Re: Parallel scan with SubTransGetTopmostTransaction assert coredump

Have you found a case that the v2 patch breaks?

The v2 patch does follow the analysis in the beginning of the
discussion, which identified that in setting up parallel workers, a
"later transaction snapshot" was taken than the one actually used in
the statement execution, and that's what eventually leads to the
observed Assertion failure.
The original problem reporter stated: "So the root cause is the
Parallel Workers process set the TransactionXmin with later
transcation snapshot".
Also, as far as I can see, it makes no sense to pass parallel workers
both an active snapshot and a (later) transaction snapshot. In the
leader, prior to execution and running parallel workers, a transaction
snapshot is obtained and pushed as the active snapshot (note: only ONE
snapshot at this point). It makes little sense to me to then obtain
ANOTHER transaction snapshot when setting up parallel workers, and
pass that (only) to the workers along with the original (earlier)
active snapshot. (If there is a reason for passing the TWO snapshots
to parallel workers, original code authors and/or snapshot experts
should speak up now, and code comments should be added accordingly to
explain the purpose and how it is MEANT to work)
This is why the v2 patch updates the code to just pass one snapshot
(the active snapshot) to the parallel workers, which gets restored in
each worker as the transaction snapshot and set as the active snapshot
(so is then consistent with the snapshot setup in the parallel
leader).