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
This commit is contained in:
Apple
2026-02-27 02:44:05 -08:00
parent a92c424845
commit c4b94a327d
19 changed files with 1075 additions and 6 deletions

View File

@@ -0,0 +1,34 @@
"""Tests for node-worker deadline handling."""
import sys
import os
import time
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "services", "node-worker"))
from models import JobRequest
def test_remaining_ms_with_future_deadline():
job = JobRequest(deadline_ts=int(time.time() * 1000) + 10000)
rem = job.remaining_ms()
assert 9000 <= rem <= 10500
def test_remaining_ms_with_past_deadline():
job = JobRequest(deadline_ts=int(time.time() * 1000) - 5000)
assert job.remaining_ms() == 0
def test_remaining_ms_no_deadline():
job = JobRequest(deadline_ts=0)
assert job.remaining_ms() == 30_000
def test_effective_idem_key_uses_idempotency_key():
job = JobRequest(job_id="j1", idempotency_key="custom-key")
assert job.effective_idem_key() == "custom-key"
def test_effective_idem_key_falls_back_to_job_id():
job = JobRequest(job_id="j2", idempotency_key="")
assert job.effective_idem_key() == "j2"

View File

@@ -0,0 +1,66 @@
"""Tests for node-worker idempotency store."""
import asyncio
import sys
import os
import time
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "services", "node-worker"))
from models import JobResponse, JobError
from idempotency import IdempotencyStore
def test_put_and_get():
store = IdempotencyStore()
resp = JobResponse(job_id="j1", status="ok", node_id="n1", provider="ollama", model="qwen3:14b")
store.put("key1", resp)
cached = store.get("key1")
assert cached is not None
assert cached.status == "ok"
assert cached.cached is True
assert cached.job_id == "j1"
def test_miss():
store = IdempotencyStore()
assert store.get("nonexistent") is None
def test_ttl_expiry():
store = IdempotencyStore()
resp = JobResponse(job_id="j2", status="ok", node_id="n1")
store.put("key2", resp)
store._cache["key2"] = (resp, time.time() - 1)
assert store.get("key2") is None
def test_timeout_shorter_ttl():
store = IdempotencyStore()
resp = JobResponse(job_id="j3", status="timeout", error=JobError(code="TO"))
store.put("key3", resp)
_, expires = store._cache["key3"]
assert expires - time.time() < 35 # timeout TTL ≈ 30s
def test_inflight_dedup():
store = IdempotencyStore()
async def run():
fut1 = await store.acquire_inflight("key4")
assert fut1 is None # first caller gets None (becomes processor)
fut2 = await store.acquire_inflight("key4")
assert fut2 is not None # second caller gets future to wait on
resp = JobResponse(job_id="j4", status="ok")
store.complete_inflight("key4", resp)
result = await asyncio.wait_for(fut2, timeout=1.0)
assert result.status == "ok"
asyncio.run(run())
def test_evict_on_max_size():
store = IdempotencyStore(max_size=3)
for i in range(5):
store.put(f"k{i}", JobResponse(job_id=f"j{i}", status="ok"))
assert len(store._cache) <= 3

View File

@@ -0,0 +1,52 @@
"""Tests for offload_client circuit breaker."""
import sys
import os
import time
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "services", "router"))
import offload_client
def setup_function():
offload_client._breakers.clear()
def test_node_available_by_default():
assert offload_client.is_node_available("noda99", "llm") is True
def test_node_unavailable_after_threshold():
for _ in range(offload_client.CB_FAILS):
offload_client.record_failure("noda99", "llm")
assert offload_client.is_node_available("noda99", "llm") is False
def test_node_available_after_open_expires():
for _ in range(offload_client.CB_FAILS):
offload_client.record_failure("noda99", "llm")
offload_client._breakers["noda99:llm"]["open_until"] = time.time() - 1
assert offload_client.is_node_available("noda99", "llm") is True
def test_success_resets_breaker():
for _ in range(offload_client.CB_FAILS):
offload_client.record_failure("noda99", "llm")
offload_client._breakers["noda99:llm"]["open_until"] = time.time() - 1
offload_client.record_success("noda99", "llm")
assert offload_client.is_node_available("noda99", "llm") is True
def test_get_unavailable_nodes():
for _ in range(offload_client.CB_FAILS):
offload_client.record_failure("bad_node", "llm")
unavail = offload_client.get_unavailable_nodes("llm")
assert "bad_node" in unavail
assert offload_client.get_unavailable_nodes("vision") == set()
def test_different_types_independent():
for _ in range(offload_client.CB_FAILS):
offload_client.record_failure("noda99", "llm")
assert offload_client.is_node_available("noda99", "llm") is False
assert offload_client.is_node_available("noda99", "vision") is True