Remote Recovery for data guarding
PostgreSQL introduces full page write mechanism to avoid inconsistency in case of system crashes. However, full page writes can amplify writes and reduce performance. Remote Recovery is designed to fetch full page from the mirror node (standby or master) during recovery to prevent torn pages in the absence of full page write. As a result, full page write can be turned off for reduced IO and improved performance.
Specifically, PolarDB PG writes a special bit in WAL when a page is first modified since the last checkpoint. During recovery, upon encountering a special bit that indicates a remote fetching is needed, a PolarDB PG database instance would fetch the page from its mirror instance and restore it locally before replaying later modifications.
Remote recovery also can leverage our parallel WAL redo framework to fetch pages and to relay WAL records in parallel.
How to use
- You can put a remote_recovery.conf in the data directory to support remote recovery. The format of remote_recovery.conf is same as recovery.conf. You should specify the mirror node address for building connection. For example, you can write "standby_conninfo = 'host=xx port=xx user=xx password=xx application_name=standby'" into remote_recovery.conf.
- The parameter of max_wal_senders in the mirror node should be configured to be larger than max_parallel_replay_workers in the recovering node.
- If the mirror node is a backup node, it should turn on hot_standby for accepting connections from the recovering node. checkpoint_sync_standby should be turned on to guarantee that the mirror node has full WAL modifications once a checkpoint is done.
