Files
microdao-daarion/ops/offload_routing.md
Apple c4b94a327d P2.2+P2.3: NATS offload node-worker + router offload integration
Node Worker (services/node-worker/):
- NATS subscriber for node.{NODE_ID}.llm.request / vision.request
- Canonical JobRequest/JobResponse envelope (Pydantic)
- Idempotency cache (TTL 10min) with inflight dedup
- Deadline enforcement (DEADLINE_EXCEEDED on expired jobs)
- Concurrency limiter (semaphore, returns busy)
- Ollama + Swapper vision providers

Router offload (services/router/offload_client.py):
- NATS req/reply with configurable retries
- Circuit breaker per node+type (3 fails/60s → open 120s)
- Concurrency semaphore for remote requests

Model selection (services/router/model_select.py):
- exclude_nodes parameter for circuit-broken nodes
- force_local flag for fallback re-selection
- Integrated circuit breaker state awareness

Router /infer pipeline:
- Remote offload path when NCS selects remote node
- Automatic fallback: exclude failed node → force_local re-select
- Deadline propagation from router to node-worker

Tests: 17 unit tests (idempotency, deadline, circuit breaker)
Docs: ops/offload_routing.md (subjects, envelope, verification)
Made-with: Cursor
2026-02-27 02:44:05 -08:00

145 lines
4.4 KiB
Markdown

# NATS Offload Routing — Operations Guide
## Architecture
```
Router (NODA1/NODA2)
├── model_select.py → selects best model from global capabilities pool
│ (local first, remote if needed, circuit breaker aware)
├── offload_client.py → NATS req/reply to remote node-worker
│ (retries, deadlines, circuit breaker)
└── global_capabilities_client.py → scatter-gather discovery
node.*.capabilities.get
```
## NATS Subjects
| Subject | Direction | Description |
|---------|-----------|-------------|
| `node.{node_id}.capabilities.get` | req/reply | NCS capabilities query |
| `node.{node_id}.llm.request` | req/reply | LLM inference offload |
| `node.{node_id}.vision.request` | req/reply | Vision inference offload |
| `node.{node_id}.stt.request` | req/reply | STT (scaffold) |
| `node.{node_id}.tts.request` | req/reply | TTS (scaffold) |
## Job Request Envelope
```json
{
"job_id": "uuid",
"trace_id": "uuid",
"actor_agent_id": "sofiia",
"target_agent_id": "helion",
"required_type": "llm",
"deadline_ts": 1740000000000,
"idempotency_key": "uuid",
"payload": {
"messages": [{"role": "user", "content": "..."}],
"prompt": "...",
"model": "qwen3:14b",
"max_tokens": 2048,
"temperature": 0.2
},
"hints": {
"prefer_models": ["qwen3:14b"]
}
}
```
## Job Response
```json
{
"job_id": "uuid",
"trace_id": "uuid",
"node_id": "noda2",
"status": "ok|busy|timeout|error",
"provider": "ollama",
"model": "qwen3:14b",
"latency_ms": 5500,
"result": {"text": "..."},
"error": null,
"cached": false
}
```
## Circuit Breaker
Per `node_id:required_type`. Opens after 3 failures in 60s, stays open 120s.
| Env | Default | Description |
|-----|---------|-------------|
| `ROUTER_OFFLOAD_CB_FAILS` | 3 | Failures to trip |
| `ROUTER_OFFLOAD_CB_WINDOW_S` | 60 | Failure window |
| `ROUTER_OFFLOAD_CB_OPEN_S` | 120 | Open duration |
## Idempotency
Node-worker caches `idempotency_key → response` for 10 minutes.
Duplicate requests return cached response immediately (< 10ms).
Inflight dedup prevents parallel execution of same job.
## Deadline Enforcement
`deadline_ts` is absolute Unix milliseconds. If already expired on arrival,
node-worker returns `status=timeout` + `error.code=DEADLINE_EXCEEDED`.
During inference, `asyncio.wait_for` enforces remaining time.
## Fallback Chain
1. NCS selects remote model (e.g. noda2)
2. Router sends NATS offload request
3. If remote fails (timeout/busy/error):
- Record circuit breaker failure
- Re-select with `exclude_nodes={failed}` + `force_local=True`
- Execute locally
## Environment Variables
### Router
| Env | Default | Description |
|-----|---------|-------------|
| `ROUTER_INFER_TIMEOUT_MS` | 25000 | Total inference deadline |
| `ROUTER_OFFLOAD_RETRIES` | 1 | NATS retry on transient |
| `ROUTER_OFFLOAD_MAX_CONCURRENCY_REMOTE` | 8 | Max parallel offloads |
### Node Worker
| Env | Default | Description |
|-----|---------|-------------|
| `NODE_ID` | noda2 | Canonical node identifier |
| `NATS_URL` | nats://dagi-nats:4222 | NATS server |
| `OLLAMA_BASE_URL` | http://host.docker.internal:11434 | Ollama API |
| `SWAPPER_URL` | http://swapper-service:8890 | Swapper API |
| `NODE_DEFAULT_LLM` | qwen3:14b | Default LLM model |
| `NODE_DEFAULT_VISION` | llava:13b | Default vision model |
| `NODE_WORKER_MAX_CONCURRENCY` | 2 | Max parallel inferences |
## Verification Commands
```bash
# Node Worker health
curl -s http://localhost:8109/healthz | jq .
# Direct NATS LLM offload
nats request node.noda2.llm.request '{"job_id":"test","required_type":"llm","payload":{"prompt":"hi","max_tokens":10}}'
# Idempotency test (same job_id)
nats request node.noda2.llm.request '{"job_id":"test","idempotency_key":"test","required_type":"llm","payload":{"prompt":"hi"}}' # → cached=true
# Deadline expired test
nats request node.noda2.llm.request '{"job_id":"expired","required_type":"llm","deadline_ts":1000,"payload":{"prompt":"hi"}}' # → DEADLINE_EXCEEDED
# Router logs (selection + offload)
docker logs dagi-router-node2 2>&1 | grep -E '\[select\]|\[offload\]|\[fallback\]'
```
## Adding New Nodes
1. Deploy `node-capabilities` + `node-worker` with unique `NODE_ID`
2. Connect NATS (leafnode or cluster member)
3. Router auto-discovers via `node.*.capabilities.get` scatter-gather
4. No config changes on Router needed