[
https://issues.apache.org/jira/browse/STORM-2153?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15595862#comment-15595862
]
P. Taylor Goetz commented on STORM-2153:
----------------------------------------
One of the things we're going to have to work out is how to properly aggregate
different metric types, if we can at all. For simple metric types like
Counters, it is straightforward. But things get more complicated with metrics
like Meters and Histograms -- there's no way to really aggregate percentiles.
The problem is that metrics maintain a certain amount of state necessary to
calculate certain values like percentiles. When a reporter reports those
values, it takes a point-in-time snapshot of those calculations, so all the
data used to calculate those values is no longer available.
As an example lets say we want to have a Histogram for execute latency (so we
can report various percentiles), and we want to aggregate those statistics
across a topology. So worker A and worker B report their histogram snapshot to
nimbus. How does nimbus aggregate those values without the datasets the
calculations were based on?
> New Metrics Reporting API
> -------------------------
>
> Key: STORM-2153
> URL: https://issues.apache.org/jira/browse/STORM-2153
> Project: Apache Storm
> Issue Type: Improvement
> Reporter: P. Taylor Goetz
>
> This is a proposal to provide a new metrics reporting API based on [Coda
> Hale's metrics library | http://metrics.dropwizard.io/3.1.0/] (AKA
> Dropwizard/Yammer metrics).
> h2. Background
> In a [discussion on the dev@ mailing list |
> http://mail-archives.apache.org/mod_mbox/storm-dev/201610.mbox/%3ccagx0urh85nfh0pbph11pmc1oof6htycjcxsxgwp2nnofukq...@mail.gmail.com%3e]
> a number of community and PMC members recommended replacing Storm’s metrics
> system with a new API as opposed to enhancing the existing metrics system.
> Some of the objections to the existing metrics API include:
> # Metrics are reported as an untyped Java object, making it very difficult to
> reason about how to report it (e.g. is it a gauge, a counter, etc.?)
> # It is difficult to determine if metrics coming into the consumer are
> pre-aggregated or not.
> # Storm’s metrics collection occurs through a specialized bolt, which in
> addition to potentially affecting system performance, complicates certain
> types of aggregation when the parallelism of that bolt is greater than one.
> In the discussion on the developer mailing list, there is growing consensus
> for replacing Storm’s metrics API with a new API based on Coda Hale’s metrics
> library. This approach has the following benefits:
> # Coda Hale’s metrics library is very stable, performant, well thought out,
> and widely adopted among open source projects (e.g. Kafka).
> # The metrics library provides many existing metric types: Meters, Gauges,
> Counters, Histograms, and more.
> # The library has a pluggable “reporter” API for publishing metrics to
> various systems, with existing implementations for: JMX, console, CSV, SLF4J,
> Graphite, Ganglia.
> # Reporters are straightforward to implement, and can be reused by any
> project that uses the metrics library (i.e. would have broader application
> outside of Storm)
> As noted earlier, the metrics library supports pluggable reporters for
> sending metrics data to other systems, and implementing a reporter is fairly
> straightforward (an example reporter implementation can be found here). For
> example if someone develops a reporter based on Coda Hale’s metrics, it could
> not only be used for pushing Storm metrics, but also for any system that used
> the metrics library, such as Kafka.
> h2. Scope of Effort
> The effort to implement a new metrics API for Storm can be broken down into
> the following development areas:
> # Implement API for Storms internal worker metrics: latencies, queue sizes,
> capacity, etc.
> # Implement API for user defined, topology-specific metrics (exposed via the
> {{org.apache.storm.task.TopologyContext}} class)
> # Implement API for storm daemons: nimbus, supervisor, etc.
> h2. Relationship to Existing Metrics
> This would be a new API that would not affect the existing metrics API. Upon
> completion, the old metrics API would presumably be deprecated, but kept in
> place for backward compatibility.
> Internally the current metrics API uses Storm bolts for the reporting
> mechanism. The proposed metrics API would not depend on any of Storm's
> messaging capabilities and instead use the [metrics library's built-in
> reporter mechanism |
> http://metrics.dropwizard.io/3.1.0/manual/core/#man-core-reporters]. This
> would allow users to use existing {{Reporter}} implementations which are not
> Storm-specific, and would simplify the process of collecting metrics.
> Compared to Storm's {{IMetricCollector}} interface, implementing a reporter
> for the metrics library is much more straightforward (an example can be found
> [here |
> https://github.com/dropwizard/metrics/blob/3.2-development/metrics-core/src/main/java/com/codahale/metrics/ConsoleReporter.java].
> The new metrics capability would not use or affect the ZooKeeper-based
> metrics used by Storm UI.
> h2. Relationship to JStorm Metrics
> [TBD]
> h2. Target Branches
> [TBD]
> h2. Performance Implications
> [TBD]
> h2. Metrics Namespaces
> [TBD]
> h2. Metrics Collected
> *Worker*
> || Namespace || Metric Type || Description ||
> *Nimbus*
> || Namespace || Metric Type || Description ||
> *Supervisor*
> || Namespace || Metric Type || Description ||
> h2. User-Defined Metrics
> [TBD]
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
