Horton

Re: speed up a logical replica setup

Amit Kapila <amit.kapila16@gmail.com>

View thread
From: Amit Kapila <amit.kapila16@gmail.com>
To: "Hayato Kuroda (Fujitsu)" <kuroda.hayato@fujitsu.com>
Cc: Euler Taveira <euler@eulerto.com>, Alexander Lakhin <exclusion@gmail.com>, "pgsql-hackers@lists.postgresql.org" <pgsql-hackers@lists.postgresql.org>, "Zhijie Hou (Fujitsu)" <houzj.fnst@fujitsu.com>
Date: 2024-07-17T10:10:08Z
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. pg_createsubscriber: Remove obsolete comment

  2. pg_createsubscriber: Fix an unpredictable recovery wait time.

  3. Fix unstable test in 040_pg_createsubscriber.

  4. Fix the testcase introduced in commit 81d20fbf7a.

  5. Further weaken new pg_createsubscriber test on Windows.

  6. Temporarily(?) weaken new pg_createsubscriber test on Windows.

  7. Make pg_createsubscriber warn if publisher has two-phase commit enabled.

  8. Make pg_createsubscriber more wary about quoting connection parameters.

  9. pg_createsubscriber: Remove failover replication slots on subscriber

  10. pg_createsubscriber: Remove replication slot check on primary

  11. pg_createsubscriber: Only --recovery-timeout controls the end of recovery process

  12. pg_createsubscriber: creates a new logical replica from a standby server

  13. Add some const decorations

  14. Add option force_initdb to PostgreSQL::Test::Cluster:init()

  15. Remove MSVC scripts

On Wed, Jul 17, 2024 at 1:23 PM Hayato Kuroda (Fujitsu)
<kuroda.hayato@fujitsu.com> wrote:
>
> I also analyzed this failure, let me share it. Here, I think events in below
> ordering were occurred.
>
> 1. Backend created a publication on $db2,
> 2. BGWriter generated RUNNING_XACT record, then
> 3. Backend created a replication slot on $db2.
>
> In this case, the recovery_target_lsn is ahead of the RUNNING_XACT record generated
> at step 3. Also, since both bgwriter and slot creation mark the record as
> *UNIMPORTANT* one, the writer won't start again even after the
> LOG_SNAPSHOT_INTERVAL_MS. The rule is written in BackgroundWriterMain():
>
> ```
>                         /*
>                          * Only log if enough time has passed and interesting records have
>                          * been inserted since the last snapshot.  Have to compare with <=
>                          * instead of < because GetLastImportantRecPtr() points at the
>                          * start of a record, whereas last_snapshot_lsn points just past
>                          * the end of the record.
>                          */
>                         if (now >= timeout &&
>                                 last_snapshot_lsn <= GetLastImportantRecPtr())
>                         {
>                                 last_snapshot_lsn = LogStandbySnapshot();
>                                 last_snapshot_ts = now;
>                         }
> ```
>
> Therefore, pg_createsubscriber waited until a new record was replicated, but no
> activities were recorded, causing a timeout. Since this is a timing issue, Alexander
> could reproduce the failure with shorter time duration and parallel running.
>

Your analysis sounds correct to me.

> IIUC, the root cause is that pg_create_logical_replication_slot() returns a LSN
> which is not generated yet. So, I think both mine [1] and Euler's approach [2]
> can solve the issue. My proposal was to add an extra WAL record after the final
> slot creation, and Euler's one was to use a restart_lsn as the recovery_target_lsn.
>

I don't think it is correct to set restart_lsn as consistent_lsn point
because the same is used to set replication origin progress. Later
when we start the subscriber, the system will use that LSN as a
start_decoding_at point which is the point after which all the commits
will be replicated. So, we will end up incorrectly using restart_lsn
(LSN from where we start reading the WAL) as start_decoding_at point.
How could that be correct?

Now, even if we use restart_lsn as recovery_target_lsn and the LSN
returned by pg_create_logical_replication_slot() as consistent LSN to
set replication progress, that also could lead to data loss because
the subscriber may never get data between restart_lsn value and
consistent LSN value.

-- 
With Regards,
Amit Kapila.