Scaling Notes

14.10.1 API tier (stateless)

• Run 3+ replicas behind the LB.
• HPA target: CPU 60%, memory 75%.
• Steady-state CPU per pod for ~100 RPS: ~250–500 m.
• Memory ~256–512 Mi per pod under steady load.
• Connection-pool size: min(50, n_postgres_connections / n_pods) — see pgxpool.Config.MaxConns.

14.10.2 Postgres

• Vertical first: start with 4 vCPU / 16 GB RAM, increase before sharding.
• Read replicas: for issuer-analytics + dashboard reads, deploy 1–2 read replicas; route via DSN logic in pkg/repository.
• Connection pooler: PgBouncer in transaction-pooling mode if total clients × MaxConns > Postgres max_connections.
• Indexes: the 002_* migrations cover the hot paths; profile new queries with EXPLAIN ANALYZE before adding.
• Vacuum: autovacuum on; tune autovacuum_vacuum_scale_factor=0.05 for the large audit_events and credentials tables.

14.10.3 Redis

• Single-instance is fine for OTP / rate-limit / session up to ~5k RPS.
• Replicated (1 primary + 2 replicas) for HA — failover via Sentinel or managed (ElastiCache).
• Cluster mode only if memory > 100 GB or > 50k RPS — IDA’s working set is small (TTL-bounded keys).
• Eviction: allkeys-lru (compose default) — safe because all data has natural TTLs except rate-limit buckets, which self-cleanup.

14.10.4 Blockchain RPC

• Self-hosted node preferred for hot workloads — avoids per-call quota.
• RPC pool: maintain 2–3 RPC URLs and round-robin in the chain client; failover on 5xx or RPC error.
• WebSocket subscription for events (DIDRegistry, RevocationRegistry) to keep off-chain cache fresh.
• Gas budgeting: monitor eth_gasPrice; alert if > 2x baseline. Bulk operations should retry with EIP-1559 fee bump.

14.10.5 Capacity planning checklist

---