On Thu, Aug 8, 2024 at 2:37 PM Hayato Kuroda (Fujitsu)
<kuroda.hayato@fujitsu.com> wrote:
>
> Thanks for discussing!
>
> I reported the issue because 1) I feared the risk of data loss and 2) simply
> because the coding looked incorrect. However, per discussion, I understood that
> it wouldn't lead to loss, and adding a global variable was unacceptable in this
> case. I modified the patch completely.
>
> The attached patch avoids using the LastCommitLSN as the local_lsn while applying
> PREPARE. get_flush_position() was not changed. Also, it contains changes that
> have not been discussed yet:
>
> - Set last_commit_end to InvaldXLogPtr in the PREPARE case.
> This causes the same result as when the stream option is not "parallel."
> - XactLastCommitEnd was replaced even ROLLBACK PREPARED case.
> Since the COMMIT PREPARED record is flushed in RecordTransactionAbortPrepared(),
> there is no need to ensure the WAL must be sent.
>
The code changes look mostly good to me. I have changed/added a few
comments in the attached modified version.
--
With Regards,
Amit Kapila.