[SPARK-25697][CORE]When zstd compression enabled, InProgress application is throwing Error in the history webui

[shahidki31]

What changes were proposed in this pull request?

When we enable event log compression and compression codec as 'zstd', we are unable to open the webui of the running application from the history server page.
The reason is that, Replay listener was unable to read from the zstd compressed eventlog due to the zstd frame was not finished yet. This causes truncated error while reading the eventLog.

So, when we try to open the WebUI from the History server page, it throws "truncated error ", and we never able to open running application in the webui, when we enable zstd compression.

In this PR, when the IO excpetion happens, and if it is a running application, we log the error,
"Failed to read Spark event log: evetLogDirAppName.inprogress", instead of throwing exception.

How was this patch tested?

Test steps:
1)spark.eventLog.compress = true
2)spark.io.compression.codec = zstd
3)restart history server
4) launch bin/spark-shell
5) run some queries
6) Open history server page
7) click on the application

Before fix:

After fix:

(Please explain how this patch was tested. E.g. unit tests, integration tests, manual tests)
(If this patch involves UI changes, please attach a screenshot; otherwise, remove this)

Please review http://spark.apache.org/contributing.html before opening a pull request.

[shahidki31]

cc @vanzin @srowen . Kindly review.

[srowen]

Should the Event Log be available for running apps? Or if it's not going to work, disable it where it can't be shown, but I suppose that could be difficult. This just silently sends you back to the jobs page?

[shahidki31]

Hi @srowen . Yes. Event logs are available for running apps, but with the extension, ".inprogress".
We can open webui from the history server page for both running and finished applications.
This is working for all compressed codecs (lz4, snappy, lzf) , also for uncompressed event logs.

[srowen]

I guess that doing nothing is better than an error screen. Is it possible to just skip reading incomplete files here? I don't know this code well. That sounds better.

[shahidki31]

@srowen . Yes. We should read only from the finished frames of zstd. When the listener try to read from the unfinished frame, zstd input reader throws an exception (unless we make set continuous true).

Currently the behavior is, it reads from the finished frames, but after that it try to read from the unfinished frame and throws exception while loading the webui. So, the solution should be, we should not parse from the unfinished frame, and load the UI based on only the finished frames.

Hi @vanzin , could you please give your inputs?

Thanks

[felixcheung]

ok to test

[SparkQA]

Test build #97233 has finished for PR 22689 at commit c309f34.

• This patch fails due to an unknown error code, -9.
• This patch merges cleanly.
• This patch adds no public classes.

[shahidki31]

retest this please.

[SparkQA]

Test build #97247 has finished for PR 22689 at commit 0924a0a.

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

[srowen]

OK, so it does read all the data it can. That seems fine

[shahidki31]

Yes. Thank you @srowen .

[srowen]

Merged to master/2.4

[shahidki31]

Thanks a lot @srowen