[SPARK-5227] [SPARK-5679] Disable FileSystem cache in WholeTextFileRecordReaderSuite

[JoshRosen]

This patch fixes two difficult-to-reproduce Jenkins test failures in InputOutputMetricsSuite (SPARK-5227 and SPARK-5679). The problem was that WholeTextFileRecordReaderSuite modifies the fs.local.block.size Hadoop configuration and this change was affecting subsequent test suites due to Hadoop's caching of FileSystem instances (see HADOOP-8490 for more details).

The fix implemented here is to disable FileSystem caching in WholeTextFileRecordReaderSuite.

[SparkQA]

Test build #27461 has started for PR 4599 at commit 47dc447.

• This patch merges cleanly.

[JoshRosen]

I SSH'ed into an AMPLab Jenkins box to reproduce the original failures and have confirmed that this patch fixes them.

/cc @ksakellis

[JoshRosen]

JIRA links, for the curious:

• https://issues.apache.org/jira/browse/HADOOP-8490
• https://issues.apache.org/jira/browse/SPARK-5679
• https://issues.apache.org/jira/browse/SPARK-5227

[ksakellis]

So, do you think we should disable it across all tests? just in case there are other tests that also modify the hadoop configuration thinking that the config objects are local to them? It might bite someone else in the butt later if we don't globally do this. I don't think there is a global test class that every tests inherits, maybe we can add it in SparkSparkContext since a lot of the new tests written use that trait?

[JoshRosen]

Good question. If we wanted to disable this in all tests, then I think the right place to do that would be in the Maven and SBT builds via system properties.

I chose not to do that here because I wasn't sure whether doing so might mask bugs, since most users of Spark will run with FileSystem caching enabled (I think that disabling it across the board may harm performance, since it sounds like a lot of Hadoop code assumes that FileSystem.get is cheap, and, accordingly, calls it many times).

[sryza]

My opinion is that doing it for all tests would be a bit drastic. It's not that modifications to Configuration objects arbitrarily affect other Configuration objects. It's that this test specifically relies on some underlying properties set on the FileSystem, and the FS Cache allows these to leak to other FS instances.

[SparkQA]

Test build #27461 has finished for PR 4599 at commit 47dc447.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/27461/
Test PASSed.

[JoshRosen]

/cc @pwendell @andrewor14

[ksakellis]

Ok, this LTGM

[pwendell]

LGTM .