UPDATE: Spark 2.x support in Dr. Elephant

[akshayrai]

Hi everyone,

Sharing some updates from Linkedin.

We are successfully analyzing all Spark 2.x jobs using Dr. Elephant.

In the last couple of months, we have been trying hard to enable Spark support in Dr. Elephant but we faced several blockers due to the stability issues in the Spark History Server (SHS). We were finally able make it work at Linkedin after deploying a custom version of SHS which mostly builds on top of the pre-existing effort that is ongoing in the open source Spark branch to improve SHS. On top of it, we added a few patches to Spark that adds more metrics which we will be contributing to open source soon.

So, essentially, if you want Dr. Elephant to analyze Spark 2.x jobs, then you need this custom setup of Spark History Server, at least until all the work is part of an official release. Most of the open source SHS improvement work was done in SPARK-18085 (closed recently). From the Dr. Elephant perspective, all the related work is already part of the open source repository (master branch).

@edwinalu /@skakker will be sharing more details on the Spark SHS setup soon.

Regards, Akshay

[shkhrgpt]

Thanks, @akshayrai for sharing this information. I am a little confused with some of the stuff, therefore I would like to clarify a few things:

1. According to SPARK-18085 JIRA ticket, it seems like that improved SHS will be part of Spark 2.3. Does this mean Dr. Elephant will not support Spark 2.0, 2.1, and 2.2?

2. Currently, Dr. Elephant does not compile against Spark 2.X. Are we going to fix this or we will use the old methodology of compiling against Spark 1.4 and run for Spark 2.3?

3. If I am correct then Spark-18085 does not add all the new metrics that some the recent PRs are using. Does that mean that Spark 2.3 would not have those metrics?

[rajeshcode]

This is still through SparkRestclient we are getting data from spark history server and analysis right. Still the focus was not to fix the compile time errors with spark2.x right.

[edwinalu]

The new executor memory metrics have been proposed in SPARK-23206.

[akshayrai]

Hi @shkhrgpt, please find answers in-line below.

1. According to SPARK-18085 JIRA ticket, it seems like that improved SHS will be part of Spark 2.3. Does this mean Dr. Elephant will not support Spark 2.0, 2.1, and 2.2?

Dr. Elephant only needs a stable version of Spark History Server. So if we can setup the improved SHS against the older versions of Spark, then Dr. Elephant could work. This hasn't been tested though. At Linkedin, we have setup the improved SHS to run with Spark 2.1. @edwinalu can correct me if I am wrong.

2. Currently, Dr. Elephant does not compile against Spark 2.X. Are we going to fix this or we will use the old methodology of compiling against Spark 1.4 and run for Spark 2.3?

This needs to be fixed. For now, we are compiling against Spark 1.x and run on Spark 2.x

3. If I am correct then Spark-18085 does not add all the new metrics that some the recent PRs are using. Does that mean that Spark 2.3 would not have those metrics?

The new metrics do not block the Dr. Elephant/Spark setup. They are only required in some of the Heuristics. But, yes, if the metrics are not part of Spark 2.3 then some heuristics would not make sense. However, I am not sure how much of an effort that will be to back-port it to the older Spark versions.

[akshayrai]

@rajeshcode, yes, we are using the REST based SparkFetcher. And for the answer to your second question refer my answer to Shekhar above (point 2).

[edwinalu]

@akshayrai is correct. At LinkedIn, we are running Spark 2.3 History Server, which processes Spark applications running Spark 2.1 and 1.6. We have not tried with 1.4.

[andrijaperovic]

Is it possible to document your steps for custom SHS configuration running on Spark 2.x in the setup article? Or is it mostly documented in the SPARK-18085 JIRA you mentioned?

It doesn't look like Cloudera is distributing Spark 2.3 (which supports https://issues.apache.org/jira/browse/SPARK-18085) as of yet, have you done the upgrade by hand? Additionally, if we want to try this out should we be building against customSHSwork branch instead of master?

[rajeshcode]

@akshayrai
Is just Upgrading our SHS to spark2.3 will fix the issue to analyze spark2.x jobs ? Do we need to add any patch or update for DR elephant related to this ? In master branch didn't find any thing recently got updated based on this.

[akshayrai]

@skakker , as discussed earlier, can you please update the Dr. Elephant release notes page and document the setup instructions for spark 2.x support and update this thread? Ensure you address all the above questions in the document.

[hereTac]

With the spark2.3 release , any work here about support spark2.x whether has begining or any plan ? please let us know . The support about spark2.x will be very helpful . (ps. so much hope about analysis spark2.x .)

[skakker]

To clarify and answer most of the questions:

There are two types changes we have made internally to SHS: Stability changes (SPARK-23608 (fixed the Jetty issue), SPARK-23607 (performance improvements)) Adding new metrics (SPARK-23206) -> For adding new spark heuristics in Dr. Elephant

If you wish to use the code from the master branch of Dr. Elephant, you can do so with your SHS as well, only issue being that it may run into Jetty issues if load on the history server increases. We have sent out a PR after fixing the jetty issue, which can be viewed in the SPARK-23608 ticket. It hasn't been merged yet but you can apply that patch on top of your SHS 2.x.

The new heuristics changes are in the "customSHS" branch in Dr. Elephant. If you wish to use these heuristics, you will have to wait for the "SPARK-23206" changes to be merged with the SHS.

[mateo41]

Has anyone tried using Dr Elephant on a version of AWS EMR with Spark 2.3.0?

[achyuthsamudrala]

What does compiling against an older version and running against a new version mean? Thats not going to analyze the logs correctly anyway. I keep running into the issues that some event logs are not being parsed. (which are part of spark 2.x) Unless I am doing something incorrectly.

[jiajie999]

@mateo41 Not working on EMR with Spark 2.3.0, Hive tez works fine.

[ritika11]

Hi

Do we have some additional cpu/executor metrics for spark jobs?

[dmateusp]

I would be interested in getting that to work on EMR 2.3.0, what issues are you guys facing ? @jiajie999

Out of the box I ran into

08-23-2018 16:07:34 ERROR [ForkJoinPool-1-worker-7] com.linkedin.drelephant.spark.fetchers.SparkRestClient : error reading failedTasks http://XXXX.ec2.internal:XXXX/api/v1/applications/application_XXXX/stages/failedTasks
javax.ws.rs.NotFoundException: HTTP 404 Not Found

[ankurchourasiya]

I would be interested in getting that to work on EMR 2.3.0, what issues are you guys facing ? @jiajie999

Out of the box I ran into

08-23-2018 16:07:34 ERROR [ForkJoinPool-1-worker-7] com.linkedin.drelephant.spark.fetchers.SparkRestClient : error reading failedTasks http://XXXX.ec2.internal:XXXX/api/v1/applications/application_XXXX/stages/failedTasks
javax.ws.rs.NotFoundException: HTTP 404 Not Found

This is Fetcher issue. Please update your fetcher in /app-conf/FetcherConf.xml to update spark fetcher as below: <fetcher> <applicationtype>spark</applicationtype> <classname>com.linkedin.drelephant.spark.fetchers.SparkFetcher</classname> <params> <use_rest_for_eventlogs>true</use_rest_for_eventlogs> <should_process_logs_locally>true</should_process_logs_locally> </params> </fetcher>

[orbiran88]

@akshayrai There are any updates about compiling against Spark 2.x? thanks!!

[t2hw]

a lengthy discussion，so the conclusion so far is " nothing is documented and nobody wants（anyone who cans） "

so , any updates ? thanks 1000 times

[mareksimunek]

Also can't see any Spark metrics in Dr. Elephant UI. SHS version 2.3.0

I got com.linkedin.drelephant.spark.SparkMetricsAggregator : applicationDurationMillis is negative. Skipping Metrics Aggregation:-1548370957424

[helenfeng737]

Any updates on this?

[ligao101]

Any updates on this? With Spark 2.4 and more upcoming Spark 3.0 what is the current progress to support newer spark versions?

[ashwin-flip]

Hi @shkhrgpt, please find answers in-line below.

1. According to SPARK-18085 JIRA ticket, it seems like that improved SHS will be part of Spark 2.3. Does this mean Dr. Elephant will not support Spark 2.0, 2.1, and 2.2?

Dr. Elephant only needs a stable version of Spark History Server. So if we can setup the improved SHS against the older versions of Spark, then Dr. Elephant could work. This hasn't been tested though. At Linkedin, we have setup the improved SHS to run with Spark 2.1. @edwinalu can correct me if I am wrong.

1. Currently, Dr. Elephant does not compile against Spark 2.X. Are we going to fix this or we will use the old methodology of compiling against Spark 1.4 and run for Spark 2.3?

This needs to be fixed. For now, we are compiling against Spark 1.x and run on Spark 2.x

Hi, I am trying to build for Spark 2.4 but since it's not compiling, I tried with the default Spark 1.4 and in build.sbt I mentioned Scala version as 2.11. But the compilation is still failing. Any work around for this ?