Stig Rohde Døssing created STORM-3121:
-----------------------------------------
Summary: Fix flaky metrics tests in storm-core
Key: STORM-3121
URL: https://issues.apache.org/jira/browse/STORM-3121
Project: Apache Storm
Issue Type: Bug
Components: storm-core
Affects Versions: 2.0.0
Reporter: Stig Rohde Døssing
Assignee: Stig Rohde Døssing
The tests are flaky, but only rarely fail. I've only seen them fail on Travis
when Travis is under load.
Example failures:
{code}
classname: org.apache.storm.metrics-test / testname:
test-custom-metric-with-multi-tasks
expected: (clojure.core/= [1 0 0 0 0 0 2] (clojure.core/subvec
(org.apache.storm.metrics-test/lookup-bucket-by-comp-id-&-metric-name! "2"
"my-custom-metric") 0 N__3207__auto__))
actual: (not (clojure.core/= [1 0 0 0 0 0 2] [1 0 0 0 0 0 0]))
at: test_runner.clj:105
{code}
{code}
classname: org.apache.storm.metrics-test / testname: test-builtin-metrics-2
expected: (clojure.core/= [1 1] (clojure.core/subvec
(org.apache.storm.metrics-test/lookup-bucket-by-comp-id-&-metric-name!
"myspout" "__emit-count/default") 0 N__3207__auto__))
actual: (not (clojure.core/= [1 1] [1 0]))
at: test_runner.clj:105
{code}
The problem is that the tests increment metrics counters in the executor async
loops, then expect the counters to end up in exact metrics buckets. The
creation of a bucket is triggered by the metrics timer. The timer is included
in time simulation and LocalCluster.waitForIdle, but the executor async loop
isn't. There isn't any guarantee that the executor async loop gets to run when
the test does a sequence like
{code}
Time.advanceClusterTime
cluster.waitForIdle
{code}
because the waitForIdle check doesn't know about the executor async loop.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)