Hi,
On 2019-05-24 19:37:15 +0300, Konstantin Knizhnik wrote:
> From my point of view releasing all temporary table locks after preparing of
> 2PC transaction is not technically possible:
> assume that this transaction has updated some tuples of temporary table - them
> are not visible to other transactions until 2PC is committed,
> but since lock is removed, other transactions can update the same tuple.
I don't think tuple level actions are the problem? Those doesn't require
table level locks to be held.
Generally, I fail to see how locks themselves are the problem. The
problem are the catalog entries for the temp table, the relation forks,
and the fact that a session basically couldn't drop (and if created in
that transaction, use) etc the temp table after the PREPARE.
Greetings,
Andres Freund