Re: [HACKERS] Speedup twophase transactions

>> Having CheckPointTwoPhase() do the flush would mean shifting the data
>> from KnownPreparedList into TwoPhaseState shmem.
>
> Er, no. For CheckPointTwoPhase(), at recovery what needs to be done is
> to have all the entries in KnownPreparedList() flushed to disk and
> have those entries removed while holding a shared memory lock.

The KnownPreparedList is constructed by the recovery process.
CheckPointTwoPhase() gets called by the checkpointer process. The
checkpointer does not have access to this valid KnownPreparedList.

>And for
> the rest we need to be careful to have PrescanPreparedTransactions,
> RecoverPreparedTransactions and StandbyRecoverPreparedTransactions
> aware of entries that are in KnownPreparedList().


Yeah, that part is straightforward. It does involve duplication of the
earlier while loops to work against KnownPreparedList. A smart single
while loop which loops over the 2PC files followed by the list would
help here :-)

> Let's leave the
> business of putting the information from KnownPreparedList to
> TwoPhaseState in RecoverPreparedTransactions, which need to be aware
> of entries in KnownPreparedList() anyway. The only thing that differs
> is how the 2PC information is fetched: from the segments or from the
> files in pg_twophase.
>

Yeah.  This part is also ok. We also got to be careful to mark the
shmem gxact entry with "ondisk = false" and need to set
prepare_start_lsn/prepare_end_lsn properly as well.

>> I wonder what's the best location for this in the common case when we
>> do shutdown of standby.  We could add code in XLOG_CHECKPOINT_SHUTDOWN
>> and XLOG_CHECKPOINT_ONLINE xlog_redo code path.
>
> ShutdownXLOG() calls CreateRestartPoint() when a standby shuts down,
> so doing all the durability work in CheckPointTwoPhase() would take
> care of any problems.
>

ShutdownXLOG() gets called from the checkpointer process. See comments
above about the checkpointer not having access to the proper
KnownPreparedList.

The following test sequence will trigger the issue:

1) start master
2) start replica
3) prepare a transaction on master
4) shutdown master
5) shutdown replica

CheckPointTwoPhase() in (5) does not sync this prepared transaction
because the checkpointer's KnownPreparedList is empty.

Regards,
Nikhils
-- Nikhil Sontakke                   http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services