Re: pg_rewind with cascade standby doesn't work well

> Consider a scenario like this,
>
> Server A: primary
> Server B :replica of A
> Server C :replica of B
>
> and somehow A down ,so B gets promoted.
> Server A: down
> Server B :new primary
> Server C :replica of B
>
> In this case, pg_rewind can be used to reconstruct the cascade; the
source is C and the target is A.
> However, we get error as belows by running pg_rewind.
>
> ```
> pg_rewind: fetched file "global/pg_control", length 8192
> pg_rewind: source and target cluster are on the same timeline
> pg_rewind: no rewind required
> ```

To fix the above mentioned behavior of pg_rewind, I suggest to change the
cascade standby's (i.e. server C's) minRecoveryPointTLI when it receives
the new timeline information from the new primary (i.e. server B).

When server B is promoted, it creates an end-of-recovery record by calling
CreateEndOfRecoveryRecord(). (in xlog.c)
And also updates B's minRecoveryPoint and minRecoveryPointTLI.
```
/*
* Update the control file so that crash recovery can follow the
timeline
* changes to this point.
*/
LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
ControlFile->minRecoveryPoint = recptr;
ControlFile->minRecoveryPointTLI = xlrec.ThisTimeLineID;
UpdateControlFile();
LWLockRelease(ControlFileLock);
```
Since C is a replica of B, the end-of-recovery record is replicated from B
to C, so the record is replayed in C by xlog_redo().
The attached patch updates minRecoveryPoint and minRecoveryPointTLI at this
point by mimicking CreateEndOfRecoveryRecord().
With this patch, you can run pg_rewind with cascade standby immediately.
(without waiting for checkpointing)

Thoughts?

Masaki Kuwamura