Re: BUG #16643: PG13 - Logical replication - initial startup never finishes and gets stuck in startup loop

On 2020-Nov-04, Amit Kapila wrote:

> On Thu, Oct 15, 2020 at 8:20 PM Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> wrote:

> > * STREAM COMMIT bug?
> > In apply_handle_stream_commit, we do CommitTransactionCommand, but
> > apparently in a tablesync worker we shouldn't do it.
>
> In the tablesync stage, we don't allow streaming. See pgoutput_startup
> where we disable streaming for the init phase. As far as I understand,
> for tablesync we create the initial slot during which streaming will
> be disabled then we will copy the table (here logical decoding won't
> be used) and then allow the apply worker to get any other data which
> is inserted in the meantime. Now, I might be missing something here
> but if you can explain it a bit more or share some test to show how we
> can reach here via tablesync worker then we can discuss the possible
> solution.

Hmm, okay, that sounds like there would be no bug then. Maybe what we
need is just an assert in apply_handle_stream_commit that
!am_tablesync_worker(), as in the attached patch. Passes tests.