Re: Stefan's bug (was: max_standby_delay considered harmful) - Mailing list pgsql-hackers

On Wed, May 26, 2010 at 9:06 PM, Fujii Masao <masao.fujii@gmail.com> wrote:
> On Wed, May 26, 2010 at 9:40 PM, Robert Haas <robertmhaas@gmail.com> wrote:
>> On Mon, May 24, 2010 at 10:35 AM, Fujii Masao <masao.fujii@gmail.com> wrote:
>>> On Mon, May 24, 2010 at 10:26 PM, Robert Haas <robertmhaas@gmail.com> wrote:
>>>> This looks pretty reasonable to me, but I guess I feel like it would
>>>> be better to drive the CancelBackup() decision off of whether we've
>>>> ever reached PM_RUN rather than consulting XLogCtl.  It just feels
>>>> cleaner to me to drive all of the postmaster decisions off of the same
>>>> signalling mechanism rather than having a separate one (that only
>>>> works because it's used very late in shutdown when we theoretically
>>>> don't need a lock) just for this one case.
>>>
>>> Okay, how about the attached patch? It uses the postmaster-local flag
>>> "ReachedEndOfRecovery" (better name?) instead of XLogCtl one.
>>
>> I've committed part of this patch, with the naming change that Tom
>> suggested.
>
> Thanks!
>
>> The parts I haven't committed are:
>>
>> 1. I don't see why we need to reset ReachedEndOfRecovery starting over
>> from PM_NO_CHILDREN.  It seems to me that once we reach PM_RUN, we
>> CAN'T go back to needing the backup label file, even if we have a
>> subsequent backend crash.  If I'm wrong, please let me know why and
>> I'll go put this back (with an appropriate comment).
>
> That reset has nothing to do with cancellation of the backup mode.
> I just added it since postmaster might use ReachedNormalRunning for
> another purpose in the future. For now, I have no objection not to
> add it.

OK, good.  I'm not sure it would be right to add it any way - reached
normal running sounds to me like it ought to mean "reached normal
running, ever" rather than "reached normal running since the last
backend crash".  In any event, it's moot for now.

>> 2. The changes to avoid launching WALReceiver except during certain
>> PM_* states.  It seems fairly sensible, but what is the case where
>> adding this logic prevents a problem?
>
> The problem is that shutdown would get stuck when walreceiver is
> invoked in PM_WAIT_BACKEND state. Imagine the following scenario:
>
> 1. Fast shutdown is requested just before the startup process calls
>   RequestXLogStreaming() which is the function to request postmaster
>   to invoke walreceiver.
>
> 2. pmdie() sends SIGTERM to the startup process, but not walreceiver
>   because it's not been started yet. Then, pmdie() switches the
>   state into PM_WAIT_BACKENDS.
>
> 3. The startup process goes through RequestXLogStreaming() and requests
>   postmaster to start walreceiver before processing SIGTERM sent from
>   pmdie(). Then the startup process exits, and postmaster invokes
>   walreceiver in PM_WAIT_BACKENDS state.
>
> 4. Once postmaster has reached PM_WAIT_BACKENDS, there is no way to send
>   SIGTERM to walreceiver. OTOH, postmaster cannot advance the state from
>   PM_WAIT_BACKENDS until walreceiver has gone away. So shutdown gets
>   stuck.

OK, makes sense.   I have committed this part also.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company