On Wed, Aug 28, 2019 at 3:05 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> I'd vote for back-patching to 10. Even if there is in fact no deadlock
> hazard, you've clearly demonstrated a significant performance hit that
> we're taking for basically no reason.
Done.
The second symptom reported in my first email looked like evidence of
high levels of spinlock backoff, which I guess might have been coming
from TransactionIdLimitedForOldSnapshots()'s hammering of
oldSnapshotControl->mutex_threshold and
oldSnapshotControl->mutex_threshold, when running
heap_page_prune_opt()-heavy workloads like the one generated by
test-indexscan.sql (from my earlier message) from many backends at the
same time on a large system. That's just an observation I'm leaving
here, I'm not planning to chase that any further for now.
--
Thomas Munro
https://enterprisedb.com