Thread: Improve warnings around CREATE INDEX CONCURRENTLY

Improve warnings around CREATE INDEX CONCURRENTLY

From
Greg Smith
Date:
When running CREATE INDEX CONCURRENTLY, the DefineIndex() code in
src/backend/commands/indexcmds.c does a few things that one would expect
from the documentation.  And then at the end it executes code described
like this:

"The index is now valid in the sense that it contains all currently
interesting tuples.  But since it might not contain tuples deleted just
before the reference snap was taken, we have to wait out any
transactions that might have older snapshots.  Obtain a list of VXIDs of
such transactions, and wait for them individually."

It's possible to end up with a long series in pg_locks waiting for
virtualxid entries at this point, for as long as some set of giant
queries takes to execute, and you'll only see them one at a time.  The
documentation warns:

"PostgreSQL must perform two scans of the table, and in addition it must
wait for all existing transactions that could potentially use the index
to terminate."

That's correct, but easy to read the wrong way.  I always assumed that
this meant it was going to wait behind anything that had a shared lock
or such on the table, things that had already accessed it.  This is the
case with some earlier parts of this same code path.  But when it comes
to the end here, the scope is actually broader than that.  And since
there's a session-level lock on the table the whole time this wait loop
is executing, that makes considerable secondary havoc possible here.
You can end up waiting an unbounded amount of time for some long-running
transaction, one not even expected to enter into the rebuild work, to
finish, which leaves you no idea what's happening unless you know just
what to look for.  (Watching such havoc actually happen is what prompted
this investigation)

What makes it worse is that the wait shows up as a virtualxid one, which
doesn't pop up on many common samples of things to look for in
pg_locks.  It would be reasonable but also incorrect for admins to
assume a table one would be visible if running into the case alluded to
in the docs.  The serial way locks are obtained is unexpected too.

Attached patch expands the warnings around this command to reflect both
issues:

-The waiting time is not necessarily limited to just things that involve
the table
-The locks it obtains while running this phase of the rebuild are unusual

--
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us



Attachment

Re: Improve warnings around CREATE INDEX CONCURRENTLY

From
Alvaro Herrera
Date:
Excerpts from Greg Smith's message of mar may 24 03:56:59 -0400 2011:

> What makes it worse is that the wait shows up as a virtualxid one, which
> doesn't pop up on many common samples of things to look for in
> pg_locks.  It would be reasonable but also incorrect for admins to
> assume a table one would be visible if running into the case alluded to
> in the docs.  The serial way locks are obtained is unexpected too.

Incidentally, this is one of the things that Jim Nasby wanted exposed
somehow so that these problems are more easily diagnosed.  I dropped the
ball on that one, but I'll be picking it up again at some point.

--
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support

Re: Improve warnings around CREATE INDEX CONCURRENTLY

From
Greg Smith
Date:
On 05/24/2011 04:48 PM, Alvaro Herrera wrote:
> Excerpts from Greg Smith's message of mar may 24 03:56:59 -0400 2011:
>
>
>> What makes it worse is that the wait shows up as a virtualxid one, which
>> doesn't pop up on many common samples of things to look for in
>> pg_locks.  It would be reasonable but also incorrect for admins to
>> assume a table one would be visible if running into the case alluded to
>> in the docs.  The serial way locks are obtained is unexpected too.
>>
> Incidentally, this is one of the things that Jim Nasby wanted exposed
> somehow so that these problems are more easily diagnosed.  I dropped the
> ball on that one, but I'll be picking it up again at some point.
>

I don't remember seeing anything about that before, but then again I
wasn't really looking for it until now.  Did you have a basic idea what
you wanted to do there?  I have enough of this work queued up now that I
may end up poking at the code myself here soon.

The first thing I was considering was dropping some DEBUG2 level logging
about the various phases of the reindex into there.  Given that
concurrent index creation has all these single instance requirements and
long runtimes, I find myself putting them into scripts that do all the
work as background processes.  If I could set client_min_messages=debug2
when starting the script, and see more info about the progress as it
happens appear in the script's output log, that is something I think
many admins would choose to do.

Not the ideal UI for exposing this info, certainly.  But a really easy
one to add, and realistically I think it would be enough to resolve most
of the transparency complaints here.  The biggest problem is not even
documenting where people should be looking toward suspiciously, which I
think the doc patch I submitted helps with.

--
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us



Re: Improve warnings around CREATE INDEX CONCURRENTLY

From
Alvaro Herrera
Date:
Excerpts from Greg Smith's message of mar may 24 17:06:07 -0400 2011:
> On 05/24/2011 04:48 PM, Alvaro Herrera wrote:

> > Incidentally, this is one of the things that Jim Nasby wanted exposed
> > somehow so that these problems are more easily diagnosed.  I dropped the
> > ball on that one, but I'll be picking it up again at some point.
>
> I don't remember seeing anything about that before, but then again I
> wasn't really looking for it until now.  Did you have a basic idea what
> you wanted to do there?  I have enough of this work queued up now that I
> may end up poking at the code myself here soon.

Please see this thread:
http://archives.postgresql.org/message-id/1290465663-sup-9908@alvh.no-ip.org

> The first thing I was considering was dropping some DEBUG2 level logging
> about the various phases of the reindex into there.

That would take this system more or less to where autovacuum was in 8.1
(well, not quite that bad because you can change it locally, but still).
It seems to me we can do something a bit better than that.

--
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support

Re: Improve warnings around CREATE INDEX CONCURRENTLY

From
Simon Riggs
Date:
On Wed, May 25, 2011 at 6:57 PM, Alvaro Herrera
<alvherre@commandprompt.com> wrote:
> Excerpts from Greg Smith's message of mar may 24 17:06:07 -0400 2011:
>> On 05/24/2011 04:48 PM, Alvaro Herrera wrote:
>
>> > Incidentally, this is one of the things that Jim Nasby wanted exposed
>> > somehow so that these problems are more easily diagnosed.  I dropped the
>> > ball on that one, but I'll be picking it up again at some point.
>>
>> I don't remember seeing anything about that before, but then again I
>> wasn't really looking for it until now.  Did you have a basic idea what
>> you wanted to do there?  I have enough of this work queued up now that I
>> may end up poking at the code myself here soon.
>
> Please see this thread:
> http://archives.postgresql.org/message-id/1290465663-sup-9908@alvh.no-ip.org
>
>> The first thing I was considering was dropping some DEBUG2 level logging
>> about the various phases of the reindex into there.
>
> That would take this system more or less to where autovacuum was in 8.1
> (well, not quite that bad because you can change it locally, but still).
> It seems to me we can do something a bit better than that.

It would be fairly straightforward to add some text to the end of the
ps message for long running DDL commands to say "phase 2" etc..
Or pg_stat_activity query text.

That way we would have some visibility into the progress of long
running DDL without too much effort.

--
 Simon Riggs                   http://www.2ndQuadrant.com/
 PostgreSQL Development, 24x7 Support, Training & Services

Re: Improve warnings around CREATE INDEX CONCURRENTLY

From
Greg Smith
Date:
On 05/25/2011 01:57 PM, Alvaro Herrera wrote:
> Please see this thread:
> http://archives.postgresql.org/message-id/1290465663-sup-9908@alvh.no-ip.org
>

Ah, now I see--I was running a CommitFest that week, so of course I
missed this message.  Those are all good ideas.

>> The first thing I was considering was dropping some DEBUG2 level logging
>> about the various phases of the reindex into there.
>>
> That would take this system more or less to where autovacuum was in 8.1
> (well, not quite that bad because you can change it locally, but still).
> It seems to me we can do something a bit better than that.
>

Sure.  Simon's command string idea might work better, and doing some
extra lock decoration as you suggested in the above thread would be
another level of improvement.  We should pick up redesign later on the
main list.  You can at least count me in as someone who wants to see
this improved now.

Back to the doc patch I submitted...is that a useful step toward making
this issue visible enough to users for now to help?

--
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us



Re: Improve warnings around CREATE INDEX CONCURRENTLY

From
Alvaro Herrera
Date:
Excerpts from Greg Smith's message of mié may 25 17:04:03 -0400 2011:

> Sure.  Simon's command string idea might work better, and doing some
> extra lock decoration as you suggested in the above thread would be
> another level of improvement.  We should pick up redesign later on the
> main list.  You can at least count me in as someone who wants to see
> this improved now.

Great

> Back to the doc patch I submitted...is that a useful step toward making
> this issue visible enough to users for now to help?

Sure, why not?  I thought I could choose my bikeshed color while I was
here, how about

+    second and third transaction.  All active transactions at the time the
+    second table scan starts, not just ones that already involve the table,
+    have the potential to block the concurrent index creation until they
+    finish.  When checking for transactions that
+    could still use the original index, concurrent index creation advances
+    through potentially interfering older transactions one at a time,
+    obtaining shared locks on their virtual transaction identifiers to wait for
+    them to complete.


--
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support

Re: Improve warnings around CREATE INDEX CONCURRENTLY

From
Greg Smith
Date:
On 05/26/2011 11:52 AM, Alvaro Herrera wrote:
> Sure, why not?  I thought I could choose my bikeshed color while I was
> here, how about
>

That text is fine too.  I don't care exactly what color the bike shed
is, I just want to get the bike out of the rain.

--
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD



Re: Improve warnings around CREATE INDEX CONCURRENTLY

From
Robert Haas
Date:
On Thu, May 26, 2011 at 11:52 AM, Alvaro Herrera
<alvherre@commandprompt.com> wrote:
> Excerpts from Greg Smith's message of mié may 25 17:04:03 -0400 2011:
>
>> Sure.  Simon's command string idea might work better, and doing some
>> extra lock decoration as you suggested in the above thread would be
>> another level of improvement.  We should pick up redesign later on the
>> main list.  You can at least count me in as someone who wants to see
>> this improved now.
>
> Great
>
>> Back to the doc patch I submitted...is that a useful step toward making
>> this issue visible enough to users for now to help?
>
> Sure, why not?  I thought I could choose my bikeshed color while I was
> here, how about
>
> +    second and third transaction.  All active transactions at the time the
> +    second table scan starts, not just ones that already involve the table,
> +    have the potential to block the concurrent index creation until they
> +    finish.  When checking for transactions that
> +    could still use the original index, concurrent index creation advances
> +    through potentially interfering older transactions one at a time,
> +    obtaining shared locks on their virtual transaction identifiers to wait for
> +    them to complete.

Alvaro, did you commit this?

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Re: Improve warnings around CREATE INDEX CONCURRENTLY

From
Alvaro Herrera
Date:
Excerpts from Robert Haas's message of lun jun 13 12:11:48 -0400 2011:
> On Thu, May 26, 2011 at 11:52 AM, Alvaro Herrera
> <alvherre@commandprompt.com> wrote:

> >> Back to the doc patch I submitted...is that a useful step toward making
> >> this issue visible enough to users for now to help?

> Alvaro, did you commit this?

I just did (9.0 and beyond), sorry for sitting on it for so long.

--
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support