verae-nats-cluster/results/20260912T055851Z/optimal-config.md

**Progress report — optimal configuration study** · `20260912T055851Z` (UTC) · all code on **NS1.GEORGELAMBERT.ORG** (`70.88.205.138`)

This document folds every ladder we have run (1-core ZFS, NS1-orchestrated, tmpfs maximize, and this exhaustive 8c/16G **ZFS** factorial) plus UDP / MQTT / reconnect probes. It recommends a lab config and a **three-box HP DL360 Gen10** projection. veth/10G was not changed.

---

## 1. Verdict (read this first)

**Keep NATS + JetStream.** Do not replace the fabric with MQTT, UDP, or a custom persistent-socket protocol for Verae jobs/events/archive. Those are either slower, less durable, or already what NATS is.

**Lab (NS1, one host, three LXC) — optimal now**

| Stream | Storage | Replicas | Why |
|--------|---------|----------|-----|
| `ZAPIER_JOBS`, `ZAPIER_WEBHOOKS`, `VERAE_ARCHIVE` | **file** (ZFS) | **3** | Survive a nats LXC death; archive must persist |
| `ZAPIER_EVENTS` | **memory** | **3** | Waiters are latency-sensitive; events rebuild from job status |
| `ZAPIER_USAGE` | file | 3 | Telemetry, limits + max-age |

Keep **8 cores / 16 GiB / `max_mem: 8G`** on 510–513 (already live). Do **not** leave JetStream on tmpfs. Do **not** drop product streams to r=1. Reuse **one NATS connection per process** (already true in middleware); never connect-per-message.

**Metal (3× DL360 Gen10) — optimal later**

Same stream table. File store on **local NVMe/M.2**, not a shared SAN. Cluster + client on **10GbE** (or 25GbE if you already have it). Dual Gold Xeon is surplus CPU for this workload; 8–16 cores dedicated to `nats-server` is enough. Expected JS file r=3: **~40–80k** 128 B pubs/s (about **3–6×** this lab’s 8c ZFS 1p, **2–4×** tmpfs 1p) — bounded by **10GbE replica RTT**, not by Xeon clocks. Core NATS will sit in the **1–3M msgs/s** band until the NIC saturates (~9 Gbit/s ≈ 8–9M × 128 B theoretical; CPU and client will hit first).

---

## 2. What we actually ran (this exhaustive pass)

Live cluster during this run: LXC 510–513 **8 cores / 16 GiB**, JetStream **on ZFS** (tmpfs from the maximize study was already unmounted). Extra factorial: file/memory × replicas 1/3, 4 KiB file r=3, reconnect-per-message ping, UDP echo 510→511, MQTT QoS0 against nats-a `:1883`. Product streams were not the bench target.

### 2.1 Cross-study history

| Study | Env | Core 1p pub | JS file r=3 1p | JS mem r=3 4p | Ping p99 |
| --- | --- | --- | --- | --- | --- |
| `20260912T051237Z` | 1c/1G ZFS (NS1 orch.) | 502,502 | 7,393 | — | 1.377ms |
| `20260912T053120Z` | 8c/16G tmpfs + mem extra | 599,004 | 17,388 | 36,355 | 0.684ms |
| `20260912T055851Z` | 8c/16G ZFS exhaustive `20260912T055851Z` | 662,227 | 14,330 | 37,736 | 1.140ms |

![JS 1p file r=3 history](charts-optimal/history-js1p.png)

### 2.2 This run — JetStream factorial

| Run | What | Pub msgs/s | Pub MB/s |
| --- | --- | --- | --- |
| `js-file-1p-20k-128-r1` | file r=1 1p 128 B | 18,888 | 2.31 |
| `js-file-4p-50k-128-r1` | file r=1 4p 128 B | 24,560 | 3.00 |
| `js-1p-20k-128-r3` | file r=3 1p 128 B | 14,330 | 1.75 |
| `js-4p-50k-128-r3` | file r=3 4p 128 B | 19,232 | 2.35 |
| `js-4p-20k-1k-r3` | file r=3 4p 1 KiB | 15,197 | 14.84 |
| `js-file-1p-20k-4k-r3` | file r=3 1p 4 KiB | 8,673 | 33.88 |
| `js-mem-1p-20k-128-r1` | memory r=1 1p 128 B | 29,972 | 3.66 |
| `js-mem-4p-50k-128-r1` | memory r=1 4p 128 B | 64,923 | 7.93 |
| `js-mem-1p-20k-128-r3` | memory r=3 1p 128 B | 20,188 | 2.46 |
| `js-mem-4p-50k-128-r3` | memory r=3 4p 128 B | 37,736 | 4.61 |
| `js-mem-4p-20k-1k-r3` | memory r=3 4p 1 KiB | 33,916 | 33.12 |

Replica **1 vs 3** on this stand (file 1p 128 B): r=1 is 18,888 vs r=3 14,330 (1.32× if r=3 is the slower one). Memory r=1 1p 29,972 vs memory r=3 20,188.

![Replica cost](charts-optimal/replicas.png)

### 2.3 Delay, reconnect tax, UDP, MQTT

| Probe | Result | Meaning |
|-------|--------|---------|
| NATS ping (persistent sockets) p50 / p99 | 0.456ms / 1.140ms | Quiet hop with a long-lived TCP conn |
| NATS **reconnect-per-message** p50 / p99 | 0.503ms / 1.750ms | TCP+NATS handshake on every pub — this is the tax to avoid |
| UDP echo 510→511 p99 | 0.363ms | Raw datagram ceiling on the same veth (no NATS) |
| MQTT QoS0 5k×128 B | 44862 pubs/s | nats-server MQTT gateway on `:1883` |

Core 1p1s 128 B this run: 662,227 pub msgs/s. Flood delay is still backlog/consume_rate, not RTT.

---

## 3. Alternative transports (why we are not switching the fabric)

NATS already **is** persistent TCP sockets with a tiny binary protocol, automatic reconnect, and optional JetStream durability. “Reduce connection overhead” is a **client** discipline: hold the connection. The reconnect probe exists to prove that opening a socket per job would dominate ping RTT.

| Idea | Fit for Verae jobs/events/archive | Throughput vs NATS core | Durability |
|------|-----------------------------------|-------------------------|------------|
| **NATS core pub/sub** | Fan-out, request-reply (`verae.billing.*`) | Highest we measured (~0.5–2M msgs/s) | None |
| **NATS JetStream file r=3** | Jobs, webhooks, archive | ~8–23k on this lab; see metal projection | Disk + 1-node loss |
| **NATS JetStream memory r=3** | Events mailbox | ~22–36k on this lab | RAM + 1-node loss; **empty on full restart** |
| **MQTT** (NATS gateway or Mosquitto) | IoT endpoints that already speak MQTT | This probe: 44862 pubs/s QoS0 — typically **well below** NATS core; QoS1 ≈ JetStream-ish with more chatter | QoS1/2 session state; not our WORM model |
| **UDP** | Telemetry that may drop | RTT 0.363ms p99 — fastest hop, **no** reliability, no cluster, no auth | None |
| **Custom persistent sockets / HTTP long-poll** | Worse NATS | You would re-implement reconnect, flow control, and fan-out | DIY |
| **WebSocket** | Browsers only | Extra framing; NATS already has WS for UIs, not for middleware | Same as core/JS behind it |
| **QUIC / WebTransport** | Lossy WAN / browsers | NATS QUIC is not the lab path; 10GbE LAN does not need it | Same |
| **Kafka / Redis streams** | Heavy log replay | Higher ops cost; not on `vmbr1` today | Yes, heavier |

**MQTT:** NATS documents MQTT as an *enabling* gateway for existing IoT, and prefers NATS end-to-end for greenfield. Zapier cloud never talks NATS or MQTT; it talks HTTPS. Putting MQTT in the middle of timestamp jobs adds protocol translation and QoS timers without helping `jobId → events`. Use MQTT only if a device already cannot speak NATS.

**UDP:** Fine as a *measurement* of veth RTT. Unusable as the job fabric (no ack, no replica, no flow control). NATS ping is already within a small multiple of UDP on this bridge.

**Persistence sockets:** Middleware and keep already keep `NATS_URL` connections open. Optimal: one connection (or a small pool) per process, `max_reconnect`, jitter, no `connect()` in the per-job path. The reconnect ladder is the anti-pattern.

---

## 4. Optimal configurations

### 4.1 NS1 lab (now)

1. **Leave 8 cores / 16 GiB** on nats-a/b/c and the worker. Host has 40 cores / 377 GiB; this is cheap.
2. **`max_mem: 8G`** stays. Required for memory streams.
3. **File r=3 on ZFS** for jobs/webhooks/archive. tmpfs doubled JS 1p (7.4k→17k) but **loses the stream on reboot** — unacceptable for archive.
4. **Memory r=3 for `ZAPIER_EVENTS`** if we accept “all three nats CTs reboot ⇒ in-flight waiters fall back to HTTP poll.” That matches the designed wait path (`GET /api/status/{jobId}`).
5. **r=1 only for throwaway benches**, never product streams. Replica=3 is the point of three guests.
6. **veth on vmbr1, no fake 10G NICs.** Already 10000Mb/s; JS does not fill it.
7. **Pin cpusets** later if keep/fleet steal; not required to beat these numbers.
8. Clients: persistent NATS connections; pull consumers with bounded `max_ack_pending` for webhooks.

### 4.2 Three HP DL360 Gen10 (projection — not measured)

Assumed bill of materials (state it in the buy):

| Piece | Assumption |
|-------|------------|
| Chassis | 3× DL360 Gen10 1U |
| CPU | Dual 2nd-gen Xeon **Gold** (e.g. 6226R 16c or 6248 20c — **32–40 cores/box**) |
| Memory | DDR4-2933, **192–384 GiB**/box (6–12×32 GiB); NATS will not use most of it |
| Storage | **NVMe M.2 or U.2** for `/var/lib/nats/jetstream` (XFS or ext4, **not** shared ZFS over the network). RAID1 of two NVMe if you want disk HA *inside* a box |
| Network | **10GbE** (FlexibleLOM or PCIe); dedicated VLAN for `:4222`+`:6222`. Do not share with public `vmbr0` traffic |
| OS | Debian/Ubuntu bare metal, `nats-server` systemd, same `nats.conf` as lab (bind private IP only) |

**What changes vs NS1 LXC**

| Factor | NS1 today | 3× DL360 | Effect on JS file r=3 |
|--------|-----------|----------|------------------------|
| Failure domain | 1 Proxmox host | 3 chassis, 3 NVMe, 3 NICs | r=3 **means** something |
| Disk | Shared ZFS SSD2 | Local NVMe fsync ~50–150 µs | Big win vs ZFS; similar to tmpfs for sequential 128 B |
| Replica path | veth/bridge (~µs–tens of µs) | 10GbE RTT typically **50–200 µs** | **Slower than same-host tmpfs**, faster than a bad SAN |
| CPU | 8 of 40 shared | 32–40 dedicated Gold cores | Headroom for many clients, not 10× JS |
| NIC | software 10G veth, already ~5 Gbit/s core | real 10GbE ~9 Gbit/s TCP | Core NATS can grow; JS r=3 stays replica-bound |

**Projected bands** (128 B, 3-node cluster, dedicated 10GbE, local NVMe, 8+ cores pinned to nats-server):

| Workload | NS1 measured (best) | DL360 projection | Confidence |
|----------|---------------------|------------------|------------|
| Core pub/sub 1p | 0.5–0.8M | **0.8–2M** | Medium — NIC + syscall, plenty of CPU |
| Core 4p4s 1 KiB | ~0.6–0.7M (~0.6 GB/s) | **~1M msgs/s / ~1 GB/s** approaching 10GbE | Medium |
| JS file r=1 | this run r=1 | **80–200k** pubs/s | Medium — NVMe + no replica wait |
| JS file r=3 | 7–23k (ZFS/tmpfs) | **40–80k** pubs/s | Medium-low — replica RTT dominates; 3 NVMe still help vs shared ZFS |
| JS memory r=3 | 22–36k | **50–100k** | Medium-low — RAM + 10GbE ack |
| Ping p99 | 0.7–1.4 ms | **0.2–0.6 ms** | Medium — real NIC but no Proxmox tax |

These are **not** DL360 measurements. Scale from: (a) our replica-1 vs replica-3 ratio once this run’s r=1 numbers exist, (b) tmpfs vs ZFS ratio (2.35× on 1p), (c) Synadia/nats bench async file r=1 ~100–400k on NVMe loopback, derated for 10GbE RTT.

**Buy notes:** M.2 via Dual uFF / enablement kit; put JetStream on NVMe **directly**, not behind a RAID controller write-through unless you measure. 1GbE onboard is a trap — use 10GbE for `:6222`. Dual Gold is for isolation (nats vs worm/tree vs OS), not because JS needs 56 cores.

---

## 5. What we are not doing

- MQTT as the Zapier or middleware transport.
- UDP for jobs.
- Emulated 10G fiber NICs on LXC.
- tmpfs as the production store.
- r=1 for product streams.
- Connect-per-job.

Re-run exhaustive: `bash scripts/exhaustive-ns1-study.sh` on NS1.