On Wed, Mar 24, 2021 at 6:18 PM Robert Haas <robertmhaas@gmail.com> wrote:
>
> On Wed, Mar 24, 2021 at 12:48 AM Andres Freund <andres@anarazel.de> wrote:
>
> > Although this specific hack doesn't seem too terrible to me. If you
> > execute a parallel insert the likelihood to end up not needing an xid is
> > pretty low. Implementing it concurrently does seem like it'd end up
> > needing another lwlock nested around xid assignment, or some more
> > complicated scheme with already holding XidGenLock or retries. But maybe
> > I'm missing an easy solution here.
>
> I don't think you need to do anything that is known outside the group
> of processes involved in the parallel query. I think you just need to
> make sure that only one of them is trying to acquire an XID at a time,
> and that all the others find out about it. I haven't thought too hard
> about the timing: if one process acquires an XID for the transaction,
> is it OK if the others do an arbitrary amount of work before they
> realize that this has happened?
>
A naive question about this scheme: What if the worker that acquires
the XID writes some row and another worker reads that row before it
gets to see the XID information? I think it won't treat such a row is
written by its own transaction. Won't such a scheme lead to different
behavior than serial inserts or where we have acquired XID before
starting parallel-operation?
--
With Regards,
Amit Kapila.
