On Wed, Dec 28, 2016 at 4:21 PM, Craig Ringer <craig@2ndquadrant.com> wrote:
> On 28 December 2016 at 08:14, Thomas Munro
> <thomas.munro@enterprisedb.com> wrote:
>
>> 3. No server must allow a transaction to be visible that hasn't been
>> flushed on N standby servers. We already prevent that on the primary
>
> Only if the primary doesn't restart. We don't persist the xact masking
> used by sync rep at the moment.
Right. Maybe you could fix that gap by making the primary wait until
the rule in synchronous_standby_names would be satisfied by the most
conservative possible synchronous_commit level (remote_apply) after
recovery and before allowing any queries to run?
> I suspect that solving that is probably tied to solving it on standbys.
Hmm. I was imagining that for standbys it might involve extra
messages flowing from the primary carrying the consensus write and
flush LSN locations (since there isn't any other kind of inter-node
communication possible, today), and then somehow teaching the standby
to see only the transactions whose commit record is <= the lastest
consensus commit LSN (precisely, and no more!) when acquiring a
snapshot if K is > 2 and N > 1 and you have synchronous_commit set to
a level >= remote_write on the standby. That could be done by simply
waiting for the consensus write or flush LSN (as appropriate) to be
applied before taking a snapshot, but aside from complicated
interlocking requirements, that would slow down snapshot acquisition
unacceptably on write-heavy systems. Another way to do it could be to
maintain separate versions of the snapshot data somehow for each
synchronous_commit level on standbys, so that you can get quickly your
hands on a snapshot that can only see xids whose commit record was <=
consensus write or flush location as appropriate. That interacts
weirdly with synchronous_commit = remote_apply on the primary though
because (assuming we want it to be useful) it needs to wait until the
LSN is applied on the standby(s) *and* they can see it in this weird
new time-delayed snapshot thing; perhaps it would require a new level
remote_apply_consensus_flush which waits for the standby(s) to apply
and and also know that the transaction has been flushed on enough
nodes to allow it to be seen...
--
Thomas Munro
http://www.enterprisedb.com