>>>>> "Peter" == Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:
>> Did you miss the fact that the issue only occurs when the top-level
>> procedure does a rollback? The problem is not with nested calls, but
>> rather with the fact that commit or rollback is leaving
>> ActiveSnapshot unset, which is (as Tom pointed out at the linked
>> post) not the expected condition inside PL code.
Peter> We need that so that we can run things like SET TRANSACTION
Peter> ISOLATION.
While testing this, I ran into another semi-related issue:
shmem_exit_inprogress isn't ever being cleared in the postmaster, which
means that if you ever have a crash-restart, any attempt to do a
rollback in a procedure will then crash or get some other form of
corruption again every time until you manually restart the cluster.
--
Andrew (irc:RhodiumToad)