Re: PATCH: logical_work_mem and logical streaming of large in-progress transactions

On Wed, Mar 4, 2020 at 3:16 AM Tomas Vondra
<tomas(dot)vondra(at)2ndquadrant(dot)com> wrote:
>
> Hi,
>
> I started looking at this patch series again, hoping to get it moving
> for PG13.
>

It is good to keep moving this forward, but there are quite a few
problems with the design which need a broader discussion. Some of
what I recall are:
a. Handling of abort of concurrent transactions. There is some code
in the patch which might work, but there is not much discussion when
it was posted.
b. Handling of partial tuples (while streaming, we came to know that
toast tuple is not complete or speculative insert is incomplete). For
this also, we have proposed a few solutions which need further
discussion. One of those is implemented in the patch series.
c. We might also need some handling for replication origins.
d. Try to minimize the performance overhead of WAL logging for
invalidations. We discussed different solutions for this and
implemented one of those.
e. How to skip already streamed transactions.

There might be a few more which I can't recall now. Apart from this,
I haven't done any detailed review of subscriber-side implementation
where we write streamed transactions to file. All of this will need
much more discussion and review before we can say it is ready to
commit, so I thought it might be better to pick it up for PG14 and
focus on other things that have a better chance for PG13 especially
because all the problems were not solved/discussed before last CF.
However, it is a good idea to keep moving this and have a discussion
on some of these issues.

> There's been a tremendous amount of work done since I last
> worked on it, and a lot was discussed on this thread, so it'll take a
> while to get familiar with the new code ...
>
> The first thing I realized that WAL-logging of assignments in v12 does
> both the "old" logging (using dedicated message) and "new" with
> toplevel-XID embedded in the first message. Yes, the patch was wrong,
> because it eliminated all calls to ProcArrayApplyXidAssignment() and so
> it was trivial to crash the replica due to KnownAssignedXids overflow.
> But I don't think re-introducing XLOG_XACT_ASSIGNMENT message is the
> right fix.
>
> I actually proposed doing this (having both ways to log assignments) so
> that there's no regression risk with (wal_level < logical). But IIRC
> Andres objected to it, argumenting that we should not log the same piece
> of information in two very different ways at the same time (IIRC it was
> discussed on the FOSDEM dev meeting, so I don't have a link to share).
> And I do agree with him ...
>

So, aren't we worried about the overhead of the amount of WAL and
performance impact for the transactions? We might want to check the
pgbench read-write test to see if that will add any significant
overhead.

> The question is, why couldn't the replica use the same assignment info
> we already write for logical decoding?
>

I haven't thought about it in detail, but we can think on those lines
if the performance overhead is in the acceptable range.

--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com