chore(ddsketch, dogstatsd): allow for higher sample rates in DSD payloads

[tobz]

Summary

This PR fixes some suboptimal behavior with the clamped sample rate support added to the DogStatsD codec by changing the count field in DDSketch to be u64, and updating the default minimum sample rate, as well as allowing it to be configured.

In #1024, we made a change to enforce a "minimum" sample rate in DogStatsD payloads. The thinking went: when payloads send very high (small numbers) sample rates, this can lead to us trying to add a number of samples to a DDSketch that overflows the count field which tracks the total number of samples in the sketch. Very high sample rates should be rare: rare in terms of intentional usage, and more often than not related to some sort of bug, or worse, an attempted denial-of-service/resource exhaustion attack on the Agent. We felt the trade-off made sense, and diligently picked a value to try and ensure that a DogStatsD payload with default settings couldn't trigger this overflow behavior.

As you might surmise, we discovered that this limit is too conservative in practice.

This PR takes a multi-faceted approach to resolving the issue: updating the default minimum sample rate and increasing the size of the count field (the one we wanted to protect from overflow) in DDSketch.

Updating the default minimum sample rate

The value we chose was, essentially, based on what a single packet could reasonably do: if you sent a multi-value payload, where we apply the sample rate to each value in the payload, and based on the maximum allowed DSD payload size by default, and so on and so forth... you arrive at a sample rate of 0.000001, or around 1 million samples. This helps prevent overflow panics on receive, although it doesn't stop potential overflow panics when fully maxed-out sketches are merged. Anyways...

In practice, the true limit is more related to the maximum number of bins allowed: if we can only ever have N bins in a sketch, and each bin can represent a maximum count of Y, then our upper bound for samples-per-sketch should be X * Y. Given the default maximum bin count of 4096, and maximum of 65,536 samples per bin... we arrive at 268,435,456, which we trim to 260,000,000 just to be safe.. and that gives us our new default minimum sample rate of 0.000000003845.

We've additionally exposed the ability to change this in the DogStatsD source itself: the default value there matches the default in the codec, but it's useful to be able to potentially ratchet down for internal testing to see where a safer limit might be.

Updating the count field in DDSketch to be u64

Increasing the sample rate doesn't help us avoid overflow if the count field is still u32, so we've bumped this to u64. "Why is this safe, though? Why was it u32 in the first place?" you may ask.

That's a good question and in the Agent codebase, the count field is an int64, which is consistent with the corresponding field type in the Protocol Buffers representation for sketches. I still don't really fully know why I originally used u32, but my best guess is that it allowed us to transparently convert to f64, without any loss of precision, when doing internal operations that worked with the count as f64.

In practice, given the bin limit and max count per bin, we need something bigger than u32 and there's just no way around that. We shouldn't hit any issues with f64 conversion loss until our counts are greater than 2^53, which is much higher than ~260M... at which point we'd have bigger issues because we would have been losing bins during post-insert trimming anyways.

Similarly, even though we're eventually converting from u64 to i64, we'd have to have counts over 2^63 for the conversion to be clamped/invalid... and we've already lost data well before that.

Change Type

• Bug fix
• New feature
• Non-functional (chore, refactoring, docs)
• Performance

How did you test this PR?

Existing unit and correctness tests.

References

AGTMETRICS-393

[tobz]

• chore(ddsketch, dogstatsd): allow for higher sample rates in DSD payloads #1087 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[pr-commenter]

Binary Size Analysis (Agent Data Plane)

Target: 53531a0 (baseline) vs 560f391 (comparison) diff
Baseline Size: 336.44 MiB
Comparison Size: 336.14 MiB
Size Change: -311.35 KiB (-0.09%)
Pass/Fail Threshold: +5%
Result: PASSED ✅

Changes by Module

Module	File Size	Symbols
[debug sections]	-289.37 KiB	7
prost	-18.01 KiB	168
rustls	+17.15 KiB	11
tokio_rustls	-16.13 KiB	6
figment	-12.61 KiB	39
otlp_protos::otlp_include::opentelemetry	+11.44 KiB	116
h2	-8.06 KiB	73
saluki_components::transforms::aggregate	+8.03 KiB	8
hyper_util	+5.78 KiB	20
[Unmapped]	-5.12 KiB	1
tokio_util	+4.94 KiB	12
anon.031a1f69be40723c9e961e9fc98b02a9.0.llvm.576639265174890831	+4.92 KiB	1
anon.8a24369cc75f65645c1607e1d511e84c.369.llvm.8512966183790337429	-4.92 KiB	1
futures_util	-4.03 KiB	24
saluki_components::transforms::host_tags	+3.82 KiB	1
saluki_components::sources::dogstatsd	-3.73 KiB	19
saluki_components::common::datadog	+3.29 KiB	22
core	+3.10 KiB	1563
serde	+2.94 KiB	26
quick_cache	-2.93 KiB	21

Detailed Symbol Changes

    FILE SIZE        VM SIZE    
 --------------  -------------- 
  +0.1%  +137Ki  [ = ]       0    [section .debug_loc]
  [NEW] +12.2Ki  [NEW] +12.1Ki    saluki_components::sources::otlp::logs::transform::transform_log_record::h663796474e3b4bce
  [NEW] +9.05Ki  [NEW] +8.83Ki    _<saluki_components::sources::dogstatsd::DogStatsDConfiguration as saluki_core::components::sources::builder::SourceBuilder>::build::_{{closure}}::h494a110866a2997d
  [NEW] +8.13Ki  [NEW] +7.89Ki    saluki_components::transforms::aggregate::config::_::_<impl serde_core::de::Deserialize for saluki_components::transforms::aggregate::config::HistogramConfiguration>::deserialize::haa284aece90cfde3
  [NEW] +7.91Ki  [NEW] +7.80Ki    rustls::conn::ConnectionCore<Data>::process_new_packets::he334f777dd629005
  [NEW] +7.80Ki  [NEW] +7.69Ki    rustls::conn::ConnectionCore<Data>::process_new_packets::h79dbc8cff7cd06be
  [NEW] +5.57Ki  [NEW] +5.41Ki    _<serde::private::de::FlatMapDeserializer<E> as serde_core::de::Deserializer>::deserialize_struct::hd2e74a7bf9e626b0
  [NEW] +5.31Ki  [NEW] +5.14Ki    _<serde::private::de::FlatMapDeserializer<E> as serde_core::de::Deserializer>::deserialize_struct::hb1e73cc751944c6e
  [NEW] +4.92Ki  [NEW]     +81    anon.031a1f69be40723c9e961e9fc98b02a9.0.llvm.576639265174890831
  [DEL] -4.87Ki  [DEL] -4.70Ki    _<serde::private::de::FlatMapDeserializer<E> as serde_core::de::Deserializer>::deserialize_struct::hc2393939d6a764b8
  [DEL] -4.92Ki  [DEL]     -81    anon.8a24369cc75f65645c1607e1d511e84c.369.llvm.8512966183790337429
 -64.8% -5.12Ki  [ = ]       0    [Unmapped]
 -87.6% -7.99Ki -88.6% -7.99Ki    tokio_rustls::common::Stream<IO,C>::read_io::hff84ac4f60c968ea
 -88.2% -8.44Ki -89.1% -8.44Ki    tokio_rustls::common::Stream<IO,C>::read_io::hb24087c6c4f8882c
  [DEL] -8.80Ki  [DEL] -8.58Ki    _<saluki_components::sources::dogstatsd::DogStatsDConfiguration as saluki_core::components::sources::builder::SourceBuilder>::build::_{{closure}}::h19f3155b19bb0e2b
 -56.7% -10.5Ki -57.0% -10.5Ki    figment::figment::Figment::extract::hef8645c13a353b1a
 -85.1% -11.6Ki -86.2% -11.6Ki    _<saluki_components::sources::otlp::logs::translator::OtlpLogsTranslator as core::iter::traits::iterator::Iterator>::next::hf976c0d95e31e00a
  -0.0% -21.3Ki  [ = ]       0    [section .debug_str]
  -0.1% -25.7Ki  -0.4% -22.0Ki    [7716 Others]
  -0.1% -27.6Ki  [ = ]       0    [section .debug_line]
  -0.3%  -373Ki  [ = ]       0    [section .debug_info]
  -0.1%  -311Ki  -0.1% -18.9Ki    TOTAL

[pr-commenter]

Regression Detector (Agent Data Plane)

Regression Detector Results

Run ID: 6c8c59c5-d13f-4304-9b60-dbc43da53b75

Baseline: 53531a0
Comparison: 560f391
Diff

❌ Experiments with retried target crashes

This is a critical error. One or more replicates failed with a non-zero exit code. These replicates may have been retried. See Replicate Execution Details for more information.

• quality_gates_rss_dsd_medium
• quality_gates_rss_dsd_ultraheavy
• otlp_ingest_metrics_adp

Optimization Goals: ❌ Regression(s) detected

perf	experiment	goal	Δ mean %	Δ mean % CI	trials	links
❌	otlp_ingest_metrics_adp	memory utilization	+6.22	[+6.00, +6.44]	1

Fine details of change detection per experiment

perf	experiment	goal	Δ mean %	Δ mean % CI	trials	links
❌	otlp_ingest_metrics_adp	memory utilization	+6.22	[+6.00, +6.44]	1
➖	dsd_uds_500mb_3k_contexts_throughput	ingress throughput	+2.07	[+1.93, +2.21]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_logs_adp	memory utilization	+0.88	[+0.42, +1.33]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_idle	memory utilization	+0.21	[+0.18, +0.24]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_dsd_ultraheavy	ingress throughput	+0.00	[-0.06, +0.07]	1	(metrics) (profiles) (logs)
➖	dsd_uds_1mb_3k_contexts_throughput	ingress throughput	+0.00	[-0.06, +0.06]	1	(metrics) (profiles) (logs)
➖	dsd_uds_512kb_3k_contexts_throughput	ingress throughput	-0.00	[-0.06, +0.05]	1	(metrics) (profiles) (logs)
➖	dsd_uds_100mb_3k_contexts_throughput	ingress throughput	-0.01	[-0.05, +0.04]	1	(metrics) (profiles) (logs)
➖	dsd_uds_10mb_3k_contexts_throughput	ingress throughput	-0.02	[-0.17, +0.13]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_dsd_medium	memory utilization	-0.08	[-0.26, +0.10]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_dsd_low	memory utilization	-0.13	[-0.27, +0.01]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_dsd_heavy	memory utilization	-0.21	[-0.34, -0.08]	1	(metrics) (profiles) (logs)

Bounds Checks: ✅ Passed

perf	experiment	bounds_check_name	replicates_passed	links
✅	quality_gates_rss_dsd_heavy	memory_usage	10/10	(metrics) (profiles) (logs)
✅	quality_gates_rss_dsd_low	memory_usage	10/10	(metrics) (profiles) (logs)
✅	quality_gates_rss_dsd_medium	memory_usage	10/10	(metrics) (profiles) (logs)
✅	quality_gates_rss_dsd_ultraheavy	memory_usage	10/10	(metrics) (profiles) (logs)
✅	quality_gates_rss_idle	memory_usage	10/10	(metrics) (profiles) (logs)

Explanation

Confidence level: 90.00%
Effect size tolerance: |Δ mean %| ≥ 5.00%

Performance changes are noted in the perf column of each table:

• ✅ = significantly better comparison variant performance
• ❌ = significantly worse comparison variant performance
• ➖ = no significant change in performance

A regression test is an A/B test of target performance in a repeatable rig, where "performance" is measured as "comparison variant minus baseline variant" for an optimization goal (e.g., ingress throughput). Due to intrinsic variability in measuring that goal, we can only estimate its mean value for each experiment; we report uncertainty in that value as a 90.00% confidence interval denoted "Δ mean % CI".

For each experiment, we decide whether a change in performance is a "regression" -- a change worth investigating further -- if all of the following criteria are true:

1. Its estimated |Δ mean %| ≥ 5.00%, indicating the change is big enough to merit a closer look.

2. Its 90.00% confidence interval "Δ mean % CI" does not contain zero, indicating that if our statistical model is accurate, there is at least a 90.00% chance there is a difference in performance between baseline and comparison variants.

3. Its configuration does not mark it "erratic".

Replicate Execution Details

We run multiple replicates for each experiment/variant. However, we allow replicates to be automatically retried if there are any failures, up to 8 times, at which point the replicate is marked dead and we are unable to run analysis for the entire experiment. We call each of these attempts at running replicates a replicate execution. This section lists all replicate executions that failed due to the target crashing or being oom killed.

Note: In the below tables we bucket failures by experiment, variant, and failure type. For each of these buckets we list out the replicate indexes that failed with an annotation signifying how many times said replicate failed with the given failure mode. In the below example the baseline variant of the experiment named experiment_with_failures had two replicates that failed by oom kills. Replicate 0, which failed 8 executions, and replicate 1 which failed 6 executions, all with the same failure mode.

Experiment	Variant	Replicates	Failure	Logs	Debug Dashboard
experiment_with_failures	baseline	0 (x8) 1 (x6)	Oom killed		Debug Dashboard

The debug dashboard links will take you to a debugging dashboard specifically designed to investigate replicate execution failures.

❌ Retried Normal Replicate Execution Failures (non-profiling)

Experiment	Variant	Replicates	Failure	Debug Dashboard
otlp_ingest_metrics_adp	comparison	6	Failed to shutdown when requested	Debug Dashboard
quality_gates_rss_dsd_medium	baseline	2	Failed to shutdown when requested	Debug Dashboard
quality_gates_rss_dsd_ultraheavy	baseline	3	Failed to shutdown when requested	Debug Dashboard
quality_gates_rss_dsd_ultraheavy	comparison	0, 3	Failed to shutdown when requested	Debug Dashboard

[rayz]

LGTM