Spark 2 - attemptFailuresValidityInterval issue

[#1]

Hi!

We are running a spark-submit with options:

--deploy-mode cluster

--conf "spark.yarn.maxAppAttempts=3"
--conf "spark.yarn.am.attemptFailuresValidityInterval=30s"

--conf...

and our application throws intentionally an exception after 70 seconds on the driver, in order to cause a manual failure.

We expected our application, with these parameters, to run forever, because the attemptFailuresValidityInterval should reset the maxAppAttempts counter sooner than the custom exception. But after 3 failures the application stops.

Our installation:

- SPARK2-2.1.0.cloudera2
- CDH 5.11

Any ideas are more than welcome!

[#2 · Accepted solution]

Sorry, this is a bug described in SPARK-22876 which suggests that the current logic of spark.yarn.am.attemptFailuresValidityInterval is flawed.

While the jira is still being worked on, looking at the comments, I don't foresee a fix anytime soon.