When using Streaming Replication in Postgres 9.4, when a query makes the standby Postgres "start to count" the
max_standby_streaming_delay, changes being made in the primary:
- Are dammed at the primary (not sent to standby) using the
wal_keep_segmentssetting to limit that; OR
- Are sent to standby and stored locally (where?) at the standby for later processing?
Scenario 1 would lead to a bigger data loss in the event of a crash in the primary. On the other hand, the scenario 2 may lead to a filesystem being completely filled if the allowed dalay is too big.