On 21.06.2017 04:48, Michael Paquier wrote:
There has not been much activity on this thread for some time, and I mentioned my intentions to some developers at the last PGCon. But I am planning to study more the work that has been done here, with as envisaged goal to present a patch for the first CF of PG11. Lots of fun ahead.
Hi Michael,Glad to see you working on this! I've been studying this topic too. Attached you can find a recently rebased version of Heikki's v4 patch. I also fixed a bug that appeared on report-receipts isolation test: XidIsConcurrent should report that a transaction is concurrent to ours when its csn equals our snapshotcsn.
There is another bug where multiple tuples with the same primary key appear after a pgbench read-write test. Querying pgbench_branches with disabled index scans, I see several tuples with the same 'bid' value. Usually they do not touch the index so there are no errors. But sometimes HOT update is not possible and it errors out with violation of unique constraint, which is how I noticed it. I can't reproduce it on my machine and have to use a 72-core one.
For now I can conclude that the oldestActiveXid is not always updated correctly. In TransactionIdAsyncCommitTree, just before the transaction sets clog status, its xid can be less than oldestActiveXid. Other transactions are seeing it as aborted for some time before it writes to clog (see TransactionIdGetCommitSeqNo). They can update a tuple it deleted, and that leads to duplication. Unfortunately, I didn't have time yet to investigate this further.
-- Alexander Kuzmenkov Postgres Professional: http://www.postgrespro.com The Russian Postgres Company
csn-5.patch.gz
Description: application/gzip
-- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
