Merge branch 'master' of github.com:Yelp/paasta into u/vit/tron-1636-…

…add-secret-volume

debian/changelog

@@ -1,3 +1,10 @@
+paasta-tools (0.174.7) xenial; urgency=medium
+
+  * 0.174.7 tagged with 'make release'
+    Commit: Merge branch "u/krall/uwsgi_autoscaling_namespace"
+
+ -- Evan Krall <[email protected]>  Tue, 30 May 2023 14:48:11 -0700
+
 paasta-tools (0.174.6) xenial; urgency=medium
 
   * 0.174.6 tagged with 'make release'

paasta_tools/__init__.py

@@ -17,4 +17,4 @@
 # setup phase, the dependencies may not exist on disk yet.
 #
 # Don't bump version manually. See `make release` docs in ./Makefile
-__version__ = "0.174.6"
+__version__ = "0.174.7"

paasta_tools/setup_prometheus_adapter_config.py

@@ -232,10 +232,18 @@ def create_instance_uwsgi_scaling_rule(
     offset_multiplier = load_system_paasta_config().get_uwsgi_offset_multiplier()
 
     deployment_name = get_kubernetes_app_name(service=service, instance=instance)
+
+    # In order for autoscaling to work safely while a service migrates from one namespace to another, the HPA needs to
+    # make sure that the deployment in the new namespace is scaled up enough to handle _all_ the load.
+    # This is because once the new deployment is 100% healthy, cleanup_kubernetes_job will delete the deployment out of
+    # the old namespace all at once, suddenly putting all the load onto the deployment in the new namespace.
+    # To ensure this, we must:
+    #  - DO NOT filter on namespace in worker_filter_terms (which is used when calculating desired_instances).
+    #  - DO filter on namespace in replica_filter_terms (which is used to calculate current_replicas).
+    # This makes sure that desired_instances includes load from all namespaces, but that the scaling ratio calculated
+    # by (desired_instances / current_replicas) is meaningful for each namespace.
     worker_filter_terms = f"paasta_cluster='{paasta_cluster}',paasta_service='{service}',paasta_instance='{instance}'"
-    replica_filter_terms = (
-        f"paasta_cluster='{paasta_cluster}',deployment='{deployment_name}'"
-    )
+    replica_filter_terms = f"paasta_cluster='{paasta_cluster}',deployment='{deployment_name}',namespace='{namespace}'"
 
     current_replicas = f"""
         sum(
@@ -323,10 +331,18 @@ def create_instance_piscina_scaling_rule(
         DEFAULT_PISCINA_AUTOSCALING_MOVING_AVERAGE_WINDOW,
     )
     deployment_name = get_kubernetes_app_name(service=service, instance=instance)
+
+    # In order for autoscaling to work safely while a service migrates from one namespace to another, the HPA needs to
+    # make sure that the deployment in the new namespace is scaled up enough to handle _all_ the load.
+    # This is because once the new deployment is 100% healthy, cleanup_kubernetes_job will delete the deployment out of
+    # the old namespace all at once, suddenly putting all the load onto the deployment in the new namespace.
+    # To ensure this, we must:
+    #  - DO NOT filter on namespace in worker_filter_terms (which is used when calculating desired_instances).
+    #  - DO filter on namespace in replica_filter_terms (which is used to calculate current_replicas).
+    # This makes sure that desired_instances includes load from all namespaces, but that the scaling ratio calculated
+    # by (desired_instances / current_replicas) is meaningful for each namespace.
     worker_filter_terms = f"paasta_cluster='{paasta_cluster}',paasta_service='{service}',paasta_instance='{instance}'"
-    replica_filter_terms = (
-        f"paasta_cluster='{paasta_cluster}',deployment='{deployment_name}'"
-    )
+    replica_filter_terms = f"paasta_cluster='{paasta_cluster}',deployment='{deployment_name}',namespace='{namespace}'"
 
     current_replicas = f"""
         sum(

yelp_package/Makefile

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 # Edit this release and run "make release"
-RELEASE=0.174.6
+RELEASE=0.174.7
 
 SHELL=/bin/bash