System Health Checks (#12193)

* Add worker health check invoke task

* Increase frequency of heartbeat task

* Adjust default threshold for worker health check

* Add server_health invoke func

src/backend/InvenTree/InvenTree/apps.py

@@ -188,13 +188,10 @@ class InvenTreeConfig(AppConfig):
     @ignore_ready_warning
     def add_heartbeat(self):
         """Ensure there is at least one background task in the queue."""
-        import django_q.models
-
         try:
-            if django_q.models.OrmQ.objects.count() == 0:
-                InvenTree.tasks.offload_task(
-                    InvenTree.tasks.heartbeat, force_async=True, group='heartbeat'
-                )
+            InvenTree.tasks.offload_task(
+                InvenTree.tasks.heartbeat, force_async=True, group='heartbeat'
+            )
         except AppRegistryNotReady:  # pragma: no cover
             pass
         except Exception:

src/backend/InvenTree/InvenTree/tasks.py

@@ -438,21 +438,30 @@ def scheduled_task(
 
 
 @tracer.start_as_current_span('heartbeat')
-@scheduled_task(ScheduledTask.MINUTES, 5)
+@scheduled_task(ScheduledTask.MINUTES, 1)
 def heartbeat():
-    """Simple task which runs at 5 minute intervals, so we can determine that the background worker is actually running.
-
-    (There is probably a less "hacky" way of achieving this)?
-    """
+    """Simple task which runs at 1 minute intervals, so we can determine that the background worker is actually running."""
     try:
         from django_q.models import OrmQ, Success
     except AppRegistryNotReady:  # pragma: no cover
         logger.info('Could not perform heartbeat task - App registry not ready')
         return
 
-    threshold = timezone.now() - timedelta(minutes=30)
+    # Write a timestamp file so that health checks can verify worker liveness
+    # without needing to start a full Django process.
+    import tempfile
+    from pathlib import Path
 
-    # Delete heartbeat results more than half an hour old,
+    try:
+        Path(tempfile.gettempdir()).joinpath('inventree_worker_heartbeat').write_text(
+            str(timezone.now().timestamp())
+        )
+    except Exception:
+        pass
+
+    threshold = timezone.now() - timedelta(minutes=15)
+
+    # Delete heartbeat results more than 15 minutes old,
     # otherwise they just create extra noise
     heartbeats = Success.objects.filter(
         func='InvenTree.tasks.heartbeat', started__lte=threshold

tasks.py

@@ -1603,6 +1603,67 @@ def worker(c, verbose: bool = False):
     manage(c, 'qcluster', pty=True, verbose=verbose)
 
 
+@task(help={'timeout': 'Maximum minutes since last heartbeat (default: 3)'})
+def worker_health(c, timeout: int = 3):
+    """Check if the background worker is healthy by reading the heartbeat file.
+
+    Exits 0 if the worker has run within the last TIMEOUT minutes, 1 otherwise.
+    No Django startup or database access is required.
+    """
+    heartbeat_file = Path(tempfile.gettempdir()) / 'inventree_worker_heartbeat'
+
+    if heartbeat_file.exists():
+        try:
+            age_seconds = time.time() - float(heartbeat_file.read_text().strip())
+            if age_seconds < timeout * 60:
+                success(
+                    f'Worker is healthy (last heartbeat {int(age_seconds) // 60}m {int(age_seconds) % 60}s ago)'
+                )
+                return
+            warning(
+                f'Heartbeat file is stale ({int(age_seconds) // 60}m {int(age_seconds) % 60}s old)'
+            )
+        except Exception as e:
+            warning(f'Could not read heartbeat file: {e}')
+    else:
+        warning(f'Heartbeat file not found: {heartbeat_file}')
+
+    error('Worker health check failed')
+    raise Exit(code=1)
+
+
+@task(
+    help={
+        'address': 'Server address to check (default: http://localhost:8000)',
+        'timeout': 'Request timeout in seconds (default: 5)',
+    }
+)
+def server_health(c, address: str = 'http://localhost:8000', timeout: int = 5):
+    """Check if the web server is healthy by requesting /api/system/health/.
+
+    Exits 0 on HTTP 200, 1 otherwise.
+    No Django startup required.
+    """
+    import urllib.error
+    import urllib.request
+
+    url = f'{address.rstrip("/")}/api/system/health/'
+
+    try:
+        with urllib.request.urlopen(url, timeout=timeout) as response:
+            if response.status == 200:
+                success(f'Server is healthy ({url})')
+                return
+            warning(f'Unexpected status {response.status} from {url}')
+    except urllib.error.URLError as e:
+        warning(f'Could not reach server at {url}: {e.reason}')
+    except Exception as e:
+        warning(f'Unexpected error checking {url}: {e}')
+
+    error('Server health check failed')
+    raise Exit(code=1)
+
+
 @task(post=[static, server])
 def test_translations(c):
     """Add a fictional language to test if each component is ready for translations."""
@@ -2457,6 +2518,8 @@ ns = Collection(
     version,
     wait,
     worker,
+    worker_health,
+    server_health,
     monitor,
     build_docs,
 )