chore: drain active calls before rolling updates

api/routes/main.py

@@ -125,3 +125,22 @@ async def health() -> HealthResponse:
             STACK_PUBLISHABLE_CLIENT_KEY if is_stack else None
         ),
     )
+
+
+class ActiveCallsResponse(BaseModel):
+    active_calls: int
+
+
+@router.get("/health/active-calls", response_model=ActiveCallsResponse)
+async def active_calls() -> ActiveCallsResponse:
+    """In-flight call count for THIS worker — the drain signal for deploys.
+
+    A deploy orchestrator polls this per worker and waits for zero before
+    sending SIGTERM, because uvicorn force-closes live call WebSockets (close
+    code 1012) on SIGTERM and would cut calls mid-conversation otherwise. The
+    count is per-process: one uvicorn per VM port (scripts/rolling_update.sh)
+    or per Kubernetes pod (preStop hook). See api/services/pipecat/active_calls.py.
+    """
+    from api.services.pipecat.active_calls import active_call_count
+
+    return ActiveCallsResponse(active_calls=active_call_count())

api/services/pipecat/active_calls.py

@@ -0,0 +1,35 @@
+"""In-process registry of active pipeline runs (live voice calls).
+
+Each uvicorn worker tracks the calls it is currently running so a deploy
+orchestrator can *drain* the worker before stopping it: poll the count, wait for
+zero, then send SIGTERM. Sending SIGTERM while calls are live makes uvicorn
+force-close their WebSockets (close code 1012), which cuts the calls instead of
+letting them finish — so the wait has to happen first.
+
+The registry is deliberately per-process. That is exactly the unit that gets
+drained: one uvicorn process per VM port (see ``scripts/rolling_update.sh``) or
+one uvicorn process per Kubernetes pod (drained via a ``preStop`` hook). The
+count is exposed read-only at ``GET /api/v1/health/active-calls`` and is also a
+natural autoscaling signal (concurrent calls per worker).
+
+Access is single-threaded (asyncio event loop), so no lock is needed. A set of
+run ids — rather than a bare counter — keeps register/unregister idempotent and
+makes the in-flight runs inspectable for debugging.
+"""
+
+_active_run_ids: set[int] = set()
+
+
+def register_active_call(workflow_run_id: int) -> None:
+    """Mark a pipeline run as active in this worker."""
+    _active_run_ids.add(workflow_run_id)
+
+
+def unregister_active_call(workflow_run_id: int) -> None:
+    """Mark a pipeline run as finished in this worker."""
+    _active_run_ids.discard(workflow_run_id)
+
+
+def active_call_count() -> int:
+    """Number of pipeline runs currently active in this worker."""
+    return len(_active_run_ids)

api/services/pipecat/run_pipeline.py

@@ -11,6 +11,10 @@ from api.services.integrations import (
     IntegrationRuntimeContext,
     create_runtime_sessions,
 )
+from api.services.pipecat.active_calls import (
+    register_active_call,
+    unregister_active_call,
+)
 from api.services.pipecat.audio_config import AudioConfig, create_audio_config
 from api.services.pipecat.event_handlers import (
     register_audio_data_handler,
@@ -870,6 +874,12 @@ async def _run_pipeline(
 
     register_audio_data_handler(audio_buffer, workflow_run_id, in_memory_audio_buffer)
 
+    # Mark this run active so a deploy can drain it before stopping this worker:
+    # the orchestrator polls GET /api/v1/health/active-calls and waits for zero
+    # before sending SIGTERM (uvicorn force-closes live call WebSockets on
+    # SIGTERM, so the wait has to come first). See api/services/pipecat/
+    # active_calls.py and scripts/rolling_update.sh.
+    register_active_call(workflow_run_id)
     try:
         # Run the pipeline
         await run_pipeline_worker(task)
@@ -877,6 +887,7 @@ async def _run_pipeline(
     except asyncio.CancelledError:
         logger.warning("Received CancelledError in _run_pipeline")
     finally:
+        unregister_active_call(workflow_run_id)
         # Close MCP sessions here, not in engine.cleanup(). The anyio cancel
         # scopes opened by MCPClient.start() in engine.initialize() are
         # task-affine; this finally runs in the same task as initialize(),

api/tests/test_active_calls.py

@@ -0,0 +1,53 @@
+"""Unit tests for the per-worker active-call registry (deploy draining).
+
+The registry backs GET /api/v1/health/active-calls, which scripts/rolling_update.sh
+(and a k8s preStop hook) polls to wait for live calls to finish before stopping a
+worker. The guarantees that matter for draining: register/unregister are
+idempotent, and the count only reaches zero when every registered run is gone.
+"""
+
+from api.services.pipecat import active_calls
+
+
+def setup_function():
+    # Module-level state — start each test from an empty registry.
+    active_calls._active_run_ids.clear()
+
+
+def test_starts_empty():
+    assert active_calls.active_call_count() == 0
+
+
+def test_register_counts_distinct_runs():
+    active_calls.register_active_call(1)
+    active_calls.register_active_call(2)
+    assert active_calls.active_call_count() == 2
+
+
+def test_register_is_idempotent():
+    # Registering the same run twice must not double-count, or the count could
+    # never drain to zero.
+    active_calls.register_active_call(1)
+    active_calls.register_active_call(1)
+    assert active_calls.active_call_count() == 1
+
+
+def test_unregister_removes_run():
+    active_calls.register_active_call(1)
+    active_calls.register_active_call(2)
+    active_calls.unregister_active_call(1)
+    assert active_calls.active_call_count() == 1
+
+
+def test_unregister_unknown_run_is_a_noop():
+    # discard() semantics: unregistering a run that was never registered (or was
+    # already removed) is safe and cannot push the count negative.
+    active_calls.unregister_active_call(999)
+    assert active_calls.active_call_count() == 0
+
+
+def test_full_lifecycle_drains_to_zero():
+    active_calls.register_active_call(42)
+    assert active_calls.active_call_count() == 1
+    active_calls.unregister_active_call(42)
+    assert active_calls.active_call_count() == 0

scripts/rolling_update.sh

@@ -28,6 +28,7 @@ NGINX_UPSTREAM_TEMPLATE="$BASE_DIR/nginx/dograh_upstream.conf.template"
 NGINX_UPSTREAM_CONF="/etc/nginx/conf.d/dograh_upstream.conf"
 
 HEALTH_CHECK_ENDPOINT="/api/v1/health"
+ACTIVE_CALLS_ENDPOINT="/api/v1/health/active-calls"
 
 # Load environment
 if [[ -f "$ENV_FILE" ]]; then
@@ -40,7 +41,9 @@ FASTAPI_WORKERS=${FASTAPI_WORKERS:-$CPU_CORES}
 ARQ_WORKERS=${ARQ_WORKERS:-1}
 
 # Tuning knobs (override via environment)
-DRAIN_TIMEOUT=${DRAIN_TIMEOUT:-300}          # seconds to wait for old workers to drain
+DRAIN_TIMEOUT=${DRAIN_TIMEOUT:-300}          # seconds to wait for active calls to finish
+DRAIN_INTERVAL=${DRAIN_INTERVAL:-5}          # seconds between active-call drain polls
+STOP_TIMEOUT=${STOP_TIMEOUT:-30}             # seconds to wait for drained workers to exit after SIGTERM
 HEALTH_MAX_ATTEMPTS=${HEALTH_MAX_ATTEMPTS:-30}  # per-worker health-check retries
 HEALTH_INTERVAL=${HEALTH_INTERVAL:-2}        # seconds between health-check retries
 
@@ -96,6 +99,20 @@ kill_process_tree() {
   fi
 }
 
+# Active in-progress call count for a single worker, via its health endpoint.
+# A worker that is unreachable (already exited) or returns no/garbage body
+# reports 0, so it never blocks the drain.
+count_active_calls_on_port() {
+  local port=$1
+  local body n
+  body=$(curl -s --max-time 3 \
+    "http://127.0.0.1:${port}${ACTIVE_CALLS_ENDPOINT}" 2>/dev/null || true)
+  n=$(printf '%s' "$body" \
+    | grep -o '"active_calls"[[:space:]]*:[[:space:]]*[0-9]\+' \
+    | grep -o '[0-9]\+$' || true)
+  printf '%s' "${n:-0}"
+}
+
 ###############################################################################
 ### ROLLBACK
 ###############################################################################
@@ -366,9 +383,46 @@ log_info "nginx reloaded — traffic now routed to band $NEW_BAND"
 ### PHASE 5: DRAIN OLD WORKERS
 ###############################################################################
 
-log_info "=== Phase 5: Draining old workers (band $OLD_BAND, timeout ${DRAIN_TIMEOUT}s) ==="
+# nginx (Phase 4) already routes new calls to the new band, so the old band only
+# holds calls still in progress. Wait for those to finish BEFORE signalling the
+# workers: SIGTERM makes uvicorn force-close live call WebSockets (close code
+# 1012), cutting calls mid-conversation. So we poll each old worker's in-flight
+# call count and only stop once it reaches zero (or DRAIN_TIMEOUT elapses).
 
-# Collect old worker PIDs
+log_info "=== Phase 5a: Draining active calls from band $OLD_BAND (timeout ${DRAIN_TIMEOUT}s) ==="
+
+drain_start=$(date +%s)
+while true; do
+  active=0
+  for ((w = 0; w < FASTAPI_WORKERS; w++)); do
+    port=$((OLD_BASE + w))
+    # Only poll workers still alive; an exited worker holds no calls.
+    pidfile="$RUN_DIR/uvicorn_${port}.pid"
+    if [[ -f "$pidfile" ]] && kill -0 "$(<"$pidfile")" 2>/dev/null; then
+      active=$((active + $(count_active_calls_on_port "$port")))
+    fi
+  done
+
+  if [[ $active -eq 0 ]]; then
+    log_info "Band $OLD_BAND fully drained — no active calls"
+    break
+  fi
+
+  elapsed=$(( $(date +%s) - drain_start ))
+  if [[ $elapsed -ge $DRAIN_TIMEOUT ]]; then
+    log_warn "Drain timeout reached (${DRAIN_TIMEOUT}s) with $active active call(s) still running — stopping anyway."
+    break
+  fi
+
+  log_info "  Waiting for $active active call(s) to finish... (${elapsed}s / ${DRAIN_TIMEOUT}s)"
+  sleep "$DRAIN_INTERVAL"
+done
+
+log_info "=== Phase 5b: Stopping old workers (band $OLD_BAND, timeout ${STOP_TIMEOUT}s) ==="
+
+# Calls are drained — now signal the workers and reap them. A drained worker
+# exits within a second or two of SIGTERM; STOP_TIMEOUT bounds stragglers (e.g.
+# a call that outlived DRAIN_TIMEOUT) before we force-kill.
 OLD_PIDS=()
 for ((w = 0; w < FASTAPI_WORKERS; w++)); do
   port=$((OLD_BASE + w))
@@ -385,7 +439,7 @@ for ((w = 0; w < FASTAPI_WORKERS; w++)); do
 done
 
 if [[ ${#OLD_PIDS[@]} -gt 0 ]]; then
-  start_time=$(date +%s)
+  stop_start=$(date +%s)
 
   while true; do
     all_dead=true
@@ -397,13 +451,13 @@ if [[ ${#OLD_PIDS[@]} -gt 0 ]]; then
     done
 
     if $all_dead; then
-      log_info "All old workers exited gracefully"
+      log_info "All old workers exited"
       break
     fi
 
-    elapsed=$(( $(date +%s) - start_time ))
-    if [[ $elapsed -ge $DRAIN_TIMEOUT ]]; then
-      log_warn "Drain timeout reached (${DRAIN_TIMEOUT}s). Force-killing remaining old workers."
+    elapsed=$(( $(date +%s) - stop_start ))
+    if [[ $elapsed -ge $STOP_TIMEOUT ]]; then
+      log_warn "Stop timeout reached (${STOP_TIMEOUT}s). Force-killing remaining old workers."
       for pid in "${OLD_PIDS[@]}"; do
         if kill -0 "$pid" 2>/dev/null; then
           kill_process_tree "$pid" "-KILL"
@@ -414,11 +468,11 @@ if [[ ${#OLD_PIDS[@]} -gt 0 ]]; then
       break
     fi
 
-    log_info "  Waiting for old workers to drain... (${elapsed}s / ${DRAIN_TIMEOUT}s)"
-    sleep 5
+    log_info "  Waiting for old workers to exit... (${elapsed}s / ${STOP_TIMEOUT}s)"
+    sleep 2
   done
 else
-  log_warn "No old worker PIDs to drain"
+  log_warn "No old worker PIDs to stop"
 fi
 
 ###############################################################################