Re: new high availability feature for the system with both asynchronous and synchronous replication

Hi all,

I create POC patch for my proposal of new feature for high availability.
I want to discuss about this feature. But this feature might be PG11
because discussion is not enough.

This patch enables walsender for async to wait until walsender for sync confirm
WAL is flashed to Disk. This feature is activated when GUC parameter
"async_walsender_delay" is set on.

I write the case when this feature is useful (this is the same as I wrote before):
1. Primary and synchronous standby are in the same center; called main center.
2. Asynchronous standby is in the another center; called backup center.
(The backup center is located far away from the main center. If replication
mode is synchronous, performance will be deteriorated. So, this replication
must be Asynchronous. )
3. Asynchronous replication is performed in the backup center too.
4. When primary in main center abnormally stops, standby in main center is
promoted, and the standby in backup center connects to the new primary.

[Main center]
|--------------------------------------------|
| |----------| synchronous |----------| |
| | | replication | | |
| | primary | <--------------> | standby1 | |
| |----------| |----------| |
|----||--------------------------------------|
||
|| asynchronous
|| replication
||
|| [Backup center]
|----||--------------------------------------|
| |----------| asynchronous |----------| |
| | | replication | | |
| | standby2 | <--------------> | standby3 | |
| |----------| |----------| |
|--------------------------------------------|

When the load in the main center becomes high, although WAL reaches standby in
backup center, WAL may not reach synchronous standby in main center for various
reasons. In other words, standby in the backup center may advance beyond
synchronous standby in main center.

When the primary abnormally stops and standby in main center promotes, two
standbys in backup center must be recovered by pg_rewind. However, it is
necessary to stop new primary for pg_rewind. If pg_basebackup is used,
recovery of backup center takes some times. This is not high availability.

Regards,
Daisuke Higuchi