Re: Failed to delete old ReorderBuffer spilled files

On 24 November 2017 at 09:20, atorikoshi <torikoshi_atsushi_z2(at)lab(dot)ntt(dot)co(dot)jp
> wrote:

>
> Thanks for letting me know.
> I think I only tested running "make check" at top directory, sorry
> for my insufficient test.
>
> The test failure happened at the beginning of replication(creating
> slot), so there are no changes yet and getting the tail of changes
> failed.
>
> Because the bug we are fixing only happens after creating files,
> I've added "txn->serialized" to the if statement condition.

Thanks.

Re-reading the patch I see

* The final_lsn of which transaction that hasn't produced an abort
* record is invalid.

which I find very hard to parse. I suggest:

We set final_lsn when we decode the commit or abort record for a
transaction,
but transactions implicitly aborted by a crash never write such a record.

then continue from there with the same text as in the patch.

Otherwise I'm happy. It passes make check, test decoding and the recovery
TAP tests too.

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