Listen on both TCP and named pipes when a pipe name is configured

[Lewis-E]

Tracers running on Azure Functions Premium (multifunction) plans need named-pipe transport so each function gets isolated trace/dogstatsd channels. Older tracers without pipe support still need to reach the agent over TCP. Today the binary picks one transport: configuring a pipe name disables TCP, which would strand any old-tracer + new-compat combination.

This PR makes the binary listen on both transports concurrently when DD_APM_WINDOWS_PIPE_NAME is set: TCP stays on DD_APM_RECEIVER_PORT (default 8126), named pipe is added alongside. When no pipe name is configured, behavior is unchanged (TCP-only, identical to before).

What changed

• mini_agent.rs: spawn TCP and (optionally) named-pipe accept loops, supervised by a single serve() function. Any task dying unexpectedly is fatal — tracers pick one transport at startup and don't switch, so silently degrading a transport would strand tracers using it.
• mini_agent.rs and main.rs: start_mini_agent now takes shutdown_rx: tokio::sync::watch::Receiver<bool> (was oneshot::Receiver<()>). Each accept loop subscribes via .clone(), so the shutdown signal fans out to both transports. Production caller in main.rs also updated.
• mini_agent.rs: graceful shutdown is preserved across the dual-transport architecture — on shutdown_rx, both accept loops drain their in-flight HTTP handlers before the supervisor signals the stats flusher (flusher_shutdown_tx) to do its final emit. The trace-stats invariant ("no AddChunk/ClientStatsPayload messages lost on shutdown") stays intact.
• config.rs: drop the override that forced dd_apm_receiver_port = 0 whenever a pipe was configured. That override was load-bearing only when pipe-mode meant "no TCP."
• Pipe code remains fully cfg-gated. On non-Windows / no windows-pipes builds (Lambda), zero pipe symbols make it into the binary.

Test plan

• Existing TCP-only and pipe-only integration tests pass (integration_test.rs)
• Unit test for default-port behavior when a pipe is configured (test_apm_windows_pipe_name)
• New test_mini_agent_dual_transport_with_real_flushers integration test (Windows + windows-pipes): boots one agent with both transports configured, sends a service-tagged trace through each, fires shutdown_tx, and asserts both traces and the final stats flush reach the mock backend. Exercises the watch fan-out and graceful drain on both accept loops in one go.
• Cross-target compile-check: macOS host, x86_64-pc-windows-gnu --features windows-pipes, and the default Lambda build
• E2e on a multifunction Premium plan

Related PRs and context

Part of the multifunction-named-pipes workstream — epic SVLS-8429, this PR's ticket SVLS-8243.

• #58 — original Windows named pipes for the trace agent (the feature this PR makes backwards-compatible)
• #51 — trace stats computation (this PR rebased on top; integrated stats_concentrator_service_handle and the graceful drain into the new serve() supervisor)
• datadog-serverless-compat-dotnet#20 — compat layer wiring the per-function pipe coordination
• dd-trace-dotnet#8164 — .NET tracer side that consumes the pipe contract

This PR's binary release must precede datadog-serverless-compat-dotnet v1.5; v1.4 stays TCP-only and is unaffected.

Notes

• A symmetric port-0 override exists for dogstatsd (config.rs, dd_dogstatsd_port). Out of scope here; flagged for a follow-up.

[Copilot]

Pull request overview

This PR updates the Windows named-pipe transport behavior in the trace mini-agent so that, when DD_APM_WINDOWS_PIPE_NAME is set, the agent listens on both TCP (DD_APM_RECEIVER_PORT, default 8126) and the configured named pipe concurrently—preserving connectivity for legacy tracers that only support TCP.

Changes:

• Refactors MiniAgent startup to spawn TCP and (optionally) named-pipe accept loops and supervise them via a centralized serve() function.
• Removes the config behavior that forced dd_apm_receiver_port = 0 when a Windows pipe name is configured.
• Updates the related config unit test to assert the default TCP port is preserved when a pipe name is set.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
crates/datadog-trace-agent/src/mini_agent.rs	Starts TCP unconditionally, optionally starts named-pipe accept loop, and centralizes task supervision.
crates/datadog-trace-agent/src/config.rs	Keeps TCP port enabled even when DD_APM_WINDOWS_PIPE_NAME is configured; updates test expectation accordingly.

Comments suppressed due to low confidence (1)

crates/datadog-trace-agent/src/mini_agent.rs:93

• Same as above: this mut is unused and suppressed via #[allow(unused_mut)]. Prefer removing the mut binding (and the allow) to avoid masking future cases where mutability is accidentally unused.

        #[allow(unused_mut)]
        let mut stats_flusher_handle = tokio::spawn(async move {
            stats_flusher
                .start_stats_flusher(stats_config, stats_rx)
                .await;
        });

[duncanpharvey]

What versions of the .NET, Java, and Node.js tracers was named pipe support added in? Can we validate that anyone is using one of these older version to justify the added complexity in listening over TCP and Named Pipes?

If we do move forward with adding this backwards compatibility for older tracers, do we have a plan for deprecation?

[Lewis-E]

@duncanpharvey We need this for the .net tracer specifically, where this support does not yet exist: DataDog/dd-trace-dotnet#8164 -- see the table in the PR description.

I actually think this complexity is worth keeping-- it's under 100 lines, I wouldn't prioritize deprecating this against any of the other work we have. At best, I would throw it at the bottom of the backlog. Do you expect performance issues or similar that would change the calculus there?

[duncanpharvey]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f89f181bbf

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid aborting the trace flusher on graceful shutdown

When shutdown is requested after a trace handler has returned but before the next trace flush interval, this aborts trace_flusher_handle immediately after the accept loops drain. start_trace_flusher has no shutdown/final-flush path (it only flushes on its timer), so any SendData just queued by the drained handlers can remain in the receiver/aggregator and never be sent; the new real-flusher tests avoid this by waiting for verify_trace_request before shutdown, but production callers can lose traces during graceful shutdown.

Useful? React with 👍 / 👎.