Re: [COMMITTERS] pgsql: Keep track of transaction commit timestamps

On 04/12/14 10:42, Heikki Linnakangas wrote:
> On 12/03/2014 04:54 PM, Alvaro Herrera wrote:
>> ir commit timestamp directly as they commit,
>> or an external transaction c
>
> Sorry, I'm late to the party, but here's some random comments on this
> after a quick review:
>
> * The whole concept of a node ID seems to be undocumented, and unused.
> No-one calls CommitTsSetDefaultNodeId(). The whole SET_TS record type is
> dead code too, when a non-default node_id is not set. Please remove, or
> explain how all that's supposed to work.
>

That's API meant to be used by extensions, maybe it would be preferable
if there was SQL API exposed also?

(It might also make more sense once replication identifiers are submitted.)

>
> * What is the definition of "latest commit", in
> pg_last_committed_xact()? It doesn't necessarily line up with the order
> of commit WAL records, nor with the order the proc array is updated
> (i.e. the order that the effects become visible to other backends). It
> seems confusing to add yet another notion of commit order. Perhaps
> that's the best we can do, but it needs to be documented.
>

It's updated on CommitTransaction (well RecordTransactionCommit but
that's only called from CommitTransaction) time so the definition of
latest commit is last CommitTransaction call.

--
Petr Jelinek http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services