Github user steveloughran commented on the issue:
https://github.com/apache/spark/pull/22186
This will eliminate a race condition between FS shutdown (in the hadoop
shutdown manager) and the hive callback. Theres a risk today that the
filesystems will be closed before that event log close()/rename() is called, so
things don't get saved âand this can happen with any FS.
registering the shutdown hook via the spark APIs, with a priority > than
the FS shutdown, guarantees that it will be called before the FS shutdown. But
it doesn't guarantee that the operation will complete within the 10s time limit
hard coded into Hadoop 2.8.x+ for any single shutdown hook to complete. It is
going to work in HDFS except in the special case of HDFS NN lock or GC pause.
The Hadoop configurable delay of
[HADOOP-15679](https://issues.apache.org/jira/browse/HADOOP-15679) needs to go
in. I've increased the default timeout to 30s there for more forgiveness with
HDFS, and for object stores with O(data) renames people should configure it
with a timeout of minutes, or, if they want to turn it off altogether, hours.
I'm backporting HADOOP-15679 to all branches 2.8.x+, so all hadoop versions
with that timeout will have the timeout configurable & the default time
extended.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
