diff --git a/docs/production-sizing.md b/docs/production-sizing.md
new file mode 100644
index 0000000..ec01b97
--- /dev/null
+++ b/docs/production-sizing.md
@@ -0,0 +1,96 @@
+# OpenBMP Production Sizing — 40 Full-Table-Edge Routers
+
+Sizing guidance for deploying the OpenBMP stack against a production ISP
+network of **40 full-table-edge routers** with gNMI streaming telemetry.
+Derived from the OpenBMP `psql-app` sizing guidance and measured lab behavior.
+
+## Workload assumptions
+
+| Parameter | Value |
+|-----------|-------|
+| Monitored routers | 40, full-table edge |
+| BMP RIB scope | Adj-RIB-In (see recommendation below) |
+| Full feeds per router | ~2–3 eBGP peers carrying the full DFZ |
+| Routes per full feed | ~1.2M (≈1M IPv4 + ~0.2M IPv6) |
+| **Estimated total NLRIs** | **~100–150M** in Adj-RIB-In |
+| Telemetry | gNMI via Telegraf → InfluxDB, ~50–200 interfaces/router, 10 s interval |
+| History retention | `ip_rib_log` 4 weeks, LS logs 4 months, `peer_event_log` 1 year |
+
+The NLRI estimate (40 × ~2.5 feeds × 1.2M) places this deployment at the top
+of the OpenBMP `psql-app` guidance tier (150M NLRIs → 64 GB heap).
+
+## BMP RIB scope — recommendation
+
+**Deploy with Adj-RIB-In only.** It is the OpenBMP default, is what every
+dashboard is built on, and captures the highest-value data — what each peer
+advertises. Alternatives and their cost:
+
+- **Loc-RIB** — adds a full post-best-path converged table per router
+  (~40 × 1.2M ≈ +48M NLRIs). Add later, selectively, only where best-path
+  analysis is needed; verify the IOS-XR release supports Loc-RIB BMP.
+- **Adj-RIB-Out** — multiplies further (per advertised peer). Not recommended
+  for the initial deployment.
+- **Post-policy Adj-RIB-In** — if inbound policy is restrictive this trims
+  volume meaningfully; with permissive import it is similar to pre-policy.
+
+## Compute & memory
+
+| Component | Lab today | Production target | Rationale |
+|-----------|-----------|-------------------|-----------|
+| **Total RAM** | 31 GB | **96–128 GB** | psql-app heap 48–64 GB + PostgreSQL shared_buffers/cache + Kafka 4–8 GB + InfluxDB + Grafana + collector |
+| **CPU** | 8 cores | **16–32 vCPU** | PostgreSQL is CPU-bound under full-table churn — lab psql already sustains ~287% (3 cores) at 18 routers |
+| `psql-app` JVM heap (`MEM`) | 3 GB | **48–64 GB** | OpenBMP guidance: 4 GB ≈ 10M NLRIs, 64 GB ≈ 150M NLRIs |
+| `psql-app` container `mem_limit` | 4 GB | **heap + ~8 GB** | Set `PSQL_APP_MEM_LIMIT` above the JVM heap |
+| `psql` container `mem_limit` | 6 GB | **48–64 GB** | Set `PSQL_MEM_LIMIT`; PostgreSQL wants ~25% as `shared_buffers` and the rest for OS cache |
+| `kafka` container `mem_limit` | 4 GB | **8–12 GB** | Set `KAFKA_MEM_LIMIT`; full-table initial dumps from 40 routers are bursty |
+
+## Storage
+
+| Store | Lab today | Production target | Notes |
+|-------|-----------|-------------------|-------|
+| **PostgreSQL** | 25 GB | **2–4 TB NVMe SSD** | `ip_rib` current state (~100–150M rows) + `ip_rib_log` history (4-week retention, the dominant grower) + `base_attrs` + `geo_ip` (~7 GB fixed). OpenBMP guidance: 500 GB main + 1 TB TimescaleDB; add headroom. |
+| **Kafka** | 0.2 GB | **100–500 GB** | 12 h retention; sized for full-table initial-dump bursts × 40 routers |
+| **InfluxDB (telemetry)** | minimal | **50–200 GB** | 40 routers × ~50–200 interfaces × 10 s gNMI × 30 d; compresses well |
+| **Total** | — | **~3–5 TB fast NVMe** | Use NVMe; PostgreSQL random-IO under churn is the bottleneck on slow disks |
+
+Put the PostgreSQL data directory and the TimescaleDB tablespace on NVMe.
+`ip_rib_log` 4-week retention is the main storage tuning knob — revisit once
+production update volume is measured.
+
+## Architecture
+
+A single host is viable only if large (**≥128 GB RAM, ≥32 vCPU, multi-TB
+NVMe**). **Preferred: split services across hosts** —
+
+| Host | Services | Profile |
+|------|----------|---------|
+| **DB host** (heaviest) | postgres | — |
+| **Pipeline host** | kafka, zookeeper, collector, psql-app | core |
+| **Presentation host** | grafana, influxdb, telegraf, whois | core + telemetry |
+
+Whichever layout: every service already carries a Compose `mem_limit` — raise
+`PSQL_MEM_LIMIT` / `PSQL_APP_MEM_LIMIT` / `KAFKA_MEM_LIMIT` in `.env` for the
+production hosts.
+
+## PostgreSQL tuning
+
+- `shared_buffers` ≈ 25% of host RAM; large `effective_cache_size`.
+- Raise `work_mem` (dashboard aggregate queries) and `maintenance_work_mem`.
+- `max_wal_size` already 10 GB — keep or raise for churn bursts.
+- Enable parallel query (`max_parallel_workers_per_gather`).
+- Aggressive autovacuum on churn tables (`ip_rib`, `base_attrs`, `ip_rib_log`)
+  — applied in the lab; persist these settings in production provisioning.
+- TimescaleDB compression is already enabled on `ip_rib_log` and the `stats_*`
+  hypertables — keep it.
+
+## Reference bill of materials (single-host option)
+
+| Resource | Spec |
+|----------|------|
+| CPU | 32 vCPU |
+| RAM | 128 GB |
+| Storage | 4 TB NVMe SSD |
+| Network | 1 GbE+ to the routers' BMP source network |
+
+For the split-host option, divide per the architecture table — the DB host
+takes the bulk of RAM and all of the fast storage.