Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[SPARK-18364][YARN] Expose metrics for YarnShuffleService #17401

Closed
wants to merge 6 commits into from

Conversation

ash211
Copy link
Contributor

@ash211 ash211 commented Mar 23, 2017

Registers the shuffle server's metrics with the Hadoop Node Manager's
DefaultMetricsSystem.

What changes were proposed in this pull request?

Expose the shuffle service metrics not only on the ExternalShuffleService as done in SPARK-16405, but also in the YarnShuffleService. Because the YARN Node Manager's metrics system is the Hadoop-internal one and not the Dropwizard Metrics system that Spark uses, this requires a conversion from DW metrics to Hadoop metrics.

How was this patch tested?

Manual deploy of new yarn-shuffle jar into a Node Manager and verifying that the metrics appear in the Node Manager-standard location. This is JMX with an query endpoint at /jmx.

Resulting metrics look like this:

[user@host ~]$ curl -sk -XGET https://`hostname -f`:8042/jmx | jq . | grep 'shuffleservice' -B 1 -A 18
    {
      "name": "Hadoop:service=NodeManager,name=shuffleservice",
      "modelerType": "shuffleservice",
      "tag.Hostname": "<redacted>",
      "openBlockRequestLatencyMillis_count": 1,
      "openBlockRequestLatencyMillis_rate15": 0.0011080303990206543,
      "openBlockRequestLatencyMillis_rate5": 0.0033057092356765017,
      "openBlockRequestLatencyMillis_rate1": 0.015991117074135343,
      "openBlockRequestLatencyMillis_rateMean": 0.003843993699021382,
      "blockTransferRateBytes_count": 118,
      "blockTransferRateBytes_rate15": 0.1307475870844372,
      "blockTransferRateBytes_rate5": 0.39007368980982715,
      "blockTransferRateBytes_rate1": 1.8869518147479705,
      "blockTransferRateBytes_rateMean": 0.45359183094454836,
      "registeredExecutorsSize": 2,
      "registerExecutorRequestLatencyMillis_count": 2,
      "registerExecutorRequestLatencyMillis_rate15": 0.001697343764758814,
      "registerExecutorRequestLatencyMillis_rate5": 0.002970701813078509,
      "registerExecutorRequestLatencyMillis_rate1": 0.0005857750515146702,
      "registerExecutorRequestLatencyMillis_rateMean": 0.007687995987242345
    },
[user@host ~]$

Registers the shuffle server's metrics with the Hadoop Node Manager's
DefaultMetricsSystem.
@SparkQA
Copy link

SparkQA commented Mar 23, 2017

Test build #75115 has finished for PR 17401 at commit 4caf4d2.

  • This patch passes all tests.
  • This patch merges cleanly.
  • This patch adds the following public classes (experimental):
  • public class YarnShuffleServiceMetrics implements MetricsSource

@jerryshao
Copy link
Contributor

Actually we follow two space indent for java code in Spark, would you please change the format?

@ash211
Copy link
Contributor Author

ash211 commented Mar 24, 2017

Thanks for taking a look @jerryshao ! I've reformatted to two-space indentation and run ./dev/lint-java to make sure this code passes the linter.

src/main/java/org/apache/spark/sql/streaming/GroupStateTimeout.java still has some errors but those are separate from this change

* @param all if true, return all metrics even if unchanged.
*/
@Override
public void getMetrics(MetricsCollector collector, boolean all) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is it possible to add a unit test to verify the correctness of converting codahale metrics to Hadoop metrics?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

should be able to, I'm working on creating one now. By correctness, I think you mostly mean that the values passed through are the same, even though the naming schemes are different?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, you're right. From my understanding the correctness means Gauge still coverts to Gauge, Meter still to Meter, not sure can it be guaranteed?

Method registerSourceMethod = metricsSystem.getClass().getDeclaredMethod("registerSource",
String.class, String.class, MetricsSource.class);
registerSourceMethod.setAccessible(true);
registerSourceMethod.invoke(metricsSystem, "shuffleservice", "Metrics on the Spark " +
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit. "shuffleService" camel case might be better.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

will change shortly

@SparkQA
Copy link

SparkQA commented Mar 24, 2017

Test build #75150 has finished for PR 17401 at commit 13aa4ff.

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

@ash211
Copy link
Contributor Author

ash211 commented Mar 24, 2017

Thanks again for the comments @jerryshao ! I've now added some tests to verify that the metrics get converted in the expected way to the collector, and camel-cased shuffleService

@SparkQA
Copy link

SparkQA commented Mar 24, 2017

Test build #75157 has finished for PR 17401 at commit 9992c10.

  • This patch fails Scala style tests.
  • This patch merges cleanly.
  • This patch adds no public classes.

@SparkQA
Copy link

SparkQA commented Mar 24, 2017

Test build #75165 has finished for PR 17401 at commit 96a0882.

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

@ash211
Copy link
Contributor Author

ash211 commented Mar 24, 2017

Ready for further review.

}

@VisibleForTesting
public static void collectMetric(MetricsRecordBuilder metricsRecordBuilder, String name, Metric metric) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is it necessary to use static here? Looks like here it is it is only for the test convenience.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I use static here to make it clear that the method does not need to be run in the context of an instance. This prevents it from accidentally accessing instance variables when I don't intend it to

import org.mockito.Matchers._
import org.mockito.Mockito.{mock, times, verify, when}
import org.scalatest.Matchers
import scala.collection.JavaConverters._
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The import ordering is not correct, Scala package should be before the third party packages.

MetricsSystemImpl metricsSystem = (MetricsSystemImpl) DefaultMetricsSystem.instance();

Method registerSourceMethod = metricsSystem.getClass().getDeclaredMethod("registerSource",
String.class, String.class, MetricsSource.class);
Copy link
Contributor

@jerryshao jerryshao Mar 27, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nit, here I think should be 2 space indent follow the above line.

String.class, String.class, MetricsSource.class);
registerSourceMethod.setAccessible(true);
registerSourceMethod.invoke(metricsSystem, "shuffleService", "Metrics on the Spark " +
"Shuffle Service", serviceMetrics);
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also here.

@SparkQA
Copy link

SparkQA commented Mar 28, 2017

Test build #75322 has finished for PR 17401 at commit 7c7d6d4.

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

@ash211
Copy link
Contributor Author

ash211 commented Mar 28, 2017

@jerryshao ready for re-review

Copy link
Contributor

@jerryshao jerryshao left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry for the delay, just some style related comments. We should ping others to take a review.

@@ -176,7 +176,8 @@ private void checkAuth(TransportClient client, String appId) {
/**
* A simple class to wrap all shuffle service wrapper metrics
*/
private class ShuffleMetrics implements MetricSet {
@VisibleForTesting
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@VisibleForTesting causes classpath issues. Please note this in the java doc instead (SPARK-11615).

This is a scalastyle output, would be better to remove this annotation.

@@ -184,7 +204,7 @@ protected void serviceInit(Configuration conf) throws Exception {
boundPort = port;
String authEnabledString = authEnabled ? "enabled" : "not enabled";
logger.info("Started YARN shuffle service for Spark on port {}. " +
"Authentication is {}. Registered executor file is {}", port, authEnabledString,
"Authentication is {}. Registered executor file is {}", port, authEnabledString,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nit: this 2 space indent looks like not necessary.

}
}

@VisibleForTesting
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also here.

Gauge m = (Gauge) metric;
Object gaugeValue = m.getValue();
if (gaugeValue instanceof Integer) {
Integer intValue = (Integer) gaugeValue;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does it mean that we could only handle integer Gauge, what if later on we add different metric in ExternalShuffleBlockHandler? Looks like we should also manually handle the case one by one here. At least we should an else branch for the fallback check.

@HyukjinKwon
Copy link
Member

gentle ping @ash211. I just wonder if it is active now.

@asfgit asfgit closed this in 3a45c7f Aug 5, 2017
jackylee-ch pushed a commit to jackylee-ch/spark that referenced this pull request Feb 18, 2019
## What changes were proposed in this pull request?

This PR is follow-up of closed  apache#17401 which only ended due to of inactivity, but its still nice feature to have.
Given review by jerryshao taken in consideration and edited:
- VisibleForTesting deleted because of dependency conflicts
- removed unnecessary reflection for `MetricsSystemImpl`
- added more available types for gauge

## How was this patch tested?

Manual deploy of new yarn-shuffle jar into a Node Manager and verifying that the metrics appear in the Node Manager-standard location. This is JMX with an query endpoint running on `hostname:port`

Resulting metrics look like this:
```
curl -sk -XGET hostname:port |  grep -v '#' | grep 'shuffleService'
hadoop_nodemanager_openblockrequestlatencymillis_rate15{name="shuffleService",} 0.31428910657834713
hadoop_nodemanager_blocktransferratebytes_rate15{name="shuffleService",} 566144.9983653595
hadoop_nodemanager_blocktransferratebytes_ratemean{name="shuffleService",} 2464409.9678099006
hadoop_nodemanager_openblockrequestlatencymillis_rate1{name="shuffleService",} 1.2893844732240272
hadoop_nodemanager_registeredexecutorssize{name="shuffleService",} 2.0
hadoop_nodemanager_openblockrequestlatencymillis_ratemean{name="shuffleService",} 1.255574678369966
hadoop_nodemanager_openblockrequestlatencymillis_count{name="shuffleService",} 315.0
hadoop_nodemanager_openblockrequestlatencymillis_rate5{name="shuffleService",} 0.7661929192569739
hadoop_nodemanager_registerexecutorrequestlatencymillis_ratemean{name="shuffleService",} 0.0
hadoop_nodemanager_registerexecutorrequestlatencymillis_count{name="shuffleService",} 0.0
hadoop_nodemanager_registerexecutorrequestlatencymillis_rate1{name="shuffleService",} 0.0
hadoop_nodemanager_registerexecutorrequestlatencymillis_rate5{name="shuffleService",} 0.0
hadoop_nodemanager_blocktransferratebytes_count{name="shuffleService",} 6.18271213E8
hadoop_nodemanager_registerexecutorrequestlatencymillis_rate15{name="shuffleService",} 0.0
hadoop_nodemanager_blocktransferratebytes_rate5{name="shuffleService",} 1154114.4881816586
hadoop_nodemanager_blocktransferratebytes_rate1{name="shuffleService",} 574745.0749848988
```

Closes apache#22485 from mareksimunek/SPARK-18364.

Lead-authored-by: marek.simunek <[email protected]>
Co-authored-by: Andrew Ash <[email protected]>
Signed-off-by: Thomas Graves <[email protected]>
prakharjain09 pushed a commit to prakharjain09/spark that referenced this pull request Nov 29, 2019
## What changes were proposed in this pull request?

This PR is follow-up of closed  apache#17401 which only ended due to of inactivity, but its still nice feature to have.
Given review by jerryshao taken in consideration and edited:
- VisibleForTesting deleted because of dependency conflicts
- removed unnecessary reflection for `MetricsSystemImpl`
- added more available types for gauge

## How was this patch tested?

Manual deploy of new yarn-shuffle jar into a Node Manager and verifying that the metrics appear in the Node Manager-standard location. This is JMX with an query endpoint running on `hostname:port`

Resulting metrics look like this:
```
curl -sk -XGET hostname:port |  grep -v '#' | grep 'shuffleService'
hadoop_nodemanager_openblockrequestlatencymillis_rate15{name="shuffleService",} 0.31428910657834713
hadoop_nodemanager_blocktransferratebytes_rate15{name="shuffleService",} 566144.9983653595
hadoop_nodemanager_blocktransferratebytes_ratemean{name="shuffleService",} 2464409.9678099006
hadoop_nodemanager_openblockrequestlatencymillis_rate1{name="shuffleService",} 1.2893844732240272
hadoop_nodemanager_registeredexecutorssize{name="shuffleService",} 2.0
hadoop_nodemanager_openblockrequestlatencymillis_ratemean{name="shuffleService",} 1.255574678369966
hadoop_nodemanager_openblockrequestlatencymillis_count{name="shuffleService",} 315.0
hadoop_nodemanager_openblockrequestlatencymillis_rate5{name="shuffleService",} 0.7661929192569739
hadoop_nodemanager_registerexecutorrequestlatencymillis_ratemean{name="shuffleService",} 0.0
hadoop_nodemanager_registerexecutorrequestlatencymillis_count{name="shuffleService",} 0.0
hadoop_nodemanager_registerexecutorrequestlatencymillis_rate1{name="shuffleService",} 0.0
hadoop_nodemanager_registerexecutorrequestlatencymillis_rate5{name="shuffleService",} 0.0
hadoop_nodemanager_blocktransferratebytes_count{name="shuffleService",} 6.18271213E8
hadoop_nodemanager_registerexecutorrequestlatencymillis_rate15{name="shuffleService",} 0.0
hadoop_nodemanager_blocktransferratebytes_rate5{name="shuffleService",} 1154114.4881816586
hadoop_nodemanager_blocktransferratebytes_rate1{name="shuffleService",} 574745.0749848988
```

Closes apache#22485 from mareksimunek/SPARK-18364.

Lead-authored-by: marek.simunek <[email protected]>
Co-authored-by: Andrew Ash <[email protected]>
Signed-off-by: Thomas Graves <[email protected]>
(cherry picked from commit a802c69)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants