From 06019ef74c64e3c9c092b911231df347e4516c3c Mon Sep 17 00:00:00 2001
From: sam <info@apodacalab.com>
Date: Tue, 19 May 2026 14:04:34 -0700
Subject: [PATCH] Add consumer-only psql-app replica for ingestion scale-out

psql-app-consumer: profile-gated (scale-out) horizontal scale-out for the
Kafka->Postgres ingestion path. Shares the primary's /config read-only so
it reuses obmp-psql.yml, whose fixed group.id makes Kafka rebalance
partitions across the primary and every replica. Its command runs ONLY
the consumer jar -- no cron, RPKI/IRR/DBIP or initdb -- so it does not
duplicate the primary's DB-maintenance jobs (config_cron wires those up
unconditionally in /usr/sbin/run). Each replica brings its own consumer
and writer threads.

Measured: one consumer-only replica took the post-storm backlog drain
from a cold-start ~3.7k msg/s to ~48k msg/s; group membership 8->16. With
2 consumers feeding it, Postgres becomes the next bottleneck (~500% CPU)
-- DB write capacity is the ceiling beyond ~2-3 consumers.

  docker compose --profile scale-out up -d --scale psql-app-consumer=2

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 docker-compose.yml | 26 ++++++++++++++++++++++++++
 1 file changed, 26 insertions(+)

diff --git a/docker-compose.yml b/docker-compose.yml
index 2205f17..411d251 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -255,6 +255,32 @@ services:
       - POSTGRES_DROP_stats_peer_rib='4 weeks'
       - POSTGRES_DROP_stats_peer_update_counts='4 weeks'
 
+  # Consumer-only psql-app replica -- horizontal ingestion scale-out.
+  # Profile-gated; bring up on demand (the host needs spare CPU+RAM for it):
+  #   docker compose --profile scale-out up -d --scale psql-app-consumer=2
+  # It shares the primary's /config (read-only) so it reuses obmp-psql.yml,
+  # whose fixed group.id "obmp-psql-consumer" makes Kafka rebalance partitions
+  # across the primary and every replica. The command runs ONLY the consumer
+  # jar -- no cron, no RPKI/IRR/DBIP, no initdb -- so a replica does NOT
+  # duplicate the primary's DB-maintenance jobs (update_global_ip_rib,
+  # update_chg_stats, retention, ...), which config_cron wires up
+  # unconditionally in /usr/sbin/run. Each replica brings its own consumer
+  # AND writer threads, so it adds real write throughput (the primary's
+  # writer_max_threads_per_type is 1).
+  psql-app-consumer:
+    profiles: ["scale-out"]
+    restart: unless-stopped
+    image: openbmp/psql-app:2.2.2
+    mem_limit: ${PSQL_APP_CONSUMER_MEM_LIMIT:-4g}
+    depends_on:
+      psql:
+        condition: service_healthy
+      kafka:
+        condition: service_started
+    volumes:
+      - ${OBMP_DATA_ROOT}/config:/config:ro
+    command: ["bash","-c","cd /var/log && exec java -Xmx3G -Xms128m -XX:+UseG1GC -XX:+UnlockExperimentalVMOptions -XX:InitiatingHeapOccupancyPercent=30 -XX:G1MixedGCLiveThresholdPercent=30 -XX:MaxGCPauseMillis=200 -XX:ParallelGCThreads=20 -XX:ConcGCThreads=5 -XX:+ExitOnOutOfMemoryError -Duser.timezone=UTC -jar /usr/local/openbmp/obmp-psql-consumer.jar -cf /config/obmp-psql.yml"]
+
   exabgp:
     restart: unless-stopped
     container_name: obmp-exabgp