WAL-G's PITR backups are insurance against data loss through erroneous data manipulations (eg: accidental DELETE/DROP/UPDATE). WAL-G's streaming approach (using pg_receivewal or similar) sends WAL records to backup storage continuously as they're generated, rather than waiting for a full segment to complete.

On top of that, for availability (and minimizing deadtime), we have 2 replicas using streaming replication. If the lead PG crashes, one of the replicas is promoted to lead (and starts accepting writes), and we "only" lose the writes that haven't been sent over the streaming replication.

You can fully eliminate that window of data loss with synchronous replication (vs the default asynchronous replication - which we use). The write slowdown (replica network round trip + 2nd write at replica) isn't worth it for us