Re: Patch for fail-back without fresh backup

On Wed, Sep 18, 2013 at 11:45 AM, Fujii Masao <masao(dot)fujii(at)gmail(dot)com> wrote:
> On Wed, Sep 18, 2013 at 10:35 AM, Sawada Masahiko <sawada(dot)mshk(at)gmail(dot)com> wrote:
>> On Tue, Sep 17, 2013 at 9:52 PM, Fujii Masao <masao(dot)fujii(at)gmail(dot)com> wrote:
>>> I set up synchronous replication with synchronous_transfer = all, and then I ran
>>> pgbench -i and executed CHECKPOINT in the master. After that, when I executed
>>> CHECKPOINT in the standby, it got stuck infinitely. I guess this was cased by
>>> synchronous_transfer feature.
>>
>> Did you set synchronous_standby_names in the standby server?
>
> Yes.
>
>> If so, the master server waits for the standby server which is set to
>> synchronous_standby_names.
>> Please let me know detail of this case.
>
> Both master and standby have the same postgresql.conf settings as follows:
>
> max_wal_senders = 4
> wal_level = hot_standby
> wal_keep_segments = 32
> synchronous_standby_names = '*'
> synchronous_transfer = all
>
>>> How does synchronous_transfer work with cascade replication? If it's set to all
>>> in the "sender-side" standby, it can resolve the data page inconsistency between
>>> two standbys?
>>>
>>
>> Currently patch supports the case which two servers are set up SYNC replication.
>> IWO, failback safe standby is the same as SYNC replication standby.
>> User can set synchronous_transfer in only master side.
>
> So, it's very strange that CHECKPOINT on the standby gets stuck infinitely.
>

I attached the patch which I have modified.

I have modified that if both synchronous replication and synchronous
transfer are requested,
but the server still in recovery(i.g., the server is in standby mode),
the server doesn't wait for
corresponding WAL replicated.
Specifically, I added condition RecoveryInProgress().

If both functions(synchronous replication and transfer) are set and
user sets up synchronous replication between two servers,
user can executes CHECKPOINT on standby side. It will not wait for
corresponding WAL replicated.
But, If both parameter are set and user doesn't set up synchronous
replication(i.g., the master server works alone),
the master server waits infinitely when user executes CHECKPOINT. This
behaviour is similar to synchronous replication.

Regards,

-------
Sawada Masahiko