feat(io): Retry transactions if secrets are enabled

[atanzu]

Summary

Retry transient 403 Forbidden responses in the Datadog forwarder when secrets management is active, so requests in flight during an API key refresh aren't dropped.

Change Type

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

How did you test this PR?

• Unit tests for StandardHttpClassifier predicate management (with_/set_/remove_status_code_predicate) and DefaultHttpRetryPolicy::with_backoff_and_classifier.
• Unit tests for is_secrets_in_use across the four secret_* config keys.
• Unit test that mutates the underlying GenericConfiguration and confirms the same retry-policy instance flips its 403 decision.
• End-to-end integration tests in lib/saluki-components/src/common/datadog/io.rs using a test TCP server, asserting 403 is not retried without secrets and is retried when secrets are in use.

References

• AGTMETRICS-504

[pr-commenter]

Regression Detector (Agent Data Plane)

Regression Detector Results

Run ID: 7b1cf41e-acd0-4d24-8a74-a4a32e128e6a

Baseline: 7c22c4d
Comparison: 001ef79
Diff

Optimization Goals: ✅ No significant changes detected

Experiments ignored for regressions

Regressions in experiments with settings containing erratic: true are ignored.

perf	experiment	goal	Δ mean %	Δ mean % CI	trials	links
➖	otlp_ingest_logs_5mb_throughput	ingress throughput	+0.02	[-0.10, +0.13]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_logs_5mb_cpu	% cpu utilization	-0.17	[-4.86, +4.52]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_logs_5mb_memory	memory utilization	-4.51	[-4.93, -4.08]	1	(metrics) (profiles) (logs)

Fine details of change detection per experiment

perf	experiment	goal	Δ mean %	Δ mean % CI	trials	links
➖	otlp_ingest_metrics_5mb_memory	memory utilization	+2.90	[+2.68, +3.11]	1	(metrics) (profiles) (logs)
➖	dsd_uds_100mb_3k_contexts_cpu	% cpu utilization	+1.67	[-3.85, +7.20]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_dsd_medium	memory utilization	+0.38	[+0.21, +0.55]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_idle	memory utilization	+0.37	[+0.34, +0.41]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_traces_ottl_transform_5mb_memory	memory utilization	+0.20	[+0.04, +0.36]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_traces_ottl_filtering_5mb_throughput	ingress throughput	+0.12	[+0.04, +0.19]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_traces_ottl_transform_5mb_throughput	ingress throughput	+0.10	[+0.03, +0.18]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_traces_5mb_throughput	ingress throughput	+0.10	[+0.02, +0.17]	1	(metrics) (profiles) (logs)
➖	dsd_uds_500mb_3k_contexts_memory	memory utilization	+0.09	[-0.07, +0.24]	1	(metrics) (profiles) (logs)
➖	dsd_uds_500mb_3k_contexts_cpu	% cpu utilization	+0.06	[-1.36, +1.49]	1	(metrics) (profiles) (logs)
➖	dsd_uds_100mb_3k_contexts_throughput	ingress throughput	+0.02	[-0.02, +0.06]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_logs_5mb_throughput	ingress throughput	+0.02	[-0.10, +0.13]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_metrics_5mb_throughput	ingress throughput	+0.01	[-0.19, +0.20]	1	(metrics) (profiles) (logs)
➖	dsd_uds_512kb_3k_contexts_memory	memory utilization	+0.00	[-0.14, +0.14]	1	(metrics) (profiles) (logs)
➖	dsd_uds_1mb_3k_contexts_throughput	ingress throughput	+0.00	[-0.05, +0.06]	1	(metrics) (profiles) (logs)
➖	dsd_uds_10mb_3k_contexts_throughput	ingress throughput	+0.00	[-0.19, +0.19]	1	(metrics) (profiles) (logs)
➖	dsd_uds_512kb_3k_contexts_throughput	ingress throughput	-0.00	[-0.05, +0.05]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_dsd_low	memory utilization	-0.06	[-0.22, +0.10]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_dsd_heavy	memory utilization	-0.11	[-0.23, +0.01]	1	(metrics) (profiles) (logs)
➖	dsd_uds_100mb_3k_contexts_memory	memory utilization	-0.15	[-0.30, -0.00]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_traces_ottl_transform_5mb_cpu	% cpu utilization	-0.15	[-2.20, +1.89]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_logs_5mb_cpu	% cpu utilization	-0.17	[-4.86, +4.52]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_traces_ottl_filtering_5mb_memory	memory utilization	-0.18	[-0.42, +0.06]	1	(metrics) (profiles) (logs)
➖	quality_gates_rss_dsd_ultraheavy	memory utilization	-0.23	[-0.36, -0.10]	1	(metrics) (profiles) (logs)
➖	dsd_uds_10mb_3k_contexts_memory	memory utilization	-0.27	[-0.42, -0.12]	1	(metrics) (profiles) (logs)
➖	dsd_uds_500mb_3k_contexts_throughput	ingress throughput	-0.31	[-0.46, -0.16]	1	(metrics) (profiles) (logs)
➖	dsd_uds_1mb_3k_contexts_memory	memory utilization	-0.34	[-0.47, -0.20]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_traces_5mb_memory	memory utilization	-0.36	[-0.52, -0.21]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_traces_5mb_cpu	% cpu utilization	-0.53	[-2.66, +1.60]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_traces_ottl_filtering_5mb_cpu	% cpu utilization	-1.45	[-3.73, +0.83]	1	(metrics) (profiles) (logs)
➖	dsd_uds_10mb_3k_contexts_cpu	% cpu utilization	-1.55	[-31.13, +28.03]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_metrics_5mb_cpu	% cpu utilization	-2.79	[-8.91, +3.32]	1	(metrics) (profiles) (logs)
➖	dsd_uds_1mb_3k_contexts_cpu	% cpu utilization	-3.03	[-55.65, +49.60]	1	(metrics) (profiles) (logs)
➖	otlp_ingest_logs_5mb_memory	memory utilization	-4.51	[-4.93, -4.08]	1	(metrics) (profiles) (logs)
➖	dsd_uds_512kb_3k_contexts_cpu	% cpu utilization	-10.84	[-64.39, +42.71]	1	(metrics) (profiles) (logs)

Bounds Checks: ✅ Passed

perf	experiment	bounds_check_name	replicates_passed	observed_value	links
✅	quality_gates_rss_dsd_heavy	memory_usage	10/10	123.74MiB ≤ 140MiB	(metrics) (profiles) (logs)
✅	quality_gates_rss_dsd_low	memory_usage	10/10	40.26MiB ≤ 50MiB	(metrics) (profiles) (logs)
✅	quality_gates_rss_dsd_medium	memory_usage	10/10	60.44MiB ≤ 75MiB	(metrics) (profiles) (logs)
✅	quality_gates_rss_dsd_ultraheavy	memory_usage	10/10	177.11MiB ≤ 200MiB	(metrics) (profiles) (logs)
✅	quality_gates_rss_idle	memory_usage	10/10	27.82MiB ≤ 40MiB	(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".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 001ef79959

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Limit 403 retries to endpoints with refreshable keys

When additional_endpoints are configured and secrets are enabled, this installs the 403 retry override for every endpoint I/O loop. However, EndpointConfiguration::build_resolved_endpoints only attaches the live GenericConfiguration to the primary endpoint; additional endpoints are built as static ResolvedEndpoints, so their API keys cannot be refreshed by ResolvedEndpoint::api_key(). A permanent 403 from a secondary endpoint will therefore be re-enqueued and retried until queue limits are hit instead of being treated as a non-retriable auth error, consuming retry queue/disk capacity and potentially delaying or dropping other transactions.

Useful? React with 👍 / 👎.

[atanzu]

Correct, I will follow-up on that as a part of #1540

[webern]

There might be a way to simplify if my comments are correct...

[webern]

Just a thought. The predicate idea could be more open to extensibility if it is just a functor that takes the response and returns a bool.

pub type HttpRetryPredicate<B> = Arc<dyn Fn(&http::Response<B>) -> bool + Send + Sync>;

fn default_should_retry<B>(resp: &http::Response<B>) -> bool {
    match resp.status() {
        StatusCode::BAD_REQUEST
        | StatusCode::UNAUTHORIZED
        | StatusCode::FORBIDDEN
        | StatusCode::PAYLOAD_TOO_LARGE => false,
        _ => resp.status().is_client_error() || resp.status().is_server_error(),
    }
}

pub struct StandardHttpClassifier<B> {
    predicates: Vec<HttpRetryPredicate<B>>,
}

impl<B: 'static> StandardHttpClassifier<B> {
    /// Starts with the default status-code predicate installed.
    pub fn new() -> Self {
        Self { predicates: vec![Arc::new(default_should_retry)] }
    }

    /// Starts with no predicates; all responses are non-retriable until one is added.
    pub fn empty() -> Self {
        Self { predicates: Vec::new() }
    }

    /// Adds a predicate. OR semantics: retry if any predicate returns `true`.
    pub fn with_predicate(mut self, predicate: HttpRetryPredicate<B>) -> Self {
        self.predicates.push(predicate);
        self
    }
}

impl<B, Error> RetryClassifier<http::Response<B>, Error> for StandardHttpClassifier<B> {
    fn should_retry(&self, response: &Result<http::Response<B>, Error>) -> bool {
        match response {
            Ok(resp) => self.predicates.iter().any(|p| p(resp)),
            Err(_) => true,
        }
    }
}

The secrets predicate in retry.rs becomes:

let secrets_gate: HttpRetryPredicate<_> = Arc::new(move |resp| {
    resp.status() == StatusCode::FORBIDDEN && is_secrets_in_use(&config)
});
let classifier = StandardHttpClassifier::new().with_predicate(secrets_gate);

This removes the HashMap<StatusCode, predicate> keying and the set_/remove_status_code_predicate
surface - predicates own their own logic and have the full response available.

[webern]

Not a request to fix in this PR, but this destructure is where the too_many_arguments warnings (suppressed a few lines down) seem to originate. All these fields are owned by self, destructured here, and then passed individually through run_io_loop and run_endpoint_io_loop.

Perhaps self could be left in-tact here and then passed through as a single argument. Not sure.

Suggested change

	let Self {
	// TODO: do not destructure self as a way to fix the #[allow(clippy::too_many_arguments)] annotations
	let Self {

[webern]

The four keys this reads (secret_backend_command, secret_backend_type, secret_refresh_interval, secret_refresh_on_api_key_failure_interval) describe whether secrets management is configured, not whether a refresh is currently in progress. I believe these are set at startup and not expected to change at runtime (right?), so the dynamic evaluation on every 403 may not be necessary and they could be read just once.

If that's right, a bool computed at construction time in from_config would behave identically:

let retry_403_on_secrets = live_config.as_ref().map_or(false, is_secrets_in_use);

That bool could then be captured directly in the predicate closure, and Option<GenericConfiguration> wouldn't need to be threaded through TransactionForwarder, run_io_loop, and run_endpoint_io_loop.

[webern]

If the is_secrets_in_use question (retry.rs:219) resolves to a bool, this parameter goes away and the threading through run_io_loop / run_endpoint_io_loop goes with it.