docs: Add HAProxy architecture analysis for EDB PostgreSQL routing

[chadmf]

Summary

Add comprehensive architectural decision record (ADR) for replacing pgBouncer with HAProxy for AAP database connection routing.

Problem

• Standard EDB reference architecture requires pgBouncer with EFM VIP management
• AAP has compatibility issues with pgBouncer that prevent its use
• Need alternative routing solution for AAP containers → PostgreSQL cluster

Solution

New Architecture Pattern:

AAP Containers → HAProxy (DB Router) → PostgreSQL VIP (EFM-managed) → PRIMARY

Clean separation of concerns:

• EFM: Database-layer VIP failover orchestration (15s)
• HAProxy: Routing layer with health checks (10s)
• Total failover detection: ~25s (well within 5-minute RTO)

Changes

New Documentation

• docs/haproxy-pgbouncer-architectural-analysis.md (500+ lines)
  • Architecture comparison: HAProxy vs pgBouncer
  • Design validation and trade-off analysis
  • Complete implementation guidance
  • External health check script (pg_is_in_recovery() validation)
  • Monitoring strategies and alerting
  • Long-term supportability considerations

Updated Documentation

• docs/aap-containerized-enterprise-dr-architecture.md
  • Updated HAProxy configuration and network topology
  • Revised AAP inventory files (pg_host points to HAProxy)
  • Updated component specifications and naming

Other

• .gitignore: Add .pub pattern

Key Architectural Trade-offs

Aspect	Impact	Mitigation
No connection pooling	+67% PostgreSQL connections	Increase max_connections to 2500, add RAM (32GB → 48GB)
Query latency	-0.5ms faster (TCP vs protocol)	✅ Benefit
Operational complexity	Simpler (standard HAProxy)	✅ Benefit
Failover detection	+5-10s lag behind EFM	Acceptable for RTO
Infrastructure cost	+$300-500/month (RAM increase)	Budget approval needed

Critical Implementation Requirements

1. PostgreSQL Resource Increase:

   • max_connections: 1500 → 2500
   • RAM per node: 32GB → 48GB
   • shared_buffers: 8GB → 12GB

2. HAProxy External Health Check:

   • Script validates pg_is_in_recovery() = false (writable node)
   • 5-second check interval with rise/fall thresholds

3. Monitoring:

   • PostgreSQL connection count (alert at 2000/2500)
   • HAProxy backend health status
   • Replication lag (existing)

Impact on RTO/RPO

• RTO: < 5 minutes ✅ (maintained)
• RPO: < 5 seconds ✅ (maintained)
• Failover detection: EFM (15s) + HAProxy health check (10s) = 25s

Test Plan

• Review architectural decision and trade-offs
• Validate HAProxy configuration syntax
• Review external health check script logic
• Confirm PostgreSQL resource requirements
• Validate integration with EFM failover
• Review monitoring and alerting strategy
• Assess long-term supportability concerns

Related Issues

Addresses AAP/pgBouncer compatibility constraints while maintaining EDB PostgreSQL high availability requirements.

---

🤖 Generated with Claude Code