Ryan Williams created SPARK-5847:
------------------------------------
Summary: Allow for configuring MetricsSystem's use of app ID to
namespace all metrics
Key: SPARK-5847
URL: https://issues.apache.org/jira/browse/SPARK-5847
Project: Spark
Issue Type: Improvement
Components: Spark Core
Affects Versions: 1.2.1
Reporter: Ryan Williams
Priority: Minor
{{MetricsSystem}} [currently prepends the app ID to all
metrics|https://github.com/apache/spark/blob/c51ab37faddf4ede23243058dfb388e74a192552/core/src/main/scala/org/apache/spark/metrics/MetricsSystem.scala#L131].
When reading Spark metrics in Graphite, I've found this to not always be
desirable. Graphite is designed to track a mostly-unchanging set of metrics
over time; it allocates large zeroed-out files for each metric it sees, and [by
default rate-limits itself from creating many of
these|https://github.com/graphite-project/carbon/blob/79158ffde5949b4056eb7fdb5e9b6b583fe21ea4/conf/carbon.conf.example#L61-L68].
App-ID namespacing means that Graphite is allocating disk-space for every
"metric" for every job it sees, when in reality some metrics may correspond to
others across jobs (e.g. driver JVM stats).
Some common Spark usage flows would be better modeled by e.g. namespacing
metrics by {{spark.app.name}}, so that successive runs of a given job would
share "metrics", from a storage perspective as well as allowing for monitoring
aspects of a job's performance over time / many runs.
There's not likely a one-size-fits-all solution here, so I'd propose allowing
the metrics config file to allow users to specify whether they'd like metrics
namespaced by {{spark.app.id}}, {{spark.app.name}}, or some other config param.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
