Skip to content

Add exponential histogram support to CloudWatch PMD Exporter #1677

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

Open
wants to merge 12 commits into
base: main
Choose a base branch
from

Conversation

dricross
Copy link
Contributor

@dricross dricross commented May 8, 2025

Description of the issue

The Cloudwatch/PMD exporter currently drops all exponential histogram metrics.

Description of changes

  • Add support for exponential histogram to the ec2tagger processor
  • Add support for exponential histogram to the CloudWatch/PMD exporter

Note

See companion PR for updating cumulativetodelta processor: amazon-contributing/opentelemetry-collector-contrib#331

License

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Tests

Note

See companion PR for integration test: aws/amazon-cloudwatch-agent-test#558

Integration test run: https://github.com/aws/amazon-cloudwatch-agent/actions/runs/16371811763
Histogram test: https://github.com/aws/amazon-cloudwatch-agent/actions/runs/16371811763/job/46261959442

Requirements

Before commit the code, please do the following steps.

  1. Run make fmt and make fmt-sh
  2. Run make lint

@dricross dricross requested a review from a team as a code owner May 8, 2025 20:55
@dricross dricross force-pushed the dricross/exponentialhistogram branch 2 times, most recently from 97c4271 to 995902e Compare May 16, 2025 21:08
return d.AddEntryWithUnit(value, weight, "")
}

func (d *ExpHistogramDistribution) AddDistribution(other *ExpHistogramDistribution) {
Copy link
Contributor

Choose a reason for hiding this comment

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

Are we trying to match the Distribution interface? It doesn't look like this would satisfy the interface as it is now. We would have to make the Distribution interface generic.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Yeah, you're right it doesn't satisfy the interface. Initially I was trying to do so but it ended up not working out because the Distribution interface is specific too HistogramDataPoint. In hindsight, I don't think it's even possible to combine exp histograms and classic histograms anyways so not really necessary at all. I already had this code though and didn't think it was worth refactoring everything to make it fit nicely together so I left it as is for now

}
// Assume function pointer is valid.
ad.expHistDistribution = exph.NewExpHistogramDistribution()
ad.expHistDistribution.ConvertFromOtel(dp, unit)
Copy link
Contributor

Choose a reason for hiding this comment

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

nit: If we store the unit in the MetricDatum already, why do we need to store it in the distribution?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

This is based off of how it currently works for the regular histograms. The unit is given to the distribution so that it can check that similar distributions are being combined. If the unit for distribution does match the unit for the incoming distribution to combine it with, a debug log is printed.

return datums
}

func (c *CloudWatch) buildMetricDataumExph(metric *aggregationDatum, dimensionsList [][]*cloudwatch.Dimension) []*cloudwatch.MetricDatum {
Copy link
Contributor

Choose a reason for hiding this comment

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

nit: Typo

Suggested change
func (c *CloudWatch) buildMetricDataumExph(metric *aggregationDatum, dimensionsList [][]*cloudwatch.Dimension) []*cloudwatch.MetricDatum {
func (c *CloudWatch) buildMetricDatumExph(metric *aggregationDatum, dimensionsList [][]*cloudwatch.Dimension) []*cloudwatch.MetricDatum {

Copy link
Contributor

This PR was marked stale due to lack of activity.

@github-actions github-actions bot added the Stale label May 28, 2025
@github-actions github-actions bot removed the Stale label Jun 28, 2025
Copy link
Contributor

github-actions bot commented Jul 7, 2025

This PR was marked stale due to lack of activity.

@@ -4,6 +4,10 @@ go 1.24.4

replace github.com/influxdata/telegraf => github.com/aws/telegraf v0.10.2-0.20250113150713-a2dfaa4cdf6d

replace collectd.org v0.4.0 => github.com/collectd/go-collectd v0.4.0
Copy link
Contributor

Choose a reason for hiding this comment

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

are these dependencies of cumulativetodeltaprocessor? I don't see go-collectd or clock used in the new code

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 was hitting some issues with downloading the collectd dependency after clearing my local go cache and I needed to redirect to github to pick it up. I'm not exactly sure what happened but it looks like collectd.org stopped vending their package via collectd.org. The issue may have been resolved by now though, so I can try again.

../../../.gvm/pkgsets/go1.22.7/global/pkg/mod/github.com/aws/[email protected]/plugins/parsers/collectd/parser.go:8:2: unrecognized import path "collectd.org": https fetch: Get "https://collectd.org/?go-get=1": dial tcp: lookup collectd.org on 10.4.4.10:53: read udp 10.169.109.191:52627->10.4.4.10:53: i/o timeout


func (d *ExpHistogramDistribution) Size() int {
size := len(d.negativeBuckets) + len(d.positiveBuckets)
if d.zeroCount > 0 {
Copy link
Contributor

Choose a reason for hiding this comment

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

what is zeroCount? the number of datapoints with 0?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Yes, pretty much. OTLP exponential histograms splits the data into three section: negative values, zero values, positive values. Positive and negative buckets are defined separately by a series of buckets+counts. The zero values don't have any buckets so its just a counter stored in the histogram structure. Side note: the definition of "0" in OTLP exponential histograms is loose. Datapoints with a magnitude less than the configurable "zero threshold" is treated as 0.

}

func (d *ExpHistogramDistribution) Resize(_ int) []*ExpHistogramDistribution {
// TODO: split data points into separate PMD requests if the number of buckets exceeds the API limit
Copy link
Contributor

Choose a reason for hiding this comment

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

what happens if we exceed the API limit?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Based on the API documentation, I believe the PMD request will be rejected with error code 400 InvalidParameterValue, though I haven't tried.


// ValuesAndCounts outputs two arrays representing the midpoints of each exponential histogram bucket and the
// counter of datapoints within the corresponding exponential histogram buckets
func (d *ExpHistogramDistribution) ValuesAndCounts() ([]float64, []float64) {
Copy link
Contributor

Choose a reason for hiding this comment

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

nit: we can make the name of the function more descriptive. Maybe GetMidpointsAndCounts?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

This naming scheme was based off of the existing function in the Distribution interface. I think ValuesAndCounts is a fairly descriptive name as that's what is actually pushed to CloudWatch in the PMD request (an array of values and an array of counts).

d.min = min(d.min, value)
d.max = max(d.max, value)

if math.Abs(value) > d.zeroThreshold {
Copy link
Contributor

Choose a reason for hiding this comment

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

so if my zero threshold is 1, then both values 2 and -2 will be counted in the zero bucket?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Good catch, this logic is incorrect. In fact this function is left over work from trying to adhere to the Distribution interface that's not used nor needed. I'll remove this function entirely.


if math.Abs(value) > d.zeroThreshold {
d.zeroCount += uint64(weight)
} else if value > d.zeroThreshold {
Copy link
Contributor

Choose a reason for hiding this comment

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

would this ever happen? Wouldn't we always hit the first case?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Incorrect logic, but will be removing this function.

distList = resize(metric.distribution, c.config.MaxValuesPerDatum)
}
datums = c.buildMetricDatumDist(metric, dimensionsList)
} else if metric.expHistDistribution != nil {
Copy link
Contributor

Choose a reason for hiding this comment

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

is there ever a chance we will have both distribution and expHistDistribution not nil? Probably shouldn't happen...

Copy link
Contributor Author

Choose a reason for hiding this comment

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

No (at least they shouldn't...). The agent will convert the different OTLP datatypes into aggregationDatums in ConvertOtelMetrics. The existing ConvertOtelHistogramDataPoints sets distribution field and the new ConvertOtelExponentialHistogramDataPoints sets the expHistDistribution field.

@dricross dricross force-pushed the dricross/exponentialhistogram branch from 5362dc8 to d71cfd4 Compare July 23, 2025 12:24
@@ -78,34 +77,6 @@ func setNewDistributionFunc(maxValuesPerDatumLimit int) {
}
}

func resize(dist distribution.Distribution, listMaxSize int) (distList []distribution.Distribution) {
Copy link
Contributor Author

Choose a reason for hiding this comment

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

refactored as functions on each distribution

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.

3 participants