after upgrade from Spark 1.5 to 1.6(CDH 5.6.0 -> 5.7.1)
some of our streaming job getting delay after long run.
with a little invesgation, here is what i found.
- the same program have no problem with Spark 1.5
- we have two kind of streaming and only those with
"updateStateByKey" was affected,
- cpu usage getting higher and higher over time ( with 1core@5% at
start and 1core@100% after a week )
- data rate is alound 100 event/s, there is no chance for the cpu
to work so hard.
- process time for a batch delay from 100ms at start to 3s after a week
- evening running the same program(for difference input data), not
all process delay with the same scale
- no warning or error message until it delay too much and went out
of memory
- process time of customer code seems have no problem
- memory/heap usage looks normal to me
Im suspecting the problem is comming from updateStateByKey but i cant
trace it down
any one experience the same problem?
--
BR
Peter Chan
---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscr...@spark.apache.org