CloudWatch Metrics Input Plugin to Support Percentile Based Metrics

[Josh-Bonello]

Feature Request

Opening a feature request kicks off a discussion.

Proposal:

The CloudWatch GetMetricData API call has the capability to use an Expression to allow for gathering pXX.XX (IE p95) statistics metrics.

Current behavior:

Based on this function, only 'minimum, maximum, sum, samplecount, average' are supported metric values. There is no extendedmetrics available which would be used by extended metrics.

Desired behavior:

Allow for pXX.XX to be valid metrics to be gathered by the CloudWatch input plugin.

Use case:

For statistical analysis, it is often better to throw out outliers of a dataset to keep from distracting from underlying trends. Using p95 allows for a higher confidence interval in what real "max" usage is, p5 similarly for "minimum" usage, and p50 for "average" usage. This allows for a statistics_input object to contain arbitrary pXX.XX values and return them from the GetMetricData API. This is suggested as an additional option, rather than complete replacement, to allow for users of this plugin to choose which statistic set they prefer.

[ssoroka]

Seems like the histogram format would be ideal for this.

[Josh-Bonello]

@ssoroka Could you provide some more details about histogram format for a potential implementer, like myself?

[ssoroka]

Sure. Take a look at the aggregators.histogram plugin, as it outputs histograms and has examples of the outputted data format for pre-bucketed data.

I think you can also make a histogram out of pretty much any value and label combo.

[reimda]

@Josh-Bonello are you still interested in this feature? #8894 asks for general expression support which I believe includes the pXX.XX expression you mention. When this is implemented we should probably aim to support more than just percentiles.

I realize it has been a long time since you made the request, but are you able to make the changes and submit a PR?

[Josh-Bonello]

I believe this would be a real benefit to your users who consume cloudwatch. The histogram functionality that is mentioned previously by Steve requires the user to understand the boundaries of their bucket. CloudWatch data is vast and has many different values depending on their use case. I believe users to be better supported by using what is available in the source API.

They won't need to guess what values are around X percentile, the API will just provide what the Xth percentile is.

I can take a look at making changes, but it would be in "spare time" level of effort.

[reimda]

Thanks for responding even though this hasn't had any attention for a couple years.

I'm not aware of anyone else working on this. If you are able to take a look, even with "spare time" level of effort, it would be a help to the project. Thanks!