Horton

Re: Parallel scan with SubTransGetTopmostTransaction assert coredump

Robert Haas <robertmhaas@gmail.com>

View thread View on postgr.es
From: Robert Haas <robertmhaas@gmail.com>
To: Greg Nancarrow <gregn4422@gmail.com>
Cc: Pavel Borisov <pashkin.elfe@gmail.com>, Tomas Vondra <tomas.vondra@enterprisedb.com>, Maxim Orlov <m.orlov@postgrespro.ru>, Michael Paquier <michael@paquier.xyz>, Pengchengliu <pengchengliu@tju.edu.cn>, Andres Freund <andres@anarazel.de>, PostgreSQL-development <pgsql-hackers@postgresql.org>
Date: 2021-08-11T19:37:21Z
Lists: pgsql-hackers 
On Wed, Aug 11, 2021 at 8:32 AM Greg Nancarrow <gregn4422@gmail.com> wrote:
> This is explained by the TransactionSnapshot being a later snapshot in
> this case.
> So this is why it seems to be wrong to call GetTransactionSnapshot()
> in InitializeParallelDSM() and use a separate, potentially later,
> snapshot than that used in the execution state for the query.

Thanks for the research. I agree with your logic here, but:

1. Then why doesn't the approach I proposed fix it?

2. Consider the case where the toplevel query is something like SELECT
complexfunc() FROM generate_series(1,10) g -- in a case like this, I
think complexfunc() can cause snapshots to be taken internally. For
example suppose we end up inside exec_eval_simple_expr, or
SPI_cursor_open_internal, in either case with read_only = false. Here
we're going to again call GetTransactionSnapshot() and then execute a
query which may use parallelism.


-- 
Robert Haas
EDB: http://www.enterprisedb.com

Commits

  1. Fix broken snapshot handling in parallel workers.

  2. Doc: move some catalogs.sgml entries to the right place.

  3. Stamp 13.2.

  4. Create an infrastructure for parallel computation in PostgreSQL.