Re: Performance degradation on concurrent COPY into a single relation in PG16.

Andres Freund <andres@anarazel.de>

From: Andres Freund <andres@anarazel.de>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Heikki Linnakangas <hlinnaka@iki.fi>, Alvaro Herrera <alvherre@alvh.no-ip.org>, Masahiko Sawada <sawada.mshk@gmail.com>, David Rowley <dgrowleyml@gmail.com>, Dean Rasheed <dean.a.rasheed@gmail.com>, Jakub Wartak <jakub.wartak@enterprisedb.com>, Andrew Dunstan <andrew@dunslane.net>, Peter Eisentraut <peter@eisentraut.org>, PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Date: 2023-10-12T16:24:19Z
Lists: pgsql-hackers
Hi,

On 2023-10-12 11:44:09 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> >> On 2023-09-25 15:42:26 -0400, Tom Lane wrote:
> >>> I just did a git bisect run to discover when the failure documented
> >>> in bug #18130 [1] started.  And the answer is commit 82a4edabd.
> 
> > Uh, huh.  The problem is that COPY uses a single BulkInsertState for multiple
> > partitions. Which to me seems to run counter to the following comment:
> >  *	The caller can also provide a BulkInsertState object to optimize many
> >  *	insertions into the same relation.  This keeps a pin on the current
> >  *	insertion target page (to save pin/unpin cycles) and also passes a
> >  *	BULKWRITE buffer selection strategy object to the buffer manager.
> >  *	Passing NULL for bistate selects the default behavior.
> 
> > The reason this doesn't cause straight up corruption due to reusing a pin from
> > another relation is that b1ecb9b3fcfb added ReleaseBulkInsertStatePin() and a
> > call to it. But I didn't make ReleaseBulkInsertStatePin() reset the bulk
> > insertion state, which is what leads to the errors from the bug report.
> 
> > Resetting the relevant BulkInsertState fields fixes the problem. But I'm not
> > sure that's the right fix. ISTM that independent of whether we fix this via
> > ReleaseBulkInsertStatePin() resetting the fields or via not reusing
> > BulkInsertState, we should add assertions defending against future issues like
> > this (e.g. by adding a relation field to BulkInsertState in cassert builds,
> > and asserting that the relation is the same as in prior calls unless
> > ReleaseBulkInsertStatePin() has been called).
> 
> Ping?  We really ought to have a fix for this committed in time for
> 16.1.

I kind of had hoped somebody would comment on the approach.  Given that nobody
has, I'll push the minimal fix of resetting the state in
ReleaseBulkInsertStatePin(), even though I think architecturally that's not
great.

Greetings,

Andres Freund



Commits

  1. Avoid edge case in pg_visibility test with small shared_buffers

  2. Fix bulk table extension when copying into multiple partitions

  3. hio: Take number of prior relation extensions into account

  4. Fix performance regression in pg_strtointNN_safe functions

  5. Fix performance problem with new COPY DEFAULT code

  6. hio: Use ExtendBufferedRelBy() to extend tables more efficiently

  7. Add VACUUM/ANALYZE BUFFER_USAGE_LIMIT option