On Fri, May 20, 2011 at 09:37:20AM +0100, Leonardo Francalanci wrote:
> I'll try to sum up what I understood:
>
> 1) the standby keeps the lock, so no problem with
> stray files coming from the unlogged->logged log
> reply, as the table can't be read during the operation
>
> 2) calling ResetUnloggedRelations before
> ProcArrayApplyRecoveryInfo would remove the problem
> of the stray files on the standby in case of master crash
> before commit/abort
>
> 3) promoting the standby shouldn't be an issue,
> since ResetUnloggedRelations is already called in
> ShutdownRecoveryTransactionEnvironment
All correct, as far as I can tell.
> Now, to move forward, some questions:
>
> - the patch is missing the "send all table pages to the
> standby" part; is there some code I can use as base?
Nothing comes to mind as especially similar.
> I guess I have to generate some special log type that
> is only "played" by standby servers.
What you described in your followup mail seemed reasonable.
> - on the standby, the commit part should be played as it
> is on the master (that is, removing the INIT fork).
> The abort case is different though: it would mean
> doing nothing on the master, while removing every forks
> but the INIT fork on the standby.
> Would it be ok to add to xl_xact_abort a new array of
> RelFileNode(s), where for each one at abort all the forks,
> except the init fork, have to be deleted by the standby
> (while the master shouldn't do anything with them)?
> I bet there's a cleaner solution...
Your "use less space in xl_xact_commit patch" seems to be going in a good
direction here. It would probably also be okay to do a ResetUnloggedRelations()
on the standby at every abort of a transaction that had started an UNLOGGED ->
LOGGED conversion. That is, just a flag might be enough.
nm
