On Wed, Aug 4, 2021 at 11:43 PM Robert Haas <robertmhaas@gmail.com> wrote:
>
> Why do you think it's right to install the serialized *active*
> snapshot as the *transaction* snapshot? I've been operating on the
> presumption that we wanted the worker to install the leader's
> transaction snapshot as its transaction snapshot as its transaction
> snapshot and the leader's active snapshot as its active snapshot,
> because in my mind the active snapshot and the transaction snapshot
> are two separate things. However, there's clearly something wrong with
> that line of reasoning, because apparently your proposal fixes the
> problem and mine doesn't.
>
In setting up the snapshot for the execution state used in command
execution, GetTransactionSnapshot() is called and (possibly a copy of)
the returned snapshot is pushed as the ActiveSnapshot.
The way I see it is that there is essentially only one snapshot here,
the last-acquired TransactionSnapshot, which the ActiveSnapshot points
to (or a copy of it). Rightly or wrongly, my v2/v5 patch is basically
changing the code to do the same snapshot setup in each of the
parallel workers.
So why (current Postgres code, no patches applied) in setting up for
parallel-worker execution (in InitializeParallelDSM) does the Postgres
code then acquire ANOTHER TransactionSnapshot (by calling
GetTransactionSnashot(), which could return CurrentSnapshot or a new
snapshot) and serialize that, as well as serializing what the
ActiveSnapshot points to, and then restore those in the workers as two
separate snapshots? Is it a mistake? Or if intentional and correct,
how so?
Regards,
Greg Nancarrow
Fujitsu Australia
