wall times are all messed up. Everything is negative hours
vijay-jangir opened this issue · comments
This is a report I generated with the JSON output. Using the latest jar to generate the reports
--packages qubole:sparklens:0.3.2-s_2.11 \
@vijay-jangir Thanks for reporting this issue.
We analysed the sparklens Json and noticed that 364 out of total 535 stages failed to report information about all the tasks. Out of expected 128152 tasks 45867 were not reported. At our end we plan to add another sparklens analyser which can report such problems. It might be good idea to grep the logs for "event dropped" to validate this claim. Increasing the size of spark.scheduler.listenerbus.eventqueue.size from default 10K to something higher like 20K or 30K will probably help. Since this will take up some memory at driver end, it will be good idea to increase the driver memory a bit if under pressure.
See https://issues.apache.org/jira/browse/SPARK-18881 and https://issues.apache.org/jira/browse/SPARK-18838 for more context around this issue.