Add delayed startup RSS shrinker

[scottopell]

What does this PR do?

Adds a delayed, one-shot startup RSS shrinker for Linux Agent processes.

The shrinker runs 2m30s after process startup and performs a best-effort reclamation pass:

• debug.FreeOSMemory() to return reclaimable Go heap pages to the OS.
• MADV_PAGEOUT over clean/read-only file-backed mappings to let Linux page out cold startup-touched code/data.
• Optional malloc_trim(0) only when DD_STARTUP_RSS_SHRINKER_MALLOC_TRIM=true is set on linux && cgo builds.

The scheduled shrinker can be disabled with:

DD_STARTUP_RSS_SHRINKER_DISABLED=true

This also extracts the existing duplicated trace-loader / installer-daemon memory release implementation into cmd/internal/rssshrinker.

Motivation

This is intentionally framed as an RSS presentation optimization, not a claim that PSS or cgroup memory pressure will necessarily improve.

Many users inspect Agent footprint through RSS-oriented tools like top and ps, even though PSS/cgroup measurements are better indicators of real memory pressure. After startup, the Agent has touched clean file-backed pages that may no longer be part of the active working set. A single delayed, best-effort reclamation pass gives Linux a chance to evict those pages so reported RSS better reflects the active working set.

The pass is delayed and one-shot to keep the behavior simple and avoid periodic page-fault churn.

Describe how you validated your changes

Used the repo Go toolchain by putting Homebrew Go first in PATH:

export PATH=/opt/homebrew/opt/go/bin:$PATH

Validated with:

dda inv test --targets=./cmd/internal/rssshrinker
dda inv agent.build --build-exclude=systemd
git diff --check

Push hooks were attempted without --no-verify. They passed go-mod-tidy, go-mod-replace, and go-test, but the local go-linter hook was blocked by a local tool version mismatch:

Expecting golangci-lint '2.12.2' but you have golangci-lint '2.9.0'.

After confirming the remaining failure was local tooling, the branch was pushed with --no-verify so CI/perf testing can run.

Additional Notes

Automated performance testing should be used to compare RSS/PSS/cgroup memory and page-fault/latency behavior before deciding whether to enable, tune, or further expand this behavior.

[datadog-datadog-us1-prod]

🎯 Code Coverage (details)
• Patch Coverage: 0.00%
• Overall Coverage: 51.73% (+0.10%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: b021bea | Docs | Datadog PR Page | Give us feedback!}

[dd-octo-sts]

Files inventory check summary

File checks results against ancestor a77444b5:

Results for datadog-agent_7.82.0~devel.git.532.4408c0f.pipeline.122421258-1_amd64.deb:

No change detected

[dd-octo-sts]

Static quality checks

✅ Please find below the results from static quality gates
Comparison made with ancestor fc18d0b
📊 Static Quality Gates Dashboard
🔗 SQG Job
SOME SIZE DELTAS ARE N/A (ANCESTOR METRICS NOT YET AVAILABLE). RETRY JOB

Successful checks

Info

	Quality gate	Change	Size (prev → curr → max)
✅	agent_deb_amd64	N/A	N/A → 750.425 → 758.200
✅	agent_deb_amd64_fips	N/A	N/A → 705.442 → 709.840
✅	agent_heroku_amd64	N/A	N/A → 307.493 → 315.230
✅	agent_msi	N/A	N/A → 645.032 → 656.640
✅	agent_rpm_amd64	N/A	N/A → 750.409 → 758.170
✅	agent_rpm_amd64_fips	N/A	N/A → 705.425 → 709.840
✅	agent_rpm_arm64	N/A	N/A → 725.438 → 729.660
✅	agent_rpm_arm64_fips	N/A	N/A → 684.190 → 688.860
✅	agent_suse_amd64	N/A	N/A → 750.409 → 758.170
✅	agent_suse_amd64_fips	N/A	N/A → 705.425 → 709.840
✅	agent_suse_arm64	N/A	N/A → 725.438 → 729.660
✅	agent_suse_arm64_fips	N/A	N/A → 684.190 → 688.860
✅	docker_agent_amd64	N/A	N/A → 809.144 → 813.790
✅	docker_agent_arm64	N/A	N/A → 809.228 → 815.030
✅	docker_agent_jmx_amd64	N/A	N/A → 1000.041 → 1004.550
✅	docker_agent_jmx_arm64	N/A	N/A → 988.778 → 994.710
✅	docker_cluster_agent_amd64	N/A	N/A → 209.691 → 210.470
✅	docker_cluster_agent_arm64	N/A	N/A → 222.765 → 222.980
✅	docker_cws_instrumentation_amd64	N/A	N/A → 7.447 → 7.480
✅	docker_cws_instrumentation_arm64	N/A	N/A → 6.877 → 7.110
✅	docker_dogstatsd_amd64	N/A	N/A → 39.196 → 39.910
✅	docker_dogstatsd_arm64	N/A	N/A → 37.305 → 38.270
✅	docker_host_profiler_amd64	N/A	N/A → 302.586 → 317.640
✅	docker_host_profiler_arm64	N/A	N/A → 314.119 → 328.900
✅	dogstatsd_deb_amd64	N/A	N/A → 29.935 → 31.150
✅	dogstatsd_deb_arm64	N/A	N/A → 27.967 → 29.530
✅	dogstatsd_rpm_amd64	N/A	N/A → 29.935 → 31.150
✅	dogstatsd_suse_amd64	N/A	N/A → 29.935 → 31.150
✅	iot_agent_deb_amd64	N/A	N/A → 46.136 → 46.380
✅	iot_agent_deb_arm64	N/A	N/A → 42.816 → 43.720
✅	iot_agent_deb_armhf	N/A	N/A → 43.596 → 43.960
✅	iot_agent_rpm_amd64	N/A	N/A → 46.136 → 46.380
✅	iot_agent_suse_amd64	N/A	N/A → 46.135 → 46.380

[cit-pr-commenter-54b7da]

Regression Detector

Regression Detector Results

Metrics dashboard
Target profiles
Run ID: 11ae1ebe-276b-41b3-aa5e-a80c827f156b

Baseline: fc18d0b
Comparison: b021bea
Diff

Optimization Goals: ✅ Improvement(s) detected

perf	experiment	goal	Δ mean %	Δ mean % CI	trials	links
✅	quality_gate_metrics_logs	memory utilization	-9.29	[-9.55, -9.03]	1	Logs bounds checks dashboard
✅	quality_gate_security_mean_fs_load	memory utilization	-18.42	[-18.61, -18.23]	1	Logs bounds checks dashboard
✅	quality_gate_security_idle	memory utilization	-18.74	[-18.97, -18.51]	1	Logs bounds checks dashboard
✅	quality_gate_security_no_fs_load	memory utilization	-20.08	[-20.35, -19.81]	1	Logs bounds checks dashboard
✅	quality_gate_idle	memory utilization	-21.35	[-21.57, -21.14]	1	Logs bounds checks dashboard
✅	quality_gate_idle_all_features	memory utilization	-29.93	[-30.23, -29.63]	1	Logs bounds checks dashboard

Fine details of change detection per experiment

perf	experiment	goal	Δ mean %	Δ mean % CI	trials	links
➖	quality_gate_logs	% cpu utilization	+1.68	[+0.69, +2.67]	1	Logs bounds checks dashboard
✅	quality_gate_metrics_logs	memory utilization	-9.29	[-9.55, -9.03]	1	Logs bounds checks dashboard
✅	quality_gate_security_mean_fs_load	memory utilization	-18.42	[-18.61, -18.23]	1	Logs bounds checks dashboard
✅	quality_gate_security_idle	memory utilization	-18.74	[-18.97, -18.51]	1	Logs bounds checks dashboard
✅	quality_gate_security_no_fs_load	memory utilization	-20.08	[-20.35, -19.81]	1	Logs bounds checks dashboard
✅	quality_gate_idle	memory utilization	-21.35	[-21.57, -21.14]	1	Logs bounds checks dashboard
✅	quality_gate_idle_all_features	memory utilization	-29.93	[-30.23, -29.63]	1	Logs bounds checks dashboard

Bounds Checks: ✅ Passed

perf	experiment	bounds_check_name	replicates_passed	observed_value	links
✅	quality_gate_idle	intake_connections	10/10	3 ≤ 4	bounds checks dashboard
✅	quality_gate_idle	memory_usage	10/10	146.98MiB ≤ 154MiB	bounds checks dashboard
✅	quality_gate_idle	total_bytes_received	10/10	579.86KiB ≤ 819.20KiB	bounds checks dashboard
✅	quality_gate_idle_all_features	intake_connections	10/10	3 ≤ 4	bounds checks dashboard
✅	quality_gate_idle_all_features	memory_usage	10/10	484.54MiB ≤ 495MiB	bounds checks dashboard
✅	quality_gate_idle_all_features	total_bytes_received	10/10	0.89MiB ≤ 1.25MiB	bounds checks dashboard
✅	quality_gate_logs	intake_connections	10/10	3 ≤ 6	bounds checks dashboard
✅	quality_gate_logs	memory_usage	10/10	180.97MiB ≤ 195MiB	bounds checks dashboard
✅	quality_gate_logs	missed_bytes	10/10	0B = 0B	bounds checks dashboard
✅	quality_gate_logs	total_bytes_received	10/10	264.01MiB ≤ 292MiB	bounds checks dashboard
✅	quality_gate_metrics_logs	cpu_usage	10/10	373.74 ≤ 2000	bounds checks dashboard
✅	quality_gate_metrics_logs	intake_connections	10/10	3 ≤ 6	bounds checks dashboard
✅	quality_gate_metrics_logs	memory_usage	10/10	394.17MiB ≤ 430MiB	bounds checks dashboard
✅	quality_gate_metrics_logs	missed_bytes	10/10	0B = 0B	bounds checks dashboard
✅	quality_gate_metrics_logs	total_bytes_received	10/10	0.86GiB ≤ 1.04GiB	bounds checks dashboard
✅	quality_gate_security_idle	cpu_usage	10/10	38.50 ≤ 40	bounds checks dashboard
✅	quality_gate_security_idle	memory_usage	10/10	299.60MiB ≤ 330MiB	bounds checks dashboard
✅	quality_gate_security_mean_fs_load	cpu_usage	10/10	62.13 ≤ 80	bounds checks dashboard
✅	quality_gate_security_mean_fs_load	memory_usage	10/10	274.84MiB ≤ 310MiB	bounds checks dashboard
✅	quality_gate_security_no_fs_load	cpu_usage	10/10	24.29 ≤ 40	bounds checks dashboard
✅	quality_gate_security_no_fs_load	memory_usage	10/10	287.37MiB ≤ 320MiB	bounds checks dashboard

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 Profiling Replicate Execution Failures (ddprof)

Note: Profiling replicas may still be executing. See the debug dashboard for up to date status.

Experiment	Variant	Replicates	Failure	Debug Dashboard
quality_gate_idle_all_features	baseline	10	Oom killed	Debug Dashboard
quality_gate_idle_all_features	comparison	10	Oom killed	Debug Dashboard
quality_gate_security_mean_fs_load	comparison	10	Oom killed	Debug Dashboard

CI Pass/Fail Decision

✅ Passed. All Quality Gates passed.

• quality_gate_security_idle, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_security_idle, bounds check cpu_usage: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check cpu_usage: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check intake_connections: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check total_bytes_received: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check missed_bytes: 10/10 replicas passed. Gate passed.
• quality_gate_idle, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_idle, bounds check intake_connections: 10/10 replicas passed. Gate passed.
• quality_gate_idle, bounds check total_bytes_received: 10/10 replicas passed. Gate passed.
• quality_gate_security_no_fs_load, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_security_no_fs_load, bounds check cpu_usage: 10/10 replicas passed. Gate passed.
• quality_gate_security_mean_fs_load, bounds check cpu_usage: 10/10 replicas passed. Gate passed.
• quality_gate_security_mean_fs_load, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_logs, bounds check intake_connections: 10/10 replicas passed. Gate passed.
• quality_gate_logs, bounds check missed_bytes: 10/10 replicas passed. Gate passed.
• quality_gate_logs, bounds check total_bytes_received: 10/10 replicas passed. Gate passed.
• quality_gate_logs, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_idle_all_features, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_idle_all_features, bounds check total_bytes_received: 10/10 replicas passed. Gate passed.
• quality_gate_idle_all_features, bounds check intake_connections: 10/10 replicas passed. Gate passed.

[pgimalac]

Is this on purpose ?

[scottopell]

quasi, I usually delete these opening a PR for review, but I also think there is some value in recording the LLM prompt/task along with the code, so I've debated making a directory for these

[pgimalac]

How did you chose this default delay ?

[scottopell]

fairly arbitrary, my goal was to hit a point in the agent startup where all ‘init’-ish things already happened, and still give plenty of good measurements within the 10min SMP run

[scottopell]

after lots of back and forth with gpt-5.5 here's where I ended up

---

What exactly did this latest experiment do?

For the latest 4408c0f run, the code:

1. Pages out every read-only file-backed mapping:
   • includes r--p
   • includes r-xp
2. Then calls:

unix.Fadvise(fd, offset, length, unix.FADV_WILLNEED)

on the same file offset/range for every mapping we paged out.

So it is not only executable mappings. It is the exact same read-only mapping set as the original high-memory-win/high-CPU version, but with FADV_WILLNEED after MADV_PAGEOUT.

Headline result

FADV_WILLNEED did not fix the CPU/refault problem.

It restored the big memory win, but CPU and page-fault/reclaim behavior look almost identical to the original bad r-xp + r--p run.

Three-way comparison: quality_gate_idle_all_features

Mode	CPU delta	PSS delta	RSS delta
old r-xp+r--p, no prefetch	+80.1 mCPU / +124%	-147.4 MiB / -32%	-145.3 MiB / -30%
skip executable, r--p only	+5.2 mCPU / +8%	-52.7 MiB / -11%	-52.5 MiB / -11%
r-xp+r--p + FADV_WILLNEED	+83.1 mCPU / +125%	-146.4 MiB / -31%	-143.9 MiB / -30%

So prefetch keeps the RSS/PSS win, but the sustained CPU problem remains.

Page fault / reclaim counters

This is the most important part.

workingset_refault_file

Mode	Comparison - baseline
old no-prefetch	+14,140,316
skip executable	+569,652
FADV_WILLNEED	+14,301,898

pgmajfault

Mode	Comparison - baseline
old no-prefetch	+1,555
skip executable	+103
FADV_WILLNEED	+1,553

pgscan_direct

Mode	Comparison - baseline
old no-prefetch	+20,890,714
skip executable	+994,158
FADV_WILLNEED	+21,429,315

pgsteal_direct

Mode	Comparison - baseline
old no-prefetch	+14,142,829
skip executable	+554,354
FADV_WILLNEED	+14,319,642

So FADV_WILLNEED did not meaningfully reduce the refault/reclaim storm.

Cgroup memory impact

You asked about cgroup impact. memory.current goes up in all variants, including the “successful” skip-exec one.

Mode	cgroup.v2.memory.current comparison - baseline
old no-prefetch	+31.2 MiB / +6%
skip executable	+29.9 MiB / +5%
FADV_WILLNEED	+34.6 MiB / +6%

So the RSS/PSS win does not translate into lower cgroup memory usage. In fact, cgroup current is modestly higher.

File cache-related stats:

Mode	memory.stat.file comparison - baseline
old no-prefetch	+10.7 MiB
skip executable	+35.4 MiB
FADV_WILLNEED	+12.4 MiB

The surprising bit: FADV_WILLNEED did not produce a dramatically larger file-cache footprint than old no-prefetch. It also did not fix refaults. That suggests the fadvise either:

• did not eagerly read enough to matter;
• got reclaimed again under cgroup pressure;
• or does not populate the specific page-table / workingset state needed to avoid later refault cost.

Interpretation

This experiment falsifies the optimistic hypothesis:

“If we page out r-xp+r--p but prefetch the backing file range, maybe we keep RSS low and avoid sustained CPU.”

Nope. At least with FADV_WILLNEED, we get:

big RSS/PSS win + big refault/reclaim CPU tax

almost exactly like the original.

The only version that actually changed the CPU/refault picture was:

skip executable mappings

Current best conclusion

Executable mappings are the key. Paging out r-xp causes the kernel to thrash on file-backed workingset refault/direct reclaim during normal idle activity. Trying to warm the file cache with FADV_WILLNEED does not prevent it.

So the best-supported implementation remains:

page out r--p only, skip r-xp

That gives:

• ~11–12% PSS/RSS reduction in idle_all_features
• no major CPU regression
• ~18–25x lower refault/reclaim counters than r-xp+r--p

Recommendation

Revert the FADV_WILLNEED experiment and keep the skip-executable version.

If we want another experiment, the next useful one is not more prefetching. It would be decomposing:

1. FreeOSMemory() only
2. r--p pageout only
3. both

But for the “can we safely reclaim executable mappings?” question, this latest data says: not with this approach.