initial changes for invoker cluster backpressure #3
Conversation
There was a problem hiding this comment.
This would really address the dynamic capacity addition aspect! Do we have the change in MesosContainerFactory somewhere which trigger the NodeStatsUpdateEvent? That would help to see complete picture once
It may be better to extract out a ResourcePolicy abstraction to keep the policy aspect separate from the scheduling purpose
| //if cluster managed resources, subscribe to events | ||
| if (poolConfig.clusterManagedResources) { | ||
| logging.info(this, "subscribing to NodeStats updates") | ||
| Events.subscribe(self, NodeStatsUpdateEvent) |
There was a problem hiding this comment.
Should it unsubscribe itself? Given its kind of singleton its fine but may be for testcases we should unsubscribe
There was a problem hiding this comment.
Looks like EventStream will handle unsubscribe on actor termination per docs, but will change this if it is a problem for tests.
| * Publishes the payload of the MsgEnvelope when the topic of the | ||
| * MsgEnvelope equals the OWEvent specified when subscribing. | ||
| */ | ||
| object Events extends EventBus with LookupClassification { |
There was a problem hiding this comment.
We can possibly use the EventStream of the ActorSystem itself. Should meet our requirement
There was a problem hiding this comment.
Sounds good! will look into making this change
|
I'm interested to know what you have in mind for |
Codecov Report
@@ Coverage Diff @@
## master #3 +/- ##
===========================================
- Coverage 79.42% 42.79% -36.63%
===========================================
Files 170 173 +3
Lines 7940 8256 +316
Branches 532 582 +50
===========================================
- Hits 6306 3533 -2773
- Misses 1634 4723 +3089
Continue to review full report at Codecov.
|
|
Was thinking of trait which abstracts the However idea would be to move your current changes to such a policy and then use that in ContainerPool decision logic. One aspect which makes it tricky is that policy need to be refreshed via event. So we may need to use some volatile state which gets updated after post processing of cluster state changes and then consumer by the |
tysonnorris
force-pushed
the
invoker-backpressure
branch
from
9d7a1a0
to
43f9028
Compare
tysonnorris
force-pushed
the
invoker-backpressure
branch
from
2c92309
to
ef31370
Compare
tysonnorris
force-pushed
the
invoker-backpressure
branch
from
6a5e9cc
to
9889d68
Compare
…e#4475) * Additional cleanup and simplifications for the ActionLoop docs * Additional cleanup and simplifications for the ActionLoop docs
* Updating runtimes to include new Node.js v12 image. - Updated runtimes manifest - Added API documentation - Minor updates to docs - Added automated test case * Fixing accidental default flag for v12
* Update docker version in the controller and guides * Update travis configuration to install docker=18.06.3 * Separate docker installation from the travis setup script * Add docker setup script * Add CoordinatedShutdown to cleanup runtime containers * add a configuration for root runc dir * Disable runc use in Jenkins environment * Add comments which explain the correlation among the version docker and runc and the type of user * Reenable Docker remote API again
…ame (apache#4488) Add `pip install docker` to Dockerfile for ow-utils to fix problem pulling docker images
* Switch to consistent indexing policy * Remove reference to Range index * Tweak indexing policy comparison to only check for included and excluded path Do not check for Index type as now there is only one which is Range for our cases * Use implicit logger * Excluding root path should be using `/*` instead of `/`
Updates to Alpakka S3 Connector v1.0.1 release. This commit also includes some fixes related to apache#4484 which were causing compatibility issues with existing setup. For existing setup indexes were still using `Hash` indexing and that caused failure while creating `IndexingPolicy` upon collection read. So added back support for `Hash` index but not using them to create new indexes now
Record collection usage stats for CosmosDB so as to enable tracking the growth of collection in terms of storage size, document count and index size over the period of time. It also enables tracking any indexing progress if any change is done in Index configuration. Note that Count stats are currently not exposed via Azure Portal Further this commit also enables emitting verbose trace for query when in debug mode. This would simplify any query performance analysis. Fixes apache#4489
`lines` method is now defined as part of java.lang.String in JDK 11. So need to use `linesIterator` for right method to be picked
Adds a configurable MetricsReporter to route Kafka metrics to Kamon once enabled. Set of metrics names which need to be captured needs to be explicitly configured
* Update to restassured v4.0 which is compatible with jdk11 * Need to initialize both truststore and keystore for ssl cert to be validated
* Add explicit return type on a few implicit values. Implicit values without an explicit return types are not guaranteed to work. They may or may not compile, depending on compilation order and they are going to be disallowed in future versions of Scala. This confuses IntelliJ as well. * Make Exec public since it's being used from tests. The call to Exec.isBinaryCode in ActionContainer.scala is from a different package and normally not visible at the call site. Due to scala/bug#11554 this may compile sometimes.
* simplify throttle code * revert to original throttle algorithm * Make the waiting time to calm down the thottle explicit
…rceManager impls; handle ClusterResourceError in case of reaching cluster capacity
…s, total memory, max memory
…dle containers; only remove idle containers that match the actual host/port of selected container
…nup of containers
tysonnorris
force-pushed
the
invoker-backpressure
branch
from
dec81c5
to
a2a1844
Compare
Description
Related issue and scope
My changes affect the following components
Types of changes
Checklist: