[
https://issues.apache.org/jira/browse/HADOOP-14989?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16223883#comment-16223883
]
Eric Yang commented on HADOOP-14989:
------------------------------------
HI [~xkrogen],
Can we keep both sinks using the same refresh rate, like 10 seconds? I would
not recommend to have different refresh rate, this is comparing data samples at
different frequency. The resulting graph will not look the same. Total is a
high watermark and it will eventually overflow. This is the reason that Hadoop
community favored gauge system to minimize compute and interested to monitor
metrics at real time only during the development phase.
If we want to produce high fidelity data samples. Time stamp, previous count,
current count, and Time passed since last sample (or refresh rate) are the
essential information to record for high fidelity data samples, but post
processing is more expensive. Gauge and average are only good for measuring
velocity of the metrics for a point in time. Most monitoring system can only
handle time precision at second or minute scale. Hence, MutableRate is heavily
dependent on time precision that the down stream can consume. One important
limitation is JMX cache reset requires JMX sink to be the last one in the chain
with slowest refresh rate to avoid accurate problem like you described. JMX
sink should not have a lower refresh rate than FileSink to avoid destroying
samples before data is sent.
> Multiple metrics2 sinks (incl JMX) result in inconsistent Mutable(Stat|Rate)
> values
> -----------------------------------------------------------------------------------
>
> Key: HADOOP-14989
> URL: https://issues.apache.org/jira/browse/HADOOP-14989
> Project: Hadoop Common
> Issue Type: Bug
> Components: metrics
> Affects Versions: 2.6.5
> Reporter: Erik Krogen
> Priority: Critical
>
> While doing some digging in the metrics2 system recently, we noticed that the
> way {{MutableStat}} values are collected (and thus {{MutableRate}}, since it
> is based off of {{MutableStat}}) mean that each sink configured (including
> JMX) only receives a portion of the average information.
> {{MutableStat}}, to compute its average value, maintains a total value since
> last snapshot, as well as operation count since last snapshot. Upon
> snapshotting, the average is calculated as (total / opCount) and placed into
> a gauge metric, and total / operation count are cleared. So the average value
> represents the average since the last snapshot. If only a single sink ever
> snapshots, this would result in the expected behavior that the value is the
> average over the reporting period. However, if multiple sinks are configured,
> or if the JMX cache is refreshed, this is another snapshot operation. So, for
> example, if you have a FileSink configured at a 60 second interval and your
> JMX cache refreshes itself 1 second before the FileSink period fires, the
> values emitted to your FileSink only represent averages _over the last one
> second_.
> A few ways to solve this issue:
> * From an operator perspective, ensure only one sink is configured. This is
> not realistic given that the JMX cache exhibits the same behavior.
> * Make {{MutableRate}} manage its own average refresh, similar to
> {{MutableQuantiles}}, which has a refresh thread and saves a snapshot of the
> last quantile values that it will serve up until the next refresh. Given how
> many {{MutableRate}} metrics there are, a thread per metric is not really
> feasible, but could be done on e.g. a per-source basis. This has some
> downsides: if multiple sinks are configured with different periods, what is
> the right refresh period for the {{MutableRate}}?
> * Make {{MutableRate}} emit two counters, one for total and one for operation
> count, rather than an average gauge and an operation count counter. The
> average could then be calculated downstream from this information. This is
> cumbersome for operators and not backwards compatible. To improve on both of
> those downsides, we could have it keep the current behavior but
> _additionally_ emit the total as a counter. The snapshotted average is
> probably sufficient in the common case (we've been using it for years), and
> when more guaranteed accuracy is required, the average could be derived from
> the total and operation count.
> Open to suggestions & input here.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
