Merge "Bump workers_pool_size to 300 and remove queueing of tasks"

ironic/conductor/base_manager.py

@@ -125,9 +125,12 @@ class BaseConductorManager(object):
         self._keepalive_evt = threading.Event()
         """Event for the keepalive thread."""
 
-        # TODO(dtantsur): make the threshold configurable?
-        rejection_func = rejection.reject_when_reached(
-            CONF.conductor.workers_pool_size)
+        # NOTE(dtantsur): do not allow queuing work. Given our model, it's
+        # better to reject an incoming request with HTTP 503 or reschedule
+        # a periodic task that end up with hidden backlog that is hard
+        # to track and debug. Using 1 instead of 0 because of how things are
+        # ordered in futurist (it checks for rejection first).
+        rejection_func = rejection.reject_when_reached(1)
         self._executor = futurist.GreenThreadPoolExecutor(
             max_workers=CONF.conductor.workers_pool_size,
             check_and_reject=rejection_func)

ironic/conf/conductor.py

@@ -22,7 +22,7 @@ from ironic.common.i18n import _
 
 opts = [
     cfg.IntOpt('workers_pool_size',
-               default=100, min=3,
+               default=300, min=3,
                help=_('The size of the workers greenthread pool. '
                       'Note that 2 threads will be reserved by the conductor '
                       'itself for handling heart beats and periodic tasks. '

releasenotes/notes/workers-20ca5c225c1474e0.yaml

@@ -0,0 +1,25 @@
+---
+issues:
+  - |
+    When configuring a single-conductor environment, make sure the number
+    of worker pools (``[conductor]worker_pool_size``) is larger than the
+    maximum parallel deployments (``[conductor]max_concurrent_deploy``).
+    This was not the case by default previously (the options used to be set
+    to 100 and 250 accordingly).
+upgrade:
+  - |
+    Because of a fix in the internal worker pool handling, you may now start
+    seeing requests rejected with HTTP 503 under a very high load earlier than
+    before. In this case, try increasing the ``[conductor]worker_pool_size``
+    option or consider adding more conductors.
+  - |
+    The default worker pool size (the ``[conductor]worker_pool_size`` option)
+    has been increased from 100 to 300. You may want to consider increasing
+    it even further if your environment allows that.
+fixes:
+  - |
+    Fixes handling new requests when the maximum number of internal workers
+    is reached. Previously, after reaching the maximum number of workers
+    (100 by default), we would queue the same number of requests (100 again).
+    This was not intentional, and now Ironic no longer queues requests if
+    there are no free threads to run them.