Horton

Re: Re: PANIC: invalid index offnum: 186 when processing BRIN indexes in VACUUM

Alvaro Herrera <alvherre@alvh.no-ip.org>

View thread View on postgr.es
From: Alvaro Herrera <alvherre@alvh.no-ip.org>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Tomas Vondra <tomas.vondra@2ndquadrant.com>, Michael Paquier <michael.paquier@gmail.com>, PostgreSQL mailing lists <pgsql-hackers@postgresql.org>
Date: 2017-11-02T19:11:29Z
Lists: pgsql-hackers 
Tom Lane wrote:
> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:

> > 2. when summarization is requested on the partial range at the end of a
> > table, we acquire extension lock on the rel, then compute relation size
> > and run summarization with the lock held.  This guarantees that we don't
> > miss any pages.  This is bad for concurrency though, so it's only done
> > in that specific scenario.
> 
> Hm, I wonder how this will play with the active proposals around
> reimplementing relation extension locks.  All that work seems to be
> assuming that the extension lock is only held for a short time and
> nothing much beyond physical extension is done while holding it.
> I'm afraid that you may be introducing a risk of e.g. deadlocks
> if you do this.

Ouch ... yeah, that could be a problem.

Another idea I had was to just insert the placeholder tuple while
holding the extension lock, then release the lock while the
summarization is done.  It would be a bit of a break of the current
separation of concerns, but I'm not convinced that the current setup is
perfect, so maybe that's okay.

> If VACUUM and brin_summarize_new_values both ignore the partial
> range, then what else would request this?  Can't we just decree
> that we don't summarize the partial range, period?

brin_summarize_range() can do it.

-- 
Álvaro Herrera                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

Commits

  1. Fix BRIN summarization concurrent with extension

  2. Fix corner-case errors in brin_doupdate().

  3. Rewrite PageIndexDeleteNoCompact into a form that only deletes 1 tuple.

  4. Invent PageIndexTupleOverwrite, and teach BRIN and GiST to use it.

  5. Close some holes in BRIN page assignment