Imran Rashid created SPARK-15976:
------------------------------------
Summary: Make Stage Numbering determinstic
Key: SPARK-15976
URL: https://issues.apache.org/jira/browse/SPARK-15976
Project: Spark
Issue Type: Improvement
Components: Scheduler
Affects Versions: 2.0.0
Reporter: Imran Rashid
Stage numbering in spark is non-deterministic. It never was deterministic, but
it *appeared* to be so in most cases. After SPARK-15927, it is far more
random. Reliable stage numbering would be helpful for internal unit tests, and
also for any client code which uses {{SparkListener}} to monitor a job and
gauge progress.
FWIW, I had never even realized that the order was non-deterministic before,
and have written plenty of code which assumes some stage numbering. I expect
users may be bitten by this too. We might even want to try to restore the
"usual" ordering from before SPARK-15927.
Finally it would be nice to restore some of the tests turned off here if
possible: https://github.com/apache/spark/pull/13688
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
