Horton

Re: snapbuild woes

Andres Freund <andres@anarazel.de>

View thread View on postgr.es
From: Andres Freund <andres@anarazel.de>
To: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Cc: PostgreSQL-development <pgsql-hackers@postgresql.org>, Erik Rijkers <er@xs4all.nl>, Robert Haas <robertmhaas@gmail.com>, Craig Ringer <craig@2ndquadrant.com>
Date: 2017-05-01T08:03:48Z
Lists: pgsql-hackers 
On 2017-05-01 03:54:49 +0200, Petr Jelinek wrote:
> I agree with adding running, I think that's good thing even for the per
> transaction tracking and snapshot exports - we could use the newly added
> field to get rid of the issue we have with 'snapshot too large' when
> there were many aborted transactions while we waited for running ones to
> finish.

I'm not sure of that - what I was proposing would only track this for
the ->running substructure.  How'd that help?


> But, I still think we need to restart the tracking after new
> xl_running_xacts. Reason for that is afaics any of the catalog snapshots
> that we assigned to transactions at the end of SnapBuildCommitTxn might
> be corrupted otherwise as they were built before we knew one of the
> supposedly running txes was actually already committed and that
> transaction might have done catalog changes.

I'm afraid you're right.  But I think this is even more complicated: The
argument in your version that this can only happen once, seems to also
be holey: Just imagine a pg_usleep(3000 * 1000000) right before
ProcArrayEndTransaction() and enjoy the picture.

Wonder if we should just (re-)add a stage between SNAPBUILD_START and
SNAPBUILD_FULL_SNAPSHOT.  Enter SNAPBUILD_BUILD_INITIAL_SNAPSHOT at the
first xl_running_xacts, wait for all transactions to end with my
approach, while populating SnapBuild->committed, only then start
collecting changes for transactions (i.e. return true from
SnapBuildProcessChange()), return true once all xacts have finished
again.  That'd presumably be a bit easier to understand, more robust -
and slower.

- Andres

Commits

  1. Fix thinko introduced in 2bef06d516460 et al.

  2. Avoid superfluous work for commits during logical slot creation.

  3. Fix race condition leading to hanging logical slot creation.

  4. Don't use on-disk snapshots for exported logical decoding snapshot.

  5. Preserve required !catalog tuples while computing initial decoding snapshot.