Re: [PATCH] add concurrent_abort callback for output plugin

On Thu, Mar 25, 2021 at 2:37 PM Markus Wanner
<markus(dot)wanner(at)enterprisedb(dot)com> wrote:
>
> here is another tidbit from our experience with using logical decoding.
> The attached patch adds a callback to notify the output plugin of a
> concurrent abort. I'll continue to describe the problem in more detail
> and how this additional callback solves it.
>
> Streamed transactions as well as two-phase commit transactions may get
> decoded before they finish. At the point the begin_cb is invoked and
> first changes are delivered to the output plugin, it is not necessarily
> known whether the transaction will commit or abort.
>
> This leads to the possibility of the transaction getting aborted
> concurrent to logical decoding. In that case, it is likely for the
> decoder to error on a catalog scan that conflicts with the abort of the
> transaction. The reorderbuffer sports a PG_CATCH block to cleanup.
> However, it does not currently inform the output plugin. From its point
> of view, the transaction is left dangling until another one comes along
> or until the final ROLLBACK or ROLLBACK PREPARED record from WAL gets
> decoded. Therefore, what the output plugin might see in this case is:
>
> * filter_prepare_cb (txn A)
> * begin_prepare_cb (txn A)
> * apply_change (txn A)
> * apply_change (txn A)
> * apply_change (txn A)
> * begin_cb (txn B)
>
> In other words, in this example, only the begin_cb of the following
> transaction implicitly tells the output plugin that txn A could not be
> fully decoded. And there's no upper time boundary on when that may
> happen. (It could also be another filter_prepare_cb, if the subsequent
> transaction happens to be a two-phase transaction as well. Or an
> explicit rollback_prepared_cb or stream_abort if there's no other
> transaction in between.)
>
> An alternative and arguably cleaner approach for streamed transactions
> may be to directly invoke stream_abort. However, the lsn argument
> passed could not be that of the abort record, as that's not known at the
> point in time of the concurrent abort. Plus, this seems like a bad fit
> for two-phase commit transactions.
>
> Again, this callback is especially important for output plugins that
> invoke further actions on downstream nodes that delay the COMMIT
> PREPARED of a transaction upstream, e.g. until prepared on other nodes.
> Up until now, the output plugin has no way to learn about a concurrent
> abort of the currently decoded (2PC or streamed) transaction (perhaps
> short of continued polling on the transaction status).
>

I think as you have noted that stream_abort or rollback_prepared will
be sent (the remaining changes in-between will be skipped) as we
decode them from WAL so it is not clear to me how it causes any delays
as opposed to where we don't detect concurrent abort say because after
that we didn't access catalog table.

--
With Regards,
Amit Kapila.