Re: Unintended restart after recovery error

On Thu, Nov 13, 2014 at 8:30 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Wed, Nov 12, 2014 at 4:52 PM, Antonin Houska <ah@cybertec.at> wrote:
>> Fujii Masao <masao.fujii@gmail.com> wrote:
>>
>>> On Wed, Nov 12, 2014 at 6:52 PM, Antonin Houska <ah@cybertec.at> wrote:
>>> > While looking at postmaster.c:reaper(), one problematic case occurred to me.
>>> >
>>> >
>>> > 1. Startup process signals PMSIGNAL_RECOVERY_STARTED.
>>> >
>>> > 2. Checkpointer process is forked and immediately dies.
>>> >
>>> > 3. reaper() catches this failure, calls HandleChildCrash() and thus sets
>>> > FatalError to true.
>>> >
>>> > 4. Startup process exits with non-zero status code too - either due to SIGQUIT
>>> > received from HandleChildCrash or due to some other failure of the startup
>>> > process itself. However, FatalError is already set, because of the previous
>>> > crash of the checkpointer. Thus reaper() does not set RecoveryError.
>>> >
>>> > 5. As RecoverError failed to be set to true, postmaster will try to restart
>>> > the cluster, although it apparently should not.
>>>
>>> Why shouldn't postmaster restart the cluster in that case?
>>>
>>
>> At least for the behavior to be consistent with simpler cases of failed
>> recovery (e.g. any FATAL error in StartupXLOG), which end up not restarting
>> the cluster.
>
> It's true that if the startup process dies we don't try to restart,
> but it's also true that if the checkpointer dies we do try to restart.
> I'm not sure why this specific situation should be an exception to
> that general rule.

442231d7f71764b8c628044e7ce2225f9aa43b6 introduced the latter rule
for hot-standby case. Maybe *during crash recovery* (i.e., hot standby
should not be enabled) it's better to treat the crash of startup process as
a catastrophic crash.

Regards,

-- 
Fujii Masao