Re: Potential data loss due to race condition during logical replication slot creation

On Tue, Apr 16, 2024 at 9:27 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>
> Pondering further, I came across the question; in what case we would
> need to restore the snapshot and jump to the consistent state when
> finding the initial start point?
>
> When creating a slot, in ReplicationSlotReserveWal(), we determine the
> restart_lsn and write a RUNNING_XACT record. This record would be the
> first RUNNING_XACT record that the logical decoding decodes in most
> cases.
>

This probably won't be true on standby as we don't LOG RUNNING_XACT
record in that case.

> In SnapBuildFindSnapshot(), if the record satisfies (a) (i.e.
> running->oldestRunningXid == running->nextXid), it can jump to the
> consistent state. If not, it means there were running transactions at
> that time of the RUNNING_XACT record being produced. Therefore, we
> must not restore the snapshot and jump to the consistent state.
> Because otherwise, we end up deciding the start point in the middle of
> a transaction that started before the RUNNING_XACT record. Probably
> the same is true for all subsequent snapbuild states.
>

One thing to consider on this is that we can only restore the snapshot
if by that time some other get changes (in your case, step
"s2_get_changes_slot0" as in your draft patch) would have serialized
the consistent snapshot at that LSN. One could question if we can't
reach a consistent state at a particular LSN then why in the first
place the snapshot has been serialized at that LSN? The answer could
be that it is okay to use such a serialized snapshot after initial
snapshot creation because we know that the restart_lsn of a slot in
such cases would be a location where we won't see the data for partial
transactions. This is because, after the very first time (after the
initdecodingcontext), the restart_lsn would be set to a location after
we reach the consistent point.

> I might be missing something but I could not find the case where we
> can or want to restore the serialized snapshot when finding the
> initial start point. If my analysis is correct, we can go with either
> (a) or (c) I proposed before[1]. Between these two options, it also
> could be an option that (a) is for backbranches for safety and (c) is
> for master.
>

The approach (a) has a downside, it will lead to tracking more
transactions (non-catalog) than required without any benefit for the
user. Considering that is true, I wouldn't prefer that approach.

-- 
With Regards,
Amit Kapila.