Fixed prevent smoke builds and keep SSE streams alive

.woodpecker/02-build-push.yml

@@ -4,8 +4,9 @@
 # here so builds are gated on a green test run.
 
 when:
-  event: [push, manual]
+  - event: [push, manual]
-  branch: [dev, stage, main]
+    branch: [dev, stage, main]
+    evaluate: 'E2E_CONVERT_JETSON != "1" && E2E_SMOKE_TEST != "1"'
 
 # Multi-arch matrix. detections is the only repo today with split per-arch
 # Dockerfiles (Jetson uses Dockerfile.jetson with TensorRT/CUDA on L4T;

e2e/tests/test_performance.py

@@ -17,7 +17,7 @@ def _percentile_ms(sorted_ms, p):
     return sorted_ms[lo] * (1 - w) + sorted_ms[hi] * w
 
 
-@pytest.mark.timeout(60)
+@pytest.mark.timeout(240)
 def test_nft_perf_01_single_image_latency_p95(
     warm_engine, image_detect, image_small
 ):
@@ -39,7 +39,7 @@ def test_nft_perf_01_single_image_latency_p95(
     assert p95 < threshold_ms
 
 
-@pytest.mark.timeout(60)
+@pytest.mark.timeout(120)
 def test_nft_perf_03_tiling_overhead_large_image(
     warm_engine, image_detect, image_small, image_large
 ):

src/main.py

@@ -415,8 +415,11 @@ async def detect_events(channel_id: str, request: Request, after_ts: Optional[in
                     if ts_ms > after_ts:
                         yield f"id: {ts_ms}\ndata: {data}\n\n"
             while True:
-                ts_ms, data = await queue.get()
+                try:
-                yield f"id: {ts_ms}\ndata: {data}\n\n"
+                    ts_ms, data = await asyncio.wait_for(queue.get(), timeout=1.0)
+                    yield f"id: {ts_ms}\ndata: {data}\n\n"
+                except asyncio.TimeoutError:
+                    yield ": heartbeat\n\n"
         except asyncio.CancelledError:
             pass
         finally: