On Mon, Apr 26, 2021 at 6:16 PM Tom Lane <tgl@xxxxxxxxxxxxx> wrote:
Hmm, that's an invalid VXID, which would ordinarily mean that nothing
is holding the lock. There is a passing mention in mvcc.sgml that
SIRead locks sometimes need to be held past transaction end, so maybe
what you're looking at is such a lock that is no longer associated
with a specific transaction. I have to disclaim knowing much of
anything about the SSI code, though.
The "Serializable Snapshot Isolation in PostgreSQL" paper (https://drkp.net/papers/ssi-vldb12.pdf) makes me think this is a reasonable line of inquiry.
Thanks,
Mike
