Merge branch '20680-default-config-updates' refs #20680

[arvados.git] / lib / config / config.default.yml

diff --git a/lib/config/config.default.yml b/lib/config/config.default.yml
index 49d62e2980dcee86825d0af8f35b5c1377505f9a..b78116255a14687208dace80d212dae452bbac80 100644 (file)
--- a/lib/config/config.default.yml
+++ b/lib/config/config.default.yml
@@ -225,7 +225,7 @@ Clusters:
 
       # Maximum number of concurrent requests to process concurrently
       # in a single service process, or 0 for no limit.
-      MaxConcurrentRequests: 64
+      MaxConcurrentRequests: 4
 
       # Maximum number of incoming requests to hold in a priority
       # queue waiting for one of the MaxConcurrentRequests slots to be
@@ -234,7 +234,7 @@ Clusters:
       #
       # If MaxQueuedRequests is 0, respond 503 immediately to
       # additional requests while at the MaxConcurrentRequests limit.
-      MaxQueuedRequests: 64
+      MaxQueuedRequests: 128
 
       # Maximum time a "lock container" request is allowed to wait in
       # the incoming request queue before returning 503.
@@ -1074,7 +1074,7 @@ Clusters:
 
       # Number of times a container can be unlocked before being
       # automatically cancelled.
-      MaxDispatchAttempts: 5
+      MaxDispatchAttempts: 10
 
       # Default value for container_count_max for container requests.  This is the
       # number of times Arvados will create a new container to satisfy a container
@@ -1413,10 +1413,30 @@ Clusters:
         # down.
         MaxInstances: 64
 
-        # Maximum fraction of CloudVMs.MaxInstances allowed to run
-        # "supervisor" containers at any given time. A supervisor is a
-        # container whose purpose is mainly to submit and manage other
-        # containers, such as arvados-cwl-runner workflow runner.
+        # The minimum number of instances expected to be runnable
+        # without reaching a provider-imposed quota.
+        #
+        # This is used as the initial value for the dispatcher's
+        # dynamic instance limit, which increases (up to MaxInstances)
+        # as containers start up successfully and decreases in
+        # response to high API load and cloud quota errors.
+        #
+        # Setting this too high creates a risk that the dispatcher
+        # will cause deadlock by starting so many supervisor
+        # containers (based on SupervisorFraction and MaxInstances)
+        # that the cloud quota prevents them from running any child
+        # containers.
+        #
+        # Setting this too low causes the dispatcher to be
+        # unnecessarily slow to start up new instances after a
+        # restart.
+        InitialQuotaEstimate: 16
+
+        # Maximum fraction of available instance capacity allowed to
+        # run "supervisor" containers at any given time. A supervisor
+        # is a container whose purpose is mainly to submit and manage
+        # other containers, such as arvados-cwl-runner workflow
+        # runner.
         #
         # If there is a hard limit on the amount of concurrent
         # containers that the cluster can run, it is important to