On Fri, Oct 2, 2015 at 11:58 PM, Alvaro Herrera
<alvherre@2ndquadrant.com> wrote:
> Fujii Masao wrote:
>
>> What happens if pg_xact_commit_timestamp() is called in standby after
>> track_commit_timestamp is disabled in master, DeactivateCommitTs() is
>> called and all commit_ts files are removed in standby? I tried that case
>> and got the following assertion failure.
>
> Ah. So the standby needs to keep the module activated if it's enabled
> locally, even when it receives a message that the master turned it off.
> Here's a patch.
I'm afraid that this behavior might confuse the users.
Please imagine the following scenario.
1. start up the server with track_commit_timestamp disabled
2. run several transactions
3. shut down the server with immediate mode
4. restart the server with track_commit_timestamp enabled
5. run "SELECT pg_last_committed_xact()"
6. run "SELECT pg_xact_commit_timestamp(xid) FROM pg_last_committed_xact()"
7. restart the server
8. run "SELECT pg_last_committed_xact()"
Firstly, in #5, pg_last_committed_xact() returns the XID and
timestamp of the last transaction which was executed in #2
(i.e., while track_commit_timestamp was disabled).
This is confusing. I think that both pg_last_committed_xact()
and pg_xact_commit_timestamp() should return only the transaction
which was executed with track_commit_timestamp enabled.
Secondly, SELECT query in #6 returns NULL. This means that
pg_xact_commit_timestamp() may not handle the transaction
which pg_last_committed_xact() handles. This is also confusing.
Finally, in #8, pg_last_committed_xact() returns NULL while
it returned non-NULL before the restart. This is also confusing.
Regards,
--
Fujii Masao