going to provide useful value-add or not.

So I think we should just stop with the blacklist test for v10,
and then see if we still get complaints (and exactly what they're
about) so that we can judge how much more work the problem deserves.
It's still ahead of where we were in previous releases, and ahead of
where we'd be if we end up reverting the patch altogether.

Or in short: having been burned by this heuristic already, I want
it out of there.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-24 Thread Andrew Dunstan 


On 09/24/2017 04:37 PM, Tom Lane wrote:
> Andrew Dunstan  writes:
>> OK, here's the finished patch. It has a pretty small footprint all
>> things considered, and I think it guarantees that nothing that could be
>> done in this area in 9.6 will be forbidden. That's probably enough to
>> get us to 10 without having to revert the whole thing, ISTM, and we can
>> leave any further refinement to the next release.
> I think this could do with some more work on the comments and test cases,
> but it's basically sound.
>
> What we still need to debate is whether to remove the heuristic
> type-is-from-same-transaction test, making the user-visible behavior
> simply "you must commit an ALTER TYPE ADD VALUE before you can use the
> new value".  I'm kind of inclined to do so; the fuzzy (and inadequately
> documented) behavior we'll have if we keep it doesn't seem very nice to
> me.
>
>   



I'd rather not. The failure cases are going to be vanishingly small, I
suspect, and we've already discussed how we might improve that test. If
you want to put some weasel words in the docs that might be ok.

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-24 Thread Tom Lane 
Andrew Dunstan  writes:
> OK, here's the finished patch. It has a pretty small footprint all
> things considered, and I think it guarantees that nothing that could be
> done in this area in 9.6 will be forbidden. That's probably enough to
> get us to 10 without having to revert the whole thing, ISTM, and we can
> leave any further refinement to the next release.

I think this could do with some more work on the comments and test cases,
but it's basically sound.

What we still need to debate is whether to remove the heuristic
type-is-from-same-transaction test, making the user-visible behavior
simply "you must commit an ALTER TYPE ADD VALUE before you can use the
new value".  I'm kind of inclined to do so; the fuzzy (and inadequately
documented) behavior we'll have if we keep it doesn't seem very nice to
me.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-24 Thread Andrew Dunstan 


On 09/23/2017 06:06 PM, Tom Lane wrote:
> Andrew Dunstan  writes:
>> OK, I think I'm convinced. Here's is the WIP code I put together for the
>> blacklist. I'm was looking for a place to put the init call, but since
>> it's possibly not going anywhere I stopped :-) . My initial thought
>> about substransactions was that we should ignore them for this purpose
>> (That's why I used TopTransactionContext for the table).
> For the blacklist, I agree we could just ignore subtransactions: all
> subtransaction levels are equally uncommitted for this purpose, and
> leaving entries from failed subtransactions in place seems like a
> non-issue, since they'd never be referenced again.  (Well, barring OID
> wraparound and an enum-value-OID collision while the transaction runs,
> but I think we can ignore that as having probability epsilon.)
>
> But you need to actually put the table in TopTransactionContext, not
> CurTransactionContext ;-).  Also, I don't think you need an init call
> so much as an end-of-transaction cleanup call.  Maybe call it
> AtEOXactEnum(), for consistency with other functions called in the
> same area.
>
>> w.r.t. table size - how large? I confess I haven't seen any systems with
>> more than a few hundred enum types. But even a million or two shouldn't
>> consume a huge amount of memory, should it?
> Dynahash tables are self-expanding, so I don't see a need to stress about
> that too much.  Anything in 10-100 seems reasonable for initial size.
>



OK, here's the finished patch. It has a pretty small footprint all
things considered, and I think it guarantees that nothing that could be
done in this area in 9.6 will be forbidden. That's probably enough to
get us to 10 without having to revert the whole thing, ISTM, and we can
leave any further refinement to the next release.

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

diff --git a/src/backend/access/transam/xact.c b/src/backend/access/transam/xact.c
index 93dca7a..1d6f774 100644
--- a/src/backend/access/transam/xact.c
+++ b/src/backend/access/transam/xact.c
@@ -32,6 +32,7 @@
 #include "access/xlogutils.h"
 #include "catalog/catalog.h"
 #include "catalog/namespace.h"
+#include "catalog/pg_enum.h"
 #include "catalog/storage.h"
 #include "commands/async.h"
 #include "commands/tablecmds.h"
@@ -2126,6 +2127,7 @@ CommitTransaction(void)
 	smgrDoPendingDeletes(true);
 
 	AtCommit_Notify();
+	AtEOXact_Enum();
 	AtEOXact_GUC(true, 1);
 	AtEOXact_SPI(true);
 	AtEOXact_on_commit_actions(true);
@@ -2405,6 +2407,7 @@ PrepareTransaction(void)
 
 	/* PREPARE acts the same as COMMIT as far as GUC is concerned */
 	AtEOXact_GUC(true, 1);
+	AtEOXact_Enum();
 	AtEOXact_SPI(true);
 	AtEOXact_on_commit_actions(true);
 	AtEOXact_Namespace(true, false);
@@ -2606,6 +2609,7 @@ AbortTransaction(void)
 			 false, true);
 		smgrDoPendingDeletes(false);
 
+		AtEOXact_Enum();
 		AtEOXact_GUC(false, 1);
 		AtEOXact_SPI(false);
 		AtEOXact_on_commit_actions(false);
diff --git a/src/backend/catalog/pg_enum.c b/src/backend/catalog/pg_enum.c
index fe61d4d..3056f68 100644
--- a/src/backend/catalog/pg_enum.c
+++ b/src/backend/catalog/pg_enum.c
@@ -28,6 +28,8 @@
 #include "utils/builtins.h"
 #include "utils/catcache.h"
 #include "utils/fmgroids.h"
+#include "utils/hsearch.h"
+#include "utils/memutils.h"
 #include "utils/syscache.h"
 #include "utils/tqual.h"
 
@@ -38,6 +40,8 @@ Oid			binary_upgrade_next_pg_enum_oid = InvalidOid;
 static void RenumberEnumType(Relation pg_enum, HeapTuple *existing, int nelems);
 static int	sort_order_cmp(const void *p1, const void *p2);
 
+/* hash table of values added in the current transaction by AddEnumLabel */
+static HTAB *enum_blacklist = NULL;
 
 /*
  * EnumValuesCreate
@@ -460,8 +464,49 @@ restart:
 	heap_freetuple(enum_tup);
 
 	heap_close(pg_enum, RowExclusiveLock);
+
+	/* Set up the blacklist hash if required */
+	if (enum_blacklist == NULL)
+	{
+		HASHCTL hash_ctl;
+		memset(_ctl, 0, sizeof(hash_ctl));
+		hash_ctl.keysize = sizeof(Oid);
+		hash_ctl.entrysize = sizeof(Oid);
+		hash_ctl.hcxt = TopTransactionContext;
+		enum_blacklist = hash_create("Enum blacklist for current transaction",
+		   32,
+		   _ctl,
+		   HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
+
+	}
+
+	/* Add the new value to the blacklist */
+	(void) hash_search(enum_blacklist, , HASH_ENTER, NULL);
 }
 
+/* Test if the enum is on the blacklist */
+bool
+EnumBlacklisted(Oid enum_id)
+{
+	bool found;
+
+	if (enum_blacklist == NULL)
+		return false;
+
+	(void) hash_search(enum_blacklist, _id, HASH_FIND, );
+	return found;
+}
+
+/*
+ * Clean up the blacklist hash at the end of the transaction. The memory will
+ * have been deallocated, so all we need to do is set the pointer back to
+ * NULL for the next transaction.
+ */
+void
+AtEOXact_Enum(void)
+{
+	enum_blacklist = NULL;
+}
 
 /*
  * RenameEnumLabel
diff --git

Re: [HACKERS] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-23 Thread Tom Lane 
Andrew Dunstan  writes:
> OK, I think I'm convinced. Here's is the WIP code I put together for the
> blacklist. I'm was looking for a place to put the init call, but since
> it's possibly not going anywhere I stopped :-) . My initial thought
> about substransactions was that we should ignore them for this purpose
> (That's why I used TopTransactionContext for the table).

For the blacklist, I agree we could just ignore subtransactions: all
subtransaction levels are equally uncommitted for this purpose, and
leaving entries from failed subtransactions in place seems like a
non-issue, since they'd never be referenced again.  (Well, barring OID
wraparound and an enum-value-OID collision while the transaction runs,
but I think we can ignore that as having probability epsilon.)

But you need to actually put the table in TopTransactionContext, not
CurTransactionContext ;-).  Also, I don't think you need an init call
so much as an end-of-transaction cleanup call.  Maybe call it
AtEOXactEnum(), for consistency with other functions called in the
same area.

> w.r.t. table size - how large? I confess I haven't seen any systems with
> more than a few hundred enum types. But even a million or two shouldn't
> consume a huge amount of memory, should it?

Dynahash tables are self-expanding, so I don't see a need to stress about
that too much.  Anything in 10-100 seems reasonable for initial size.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-23 Thread Andrew Dunstan 


On 09/23/2017 03:52 PM, Tom Lane wrote:
> Andrew Dunstan  writes:
>> On 09/23/2017 02:00 PM, Tom Lane wrote:
>>> So I'm back to not being sure about the path forward.  Maybe it would be
>>> all right to say "the value added by ADD VALUE can't be used in the same
>>> transaction, period".  That's still a step forward compared to the pre-v10
>>> prohibition on doing it at all.  I don't remember if there were use-cases
>>> where we really needed the exception for new-in-transaction types.
>> Well, my idea was to have the test run like this:
>>   * is the value an old one? Test txnid of tuple. If yes it's ok
>>   * is the value one created by ALTER TYPE ADD VALUE? Test
>> blacklist. If no, it's ok.
>>   * is the enum a new one? Test whitelist. If yes, it's ok.
>>   * anything else is not ok.
> My point is that if you do 1 and 3, you don't need 2.  Or if you do
> 2 and 3, you don't need 1.  But in most cases, testing the tuple
> hint bits is cheap, so you don't really want that option.
>
> In any case, what I'm worried about is the amount of bookkeeping
> overhead added by keeping a whitelist of enum-types-created-in-
> current-transaction.  That's less than trivial, especially since
> you have to account correctly for subtransactions.  And there are
> common use-cases where that table will become large.
>
>> If we just did the blacklist and stuck with our current heuristic test
>> for enum being created in the current transaction, we'd still probably
>> avoid 99% of the problems, including specifically the one that gave rise
>> to the bug report.
> True.  But I'm not sure whether the heuristic test is adding anything
> meaningful if we use a blacklist first.  The case where it could help
> is
>
>   begin;
>   create type t as enum();
>   alter type t add value 'v';
>   -- do something with 'v'
>   commit;
>
> That perhaps is worth something, but if somebody is trying to build a new
> enum type in pieces like that, doesn't it seem fairly likely that they
> might throw in an ALTER OWNER or GRANT as well?  My feeling is that the
> lesson we need to learn is that the heuristic test isn't good enough.
>
>   


OK, I think I'm convinced. Here's is the WIP code I put together for the
blacklist. I'm was looking for a place to put the init call, but since
it's possibly not going anywhere I stopped :-) . My initial thought
about substransactions was that we should ignore them for this purpose
(That's why I used TopTransactionContext for the table).

I agree the heuristic test isn't good enough, and if we can get a 100%
accurate test for the newness of the enum type then the blacklist would
be redundant.

w.r.t. table size - how large? I confess I haven't seen any systems with
more than a few hundred enum types. But even a million or two shouldn't
consume a huge amount of memory, should it?

cheers

andrew

-- 

Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

diff --git a/src/backend/catalog/pg_enum.c b/src/backend/catalog/pg_enum.c
index fe61d4d..52c1271 100644
--- a/src/backend/catalog/pg_enum.c
+++ b/src/backend/catalog/pg_enum.c
@@ -28,6 +28,8 @@
 #include "utils/builtins.h"
 #include "utils/catcache.h"
 #include "utils/fmgroids.h"
+#include "utils/hsearch.h"
+#include "utils/memutils.h"
 #include "utils/syscache.h"
 #include "utils/tqual.h"
 
@@ -38,6 +40,9 @@ Oid			binary_upgrade_next_pg_enum_oid = InvalidOid;
 static void RenumberEnumType(Relation pg_enum, HeapTuple *existing, int nelems);
 static int	sort_order_cmp(const void *p1, const void *p2);
 
+/* hash table of values added in current transaction by AddEnumLabel */
+
+static HTAB *enum_blacklist = NULL;
 
 /*
  * EnumValuesCreate
@@ -460,8 +465,44 @@ restart:
 	heap_freetuple(enum_tup);
 
 	heap_close(pg_enum, RowExclusiveLock);
+
+	/* set up blacklist hash if required */
+	if (enum_blacklist == NULL)
+	{
+		HASHCTL hash_ctl;
+		memset(_ctl, 0, sizeof(hash_ctl));
+		hash_ctl.keysize = sizeof(Oid);
+		hash_ctl.entrysize = sizeof(Oid);
+		hash_ctl.hcxt = CurTransactionContext;
+		enum_blacklist = hash_create("Enum blacklist for current transaction",
+		   32,
+		   _ctl,
+		   HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
+
+	}
+
+	/* and add the new value to the blacklist */
+
+	(void) hash_search(enum_blacklist, , HASH_ENTER, NULL);
 }
 
+bool
+EnumBlacklisted(Oid enum_id)
+{
+	bool found;
+
+	if (enum_blacklist == NULL)
+		return false;
+
+	(void) hash_search(enum_blacklist, _id, HASH_FIND, );
+	return found;
+}
+
+void
+InitEnumBlacklist(void)
+{
+	enum_blacklist = NULL;
+}
 
 /*
  * RenameEnumLabel
diff --git a/src/backend/utils/adt/enum.c b/src/backend/utils/adt/enum.c
index 973397c..a7ba3d0 100644
--- a/src/backend/utils/adt/enum.c
+++ b/src/backend/utils/adt/enum.c
@@ -76,6 +76,10 @@ check_safe_enum_use(HeapTuple enumval_tup)
 		TransactionIdDidCommit(xmin))

Re: [HACKERS] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-23 Thread Tom Lane 
Andrew Dunstan  writes:
> On 09/23/2017 02:00 PM, Tom Lane wrote:
>> So I'm back to not being sure about the path forward.  Maybe it would be
>> all right to say "the value added by ADD VALUE can't be used in the same
>> transaction, period".  That's still a step forward compared to the pre-v10
>> prohibition on doing it at all.  I don't remember if there were use-cases
>> where we really needed the exception for new-in-transaction types.

> Well, my idea was to have the test run like this:

>   * is the value an old one? Test txnid of tuple. If yes it's ok
>   * is the value one created by ALTER TYPE ADD VALUE? Test
> blacklist. If no, it's ok.
>   * is the enum a new one? Test whitelist. If yes, it's ok.
>   * anything else is not ok.

My point is that if you do 1 and 3, you don't need 2.  Or if you do
2 and 3, you don't need 1.  But in most cases, testing the tuple
hint bits is cheap, so you don't really want that option.

In any case, what I'm worried about is the amount of bookkeeping
overhead added by keeping a whitelist of enum-types-created-in-
current-transaction.  That's less than trivial, especially since
you have to account correctly for subtransactions.  And there are
common use-cases where that table will become large.

> If we just did the blacklist and stuck with our current heuristic test
> for enum being created in the current transaction, we'd still probably
> avoid 99% of the problems, including specifically the one that gave rise
> to the bug report.

True.  But I'm not sure whether the heuristic test is adding anything
meaningful if we use a blacklist first.  The case where it could help
is

begin;
create type t as enum();
alter type t add value 'v';
-- do something with 'v'
commit;

That perhaps is worth something, but if somebody is trying to build a new
enum type in pieces like that, doesn't it seem fairly likely that they
might throw in an ALTER OWNER or GRANT as well?  My feeling is that the
lesson we need to learn is that the heuristic test isn't good enough.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-23 Thread Andrew Dunstan 


On 09/23/2017 02:00 PM, Tom Lane wrote:
> I wrote:
>> Andrew Dunstan  writes:
>>> I see what you're saying, but my idea was slightly different. We would
>>> only add to the hashtable I had in mind at the bottom AddEnumLabel().
>>> Any other value, whether added in the current transaction or not, should
>>> be safe, AIUI.
>> Oh, I see: a list of enum values we need to blacklist, not whitelist.
>> Yes, that's a significantly better idea than mine, because in common
>> use-cases that would be empty or have a very small number of entries.
> Oh, wait a minute ... that's not where the problem is, really.  We can
> already tell reliably whether an enum value was created in the current
> transaction: the is-it-committed check in check_safe_enum_use is
> perfectly safe for that AFAICS.  The hard part of this is the part about
> "was the enum type created in the current transaction?".  We could make
> that reliable with the table I proposed of enum types created in the
> current transaction, but the possible performance drag is a concern.
>
> What I understand you to be proposing is to blacklist individual
> enum values added by ALTER ADD VALUE, but *not* values created
> aboriginally by CREATE TYPE AS ENUM.  (The latter are surely safe,
> because the type must disappear if they do.)  However, that would
> require dropping the second part of the current documentation promise:
>
>If ALTER TYPE ... ADD VALUE (the form that adds a new value to
>an enum type) is executed inside a transaction block, the new value cannot
>be used until after the transaction has been committed, except in the case
>that the enum type itself was created earlier in the same transaction.
>
> We'd have to just say "it can't be used till the transaction has been
> committed", full stop.  Otherwise we're right back into the problem of
> needing to detect whether the enum type is new in transaction.
>
>>> Maybe we should also keep a cache of whitelisted enums
>>> created in the current transaction.
>> What for?
> I now realize that what you probably meant here was to track enum *types*,
> not values, created in the current transaction.  But if we're doing that
> then we don't really need the per-enum-value-blacklist part of it.
>
> So I'm back to not being sure about the path forward.  Maybe it would be
> all right to say "the value added by ADD VALUE can't be used in the same
> transaction, period".  That's still a step forward compared to the pre-v10
> prohibition on doing it at all.  I don't remember if there were use-cases
> where we really needed the exception for new-in-transaction types.
>
>   



Well, my idea was to have the test run like this:

  * is the value an old one? Test txnid of tuple. If yes it's ok
  * is the value one created by ALTER TYPE ADD VALUE? Test
blacklist. If no, it's ok.
  * is the enum a new one? Test whitelist. If yes, it's ok.
  * anything else is not ok.

I think that would let us keep our promise.

If we just did the blacklist and stuck with our current heuristic test
for enum being created in the current transaction, we'd still probably
avoid 99% of the problems, including specifically the one that gave rise
to the bug report.

Am I missing something?


cheers


andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-23 Thread Tom Lane 
I wrote:
> Andrew Dunstan  writes:
>> I see what you're saying, but my idea was slightly different. We would
>> only add to the hashtable I had in mind at the bottom AddEnumLabel().
>> Any other value, whether added in the current transaction or not, should
>> be safe, AIUI.

> Oh, I see: a list of enum values we need to blacklist, not whitelist.
> Yes, that's a significantly better idea than mine, because in common
> use-cases that would be empty or have a very small number of entries.

Oh, wait a minute ... that's not where the problem is, really.  We can
already tell reliably whether an enum value was created in the current
transaction: the is-it-committed check in check_safe_enum_use is
perfectly safe for that AFAICS.  The hard part of this is the part about
"was the enum type created in the current transaction?".  We could make
that reliable with the table I proposed of enum types created in the
current transaction, but the possible performance drag is a concern.

What I understand you to be proposing is to blacklist individual
enum values added by ALTER ADD VALUE, but *not* values created
aboriginally by CREATE TYPE AS ENUM.  (The latter are surely safe,
because the type must disappear if they do.)  However, that would
require dropping the second part of the current documentation promise:

   If ALTER TYPE ... ADD VALUE (the form that adds a new value to
   an enum type) is executed inside a transaction block, the new value cannot
   be used until after the transaction has been committed, except in the case
   that the enum type itself was created earlier in the same transaction.

We'd have to just say "it can't be used till the transaction has been
committed", full stop.  Otherwise we're right back into the problem of
needing to detect whether the enum type is new in transaction.

>> Maybe we should also keep a cache of whitelisted enums
>> created in the current transaction.

> What for?

I now realize that what you probably meant here was to track enum *types*,
not values, created in the current transaction.  But if we're doing that
then we don't really need the per-enum-value-blacklist part of it.

So I'm back to not being sure about the path forward.  Maybe it would be
all right to say "the value added by ADD VALUE can't be used in the same
transaction, period".  That's still a step forward compared to the pre-v10
prohibition on doing it at all.  I don't remember if there were use-cases
where we really needed the exception for new-in-transaction types.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-23 Thread Andrew Dunstan 


On 09/23/2017 11:16 AM, Tom Lane wrote:
> Andrew Dunstan  writes:
>
>>> The immediate question is do we care to design/implement such a thing
>>> post-RC1.  I'd have to vote "no".  I think the most prudent thing to
>>> do is revert 15bc038f9 and then have another go at it during the v11
>>> cycle.
>> Sadly I agree. We've made decisions like this in the past, and I have
>> generally been supportive of them. I think this is the first time I have
>> been on the receiving end of one so late in the process :-(
> Unless you want to try writing a patch for this in the next day or two,
> I think we have to do that.  But now that I see the plan clearly, maybe
> we could get away with a post-RC1 fix.


OK, I'll give it a shot.

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-23 Thread Tom Lane 
Andrew Dunstan  writes:
> On 09/22/2017 11:19 PM, Tom Lane wrote:
>> Yeah, I was considering the same thing over dinner, though I'd phrase
>> it oppositely: keep a list of enum type OIDs created in the current
>> transaction, so that we could whitelist them.  This could maybe become
>> a problem if someone created a zillion enums in one xact, though.

> I see what you're saying, but my idea was slightly different. We would
> only add to the hashtable I had in mind at the bottom AddEnumLabel().
> Any other value, whether added in the current transaction or not, should
> be safe, AIUI.

Oh, I see: a list of enum values we need to blacklist, not whitelist.
Yes, that's a significantly better idea than mine, because in common
use-cases that would be empty or have a very small number of entries.
In particular that fixes the "pg_restore -1" scenario, because no
matter how many enums you're restoring, pg_dump doesn't use ALTER
TYPE ADD VALUE.  (Well, it does in --binary-upgrade mode, but those
scripts are run in transaction-per-statement mode so it's fine.)

> Maybe we should also keep a cache of whitelisted enums
> created in the current transaction.

What for?  Wouldn't be any faster to search, in fact likely slower
because it could get large in common use-cases.

>> The immediate question is do we care to design/implement such a thing
>> post-RC1.  I'd have to vote "no".  I think the most prudent thing to
>> do is revert 15bc038f9 and then have another go at it during the v11
>> cycle.

> Sadly I agree. We've made decisions like this in the past, and I have
> generally been supportive of them. I think this is the first time I have
> been on the receiving end of one so late in the process :-(

Unless you want to try writing a patch for this in the next day or two,
I think we have to do that.  But now that I see the plan clearly, maybe
we could get away with a post-RC1 fix.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-23 Thread Andrew Dunstan 


On 09/22/2017 11:19 PM, Tom Lane wrote:
> Andrew Dunstan  writes:
>> On 09/22/2017 05:46 PM, Tom Lane wrote:
>>> I'm not sure if that qualifies as a stop-ship problem, but it ain't
>>> good, for sure.  We need to look at whether we should revert 15bc038f9
>>> or somehow revise its rules.
>> I wonder if we wouldn't be better
>> doing this more directly, keeping a per-transaction hash of unsafe enum
>> values (which will almost always be empty). It might even speed up the
>> check.
> Yeah, I was considering the same thing over dinner, though I'd phrase
> it oppositely: keep a list of enum type OIDs created in the current
> transaction, so that we could whitelist them.  This could maybe become
> a problem if someone created a zillion enums in one xact, though.


I see what you're saying, but my idea was slightly different. We would
only add to the hashtable I had in mind at the bottom AddEnumLabel().
Any other value, whether added in the current transaction or not, should
be safe, AIUI. Maybe we should also keep a cache of whitelisted enums
created in the current transaction.

I'm not to worried about people creating a zillion enums (or enum labels
being added for the solution I had in mind). Even a hash of a million
Oids will only consume a few megabytes, won't it?

>
> The immediate question is do we care to design/implement such a thing
> post-RC1.  I'd have to vote "no".  I think the most prudent thing to
> do is revert 15bc038f9 and then have another go at it during the v11
> cycle.
>
>   


Sadly I agree. We've made decisions like this in the past, and I have
generally been supportive of them. I think this is the first time I have
been on the receiving end of one so late in the process :-(

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-22 Thread Tom Lane 
Andrew Dunstan  writes:
> On 09/22/2017 05:46 PM, Tom Lane wrote:
>> I'm not sure if that qualifies as a stop-ship problem, but it ain't
>> good, for sure.  We need to look at whether we should revert 15bc038f9
>> or somehow revise its rules.

> I wonder if we wouldn't be better
> doing this more directly, keeping a per-transaction hash of unsafe enum
> values (which will almost always be empty). It might even speed up the
> check.

Yeah, I was considering the same thing over dinner, though I'd phrase
it oppositely: keep a list of enum type OIDs created in the current
transaction, so that we could whitelist them.  This could maybe become
a problem if someone created a zillion enums in one xact, though.

The immediate question is do we care to design/implement such a thing
post-RC1.  I'd have to vote "no".  I think the most prudent thing to
do is revert 15bc038f9 and then have another go at it during the v11
cycle.

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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-22 Thread Andrew Dunstan 


On 09/22/2017 05:46 PM, Tom Lane wrote:
> bal...@obiserver.hu writes:
>> PostgreSQL version: 10beta4
>> testdb=# begin;
>> BEGIN
>> testdb=# create type enumtype as enum ('v1', 'v2');
>> CREATE TYPE
>> testdb=# alter type enumtype owner to testrole;
>> ALTER TYPE
>> testdb=# create table testtable (enumcolumn enumtype not null default 'v1');
>> ERROR:  unsafe use of new value "v1" of enum type enumtype
>> HINT:  New enum values must be committed before they can be used.
> Hmm, that's pretty annoying.  It's happening be

> cause check_safe_enum_use
> insists that the enum's pg_type entry not be updated-in-transaction.
> We thought that the new rules instituted by 15bc038f9 would be generally
> more convenient to use than the old ones --- but this example shows
> that, for example, pg_dump scripts involving enums often could not
> be restored inside a single transaction.
>
> I'm not sure if that qualifies as a stop-ship problem, but it ain't
> good, for sure.  We need to look at whether we should revert 15bc038f9
> or somehow revise its rules.



:-(


The only real problem comes from adding a value to an enum that has been
created in an earlier transaction and then using that enum value. What
we're doing here is essentially a heuristic test for that condition, and
we're getting some false positives. I wonder if we wouldn't be better
doing this more directly, keeping a per-transaction hash of unsafe enum
values (which will almost always be empty). It might even speed up the
check.

cheers

andrew

-- 
Andrew Dunstanhttps://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, 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] [BUGS] BUG #14825: enum type: unsafe use?

2017-09-22 Thread Tom Lane 
bal...@obiserver.hu writes:
> PostgreSQL version: 10beta4

> testdb=# begin;
> BEGIN
> testdb=# create type enumtype as enum ('v1', 'v2');
> CREATE TYPE
> testdb=# alter type enumtype owner to testrole;
> ALTER TYPE
> testdb=# create table testtable (enumcolumn enumtype not null default 'v1');
> ERROR:  unsafe use of new value "v1" of enum type enumtype
> HINT:  New enum values must be committed before they can be used.

Hmm, that's pretty annoying.  It's happening because check_safe_enum_use
insists that the enum's pg_type entry not be updated-in-transaction.
We thought that the new rules instituted by 15bc038f9 would be generally
more convenient to use than the old ones --- but this example shows
that, for example, pg_dump scripts involving enums often could not
be restored inside a single transaction.

I'm not sure if that qualifies as a stop-ship problem, but it ain't
good, for sure.  We need to look at whether we should revert 15bc038f9
or somehow revise its rules.

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