The behavior of PeerDB while consuming changes from the logical replication slot is as follows:
  1. PeerDB would start reading the slot and keep waiting until the first record appears. It never exits before the first record appears.
  2. Once the first record appears - SyncFlow exits based on whichever comes first:
  3. PEERDB_CDC_IDLE_TIMEOUT_SECONDS defaults to 60s. This is set as an environment variable for the flow worker. OR
  4. max_batch_size defaults to 100K. This is set as a part an option (WITH part) of the CREATE MIRROR command.
Below are the list of a few recommendations that you can follow while tuning PEERDB_CDC_IDLE_TIMEOUT_SECONDS and max_batch_size

​
Latency / Lag Sensitive Workloads

  1. Low PEERDB_CDC_IDLE_TIMEOUT_SECONDS . You can leave this as default i.e. 60s. Other sane defaults incl. 30s, 40s etc.
  2. High enough max_batch_size  to avoid too many roundtrips and support larger throughputs. You can leave this as default i.e. 100K or go a bit higher say 500K

​
Others

  1. High PEERDB_CDC_IDLE_TIMEOUT_SECONDS . Can be 5mins, 10mins, 30mins based on what acceptable lag or realtime-ness you are aiming for.
  2. High enough max_batch_size to avoid too many roundtrips and support larger throughputs. You can decide this based on the memory available on the PeerDB instance. For example 100K for instances with low memory (<=16GB) and 1 million for vms with higher RAM (>=32GB RAM)

​
Postgres configs

  1. Make sure to set wal_sender_timeout to 0
  2. If you constantly observing IO wait_event_type in pg_stat_activity while PeerDB is consuming the slot, consider increasing (say doubling) logical_decoding_work_mem