Horton

Re: row filtering for logical replication

Amit Kapila <amit.kapila16@gmail.com>

View thread
From: Amit Kapila <amit.kapila16@gmail.com>
To: Tomas Vondra <tomas.vondra@enterprisedb.com>
Cc: Ajin Cherian <itsajin@gmail.com>, Euler Taveira <euler@eulerto.com>, Peter Smith <smithpb2250@gmail.com>, Rahila Syed <rahilasyed90@gmail.com>, Peter Eisentraut <peter.eisentraut@enterprisedb.com>, Önder Kalacı <onderkalaci@gmail.com>, japin <japinli@hotmail.com>, Michael Paquier <michael@paquier.xyz>, David Steele <david@pgmasters.net>, Craig Ringer <craig@2ndquadrant.com>, Tomas Vondra <tomas.vondra@2ndquadrant.com>, Amit Langote <amitlangote09@gmail.com>, PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Date: 2021-09-25T05:22:47Z
Lists: pgsql-hackers

Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →

  1. Release cache tuple when no longer needed

  2. Add some additional tests for row filters in logical replication.

  3. Fix one of the tests introduced in commit 52e4f0cd47.

  4. Allow specifying row filters for logical replication of tables.

  5. Move scanint8() to numutils.c

  6. Replace Test::More plans with done_testing

  7. Reduce relcache access in WAL sender streaming logical changes

  8. Small cleanups related to PUBLICATION framework code

  9. Add a view to show the stats of subscription workers.

  10. Allow publishing the tables of schema.

  11. Doc: improve documentation of CREATE/ALTER SUBSCRIPTION.

  12. Add PublicationTable and PublicationRelInfo structs

  13. Remove unused argument "txn" in maybe_send_schema().

  14. Add prepare API support for streaming transactions in logical replication.

  15. Unify PostgresNode's new() and get_new_node() methods

  16. Use l*_node() family of functions where appropriate

  17. Add support for prepared transactions to built-in logical replication.

  18. Restore the portal-level snapshot after procedure COMMIT/ROLLBACK.

  19. Rename a parse node to be more general

  20. Remove unused column atttypmod from initial tablesync query

  21. SEARCH and CYCLE clauses

On Sat, Sep 25, 2021 at 3:07 AM Tomas Vondra
<tomas.vondra@enterprisedb.com> wrote:
>
> On 9/24/21 8:09 AM, Amit Kapila wrote:
> > On Thu, Sep 23, 2021 at 6:03 PM Tomas Vondra
> > <tomas.vondra@enterprisedb.com> wrote:
> >>
> >> 13) turning update into insert
> >>
> >> I agree with Ajin Cherian [4] that looking at just old or new row for
> >> updates is not the right solution, because each option will "break" the
> >> replica in some case. So I think the goal "keeping the replica in sync"
> >> is the right perspective, and converting the update to insert/delete if
> >> needed seems appropriate.
> >>
> >> This seems a somewhat similar to what pglogical does, because that may
> >> also convert updates (although only to inserts, IIRC) when handling
> >> replication conflicts. The difference is pglogical does all this on the
> >> subscriber, while this makes the decision on the publisher.
> >>
> >> I wonder if this might have some negative consequences, or whether
> >> "moving" this to downstream would be useful for other purposes in the
> >> fuure (e.g. it might be reused for handling other conflicts).
> >>
> >
> > Apart from additional traffic, I am not sure how will we handle all
> > the conditions on subscribers, say if the new row doesn't match, how
> > will subscribers know about this unless we pass row_filter or some
> > additional information along with tuple. Previously, I have done some
> > research and shared in one of the emails above that IBM's InfoSphere
> > Data Replication [1] performs filtering in this way which also
> > suggests that we won't be off here.
> >
>
> I'm certainly not suggesting what we're doing is wrong. Given the design
> of built-in logical replication it makes sense doing it this way, I was
> just thinking aloud about what we might want to do in the future (e.g.
> pglogical uses this to deal with conflicts between multiple sources, and
> so on).
>

Fair enough.

> >>
> >>
> >> 15) pgoutput_row_filter initializing filter
> >>
> >> I'm not sure I understand why the filter initialization gets moved from
> >> get_rel_sync_entry. Presumably, most of what the replication does is
> >> replicating rows, so I see little point in not initializing this along
> >> with the rest of the rel_sync_entry.
> >>
> >
> > Sorry, IIRC, this has been suggested by me and I thought it was best
> > to do any expensive computation the first time it is required. I have
> > shared few cases like in [2] where it would lead to additional cost
> > without any gain. Unless I am missing something, I don't see any
> > downside of doing it in a delayed fashion.
> >
>
> Not sure, but the arguments presented there seem a bit wonky ...
>
> Yes, the work would be wasted if we discard the cached data without
> using it (it might happen for truncate, I'm not sure). But how likely is
> it that such operations happen *in isolation*? I'd bet the workload is
> almost never just a stream of truncates - there are always some
> operations in between that would actually use this.
>

It could also happen with a mix of truncate and other operations as we
decide whether to publish an operation or not after
get_rel_sync_entry.

> Similarly for the errors - IIRC hitting an error means the replication
> restarts, which is orders of magnitude more expensive than anything we
> can save by this delayed evaluation.
>
> I'd keep it simple, for the sake of simplicity of the whole patch.
>

The current version proposed by Peter is not reviewed yet and by
looking at it I have some questions too which I'll clarify in a
separate email. I am not sure if you are against delaying the
expression initialization because of the current code or concept as a
general because if it is later then we have other instances as well
when we don't do all the work in get_rel_sync_entry like building
tuple conversion map which is cached as well.

-- 
With Regards,
Amit Kapila.