On Thu, Sep 10, 2020 at 4:54 PM Alvaro Herrera <alvherre@2ndquadrant.com> wrote:
> Interesting example, thanks. It seems this can be fixed without
> breaking anything else by changing the planner so that it includes
> detached partitions when we are in a snapshot-isolation transaction.
> Indeed, the results from the detach-partition-concurrently-1.spec
> isolation test are more satisfying with this change.
Hmm, so I think the idea here is that since we're out-waiting plans
with the old partition descriptor by waiting for lock release, it's OK
for anyone who has a lock to keep using the old partition descriptor
as long as they continuously hold the lock. Is that right? I can't
think of a hole in that logic, but it's probably worth noting in the
comments, in case someone is tempted to change the way that we
out-wait plans with the old partition descriptor to some other
mechanism.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
