enhancement: allow configuring maximum number of metrics per DD metrics request payloads #476
Conversation
…s request payload
github-actions
bot
added
area/core
Regression Detector (DogStatsD)Regression Detector ResultsRun ID: 78007058-5937-40e4-898e-5f8ddb9d4c53 Baseline: 7.63.0-rc.2 Optimization Goals: ✅ No significant changes detected
|
| perf | experiment | goal | Δ mean % | Δ mean % CI | trials | links |
|---|---|---|---|---|---|---|
| ➖ | quality_gates_idle_rss | memory utilization | +0.42 | [+0.31, +0.53] | 1 | |
| ➖ | dsd_uds_10mb_3k_contexts | ingress throughput | +0.01 | [-0.00, +0.02] | 1 | |
| ➖ | dsd_uds_1mb_3k_contexts | ingress throughput | +0.00 | [-0.00, +0.01] | 1 | |
| ➖ | dsd_uds_100mb_250k_contexts | ingress throughput | +0.00 | [-0.00, +0.00] | 1 | |
| ➖ | dsd_uds_512kb_3k_contexts | ingress throughput | +0.00 | [-0.01, +0.01] | 1 | |
| ➖ | dsd_uds_1mb_50k_contexts | ingress throughput | +0.00 | [-0.00, +0.00] | 1 | |
| ➖ | dsd_uds_1mb_50k_contexts_memlimit | ingress throughput | +0.00 | [-0.00, +0.00] | 1 | |
| ➖ | dsd_uds_1mb_3k_contexts_dualship | ingress throughput | +0.00 | [-0.00, +0.00] | 1 | |
| ➖ | dsd_uds_100mb_3k_contexts | ingress throughput | -0.00 | [-0.04, +0.04] | 1 | |
| ➖ | dsd_uds_40mb_12k_contexts_40_senders | ingress throughput | -0.00 | [-0.01, +0.00] | 1 | |
| ➖ | dsd_uds_100mb_3k_contexts_distributions_only | memory utilization | -1.34 | [-1.49, -1.19] | 1 | |
| ➖ | dsd_uds_500mb_3k_contexts | ingress throughput | -2.59 | [-2.73, -2.45] | 1 |
Bounds Checks: ❌ Failed
| perf | experiment | bounds_check_name | replicates_passed | links |
|---|---|---|---|---|
| ❌ | quality_gates_idle_rss | memory_usage | 0/10 |
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:
-
Its estimated |Δ mean %| ≥ 5.00%, indicating the change is big enough to merit a closer look.
-
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.
-
Its configuration does not mark it "erratic".
Regression Detector (Saluki)Regression Detector ResultsRun ID: a3c92e88-6b12-403a-8b10-d3a49a649abc Baseline: 4cfc44e Optimization Goals: ✅ No significant changes detected
|
| perf | experiment | goal | Δ mean % | Δ mean % CI | trials | links |
|---|---|---|---|---|---|---|
| ➖ | quality_gates_idle_rss | memory utilization | +0.60 | [+0.57, +0.63] | 1 | |
| ➖ | dsd_uds_1mb_50k_contexts_memlimit | ingress throughput | +0.59 | [+0.33, +0.84] | 1 | |
| ➖ | dsd_uds_100mb_3k_contexts | ingress throughput | +0.01 | [-0.04, +0.06] | 1 | |
| ➖ | dsd_uds_40mb_12k_contexts_40_senders | ingress throughput | +0.01 | [-0.02, +0.04] | 1 | |
| ➖ | dsd_uds_50mb_10k_contexts_no_inlining | ingress throughput | +0.00 | [-0.06, +0.06] | 1 | |
| ➖ | dsd_uds_10mb_3k_contexts | ingress throughput | +0.00 | [-0.03, +0.03] | 1 | |
| ➖ | dsd_uds_1mb_3k_contexts | ingress throughput | -0.00 | [-0.00, +0.00] | 1 | |
| ➖ | dsd_uds_1mb_50k_contexts | ingress throughput | -0.00 | [-0.01, +0.00] | 1 | |
| ➖ | dsd_uds_1mb_3k_contexts_dualship | ingress throughput | -0.00 | [-0.01, +0.00] | 1 | |
| ➖ | dsd_uds_512kb_3k_contexts | ingress throughput | -0.00 | [-0.01, +0.01] | 1 | |
| ➖ | dsd_uds_100mb_250k_contexts | ingress throughput | -0.01 | [-0.05, +0.03] | 1 | |
| ➖ | dsd_uds_50mb_10k_contexts_no_inlining_no_allocs | ingress throughput | -0.02 | [-0.07, +0.04] | 1 | |
| ➖ | dsd_uds_100mb_3k_contexts_distributions_only | memory utilization | -0.15 | [-0.26, -0.04] | 1 | |
| ➖ | dsd_uds_500mb_3k_contexts | ingress throughput | -0.63 | [-0.75, -0.50] | 1 |
Bounds Checks: ✅ Passed
| perf | experiment | bounds_check_name | replicates_passed | links |
|---|---|---|---|---|
| ✅ | quality_gates_idle_rss | memory_usage | 10/10 |
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:
-
Its estimated |Δ mean %| ≥ 5.00%, indicating the change is big enough to merit a closer look.
-
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.
-
Its configuration does not mark it "erratic".
Regression Detector LinksExperiment Result Links
|
Summary
In #457, we changed around how we handle the ability to split oversized request payloads in the Datadog Metrics destination from storing the raw encoded metrics to storing the
Metricvalues themselves. This was done in order to improve the average memory consumed by the destination, as large event batches would tend to drive the buffers used to hold the encoded metrics up in size over time, which could waste significant amounts of memory in the long run.While switching to holding
Metricvalues directly provided more determinism --Metricis only ever X bytes, not variable -- it worsened the worst-case behavior because metrics can easily be encoded to a smaller size than that ofMetric, meaning that after encoding a certain number of metrics, holding theirMetricrepresentation becomes inefficient.In order to put an upper bound on this, we've introduced a "maximum metrics per payload" configuration that the request builders use. This means that we'll flush a request either when it's hit the (un)compressed size limits, or when it hits the maximum-metrics-per-payload limit.
This new configuration value --
serializer_max_metrics_per_payload-- operates slightly different from a nearly equivalent configuration value in the Datadog Agent:serializer_max_series_points_per_payload. This is due to the fact that the Datadog Agent is tracking the points that have been serialized, whereas we have to hold on to the entireMetric, so I wanted to keep the configuration setting named in a way that's faithful to the underlying behavior. However, all of this said, series/sketches generally have one point on average when flushed, so the number of metrics in a payload is also generally equal to the number of points in a payload. As such, we have the same default value of10000, meaning we'll allow us to 10,000 metrics per request payload.With this change, our calculated firm bounds for the Datadog Metrics component have dropped significantly, from ~69MB to ~6.6MB. In reality, after merging #457, the theoretical firm bound was closer to ~415MB, but I didn't bother trying to bring it true-to-life because it depended on an annoying calculation to determine the smallest valid metric we could encode, and how many of those we could fit per endpoint, and so on... easier to just make this follow-up PR. :)
Change Type
How did you test this PR?
This PR includes a unit test that asserts that the configured limit is obeyed. I also tested this out locally by sending a small number of metrics through DogStatsD and observing that multiple payloads were built, indicating that we were flushing earlier than we normally would have otherwise, since all metrics would fit within the configured (un)compressed size limits.
References
N/A