[RaphaelManke]

Summary

• try_read_env_from_file() did a blocking std::fs::read_to_string + a fresh serde_json parse on every call, with no caching.
• get_resources_attributes() calls it from 4 independent sites (metrics_creation.rs, span_creation.rs, span_mutations.rs, exporter.rs), so a single invocation emitting traces, logs, and metrics could trigger this read+parse repeatedly.
• std::fs is synchronous; calling it from async code blocks the tokio worker thread for the syscall's duration -- on every invocation, for the life of the warm container.
• The file is written once by opt/shared.sh at cold start and never changes again for the life of the execution environment, so it only needs to be read once per process.

Fix

Cached behind Mutex<Option<Arc<Option<serde_json::Value>>>>, matching the existing shared-state pattern already used in state/global.rs. Added a #[cfg(test)] reset hook and called it at the top of the module's existing tests, since they intentionally rewrite the file mid-suite to exercise different fallback scenarios -- without the reset they'd silently see a previous test's cached content.

Measured

This crate has no [lib] target for a real criterion/divan bench to attach to, so measured in isolation by replicating the exact read+parse call in a throwaway benchmark: 9.4us/call uncached vs. 35ns/call cached -- a 265x difference.

Test plan

• cargo build --release
• cargo test --release (325/325 passing, including all 7 otlp::resources::tests::*)
• Ran the full suite twice to confirm no test-order flakiness from the cache

[dash0-dev]

Darkplane Automerge is enabled for this repository (mode: Dry run).

Trigger an evaluation for this pull request when it is ready:

Evaluations on this repository start manually. To have Darkplane Automerge evaluate each pull request on its own, set Trigger to Automatic in the repository's configuration.

[dash0-dev]

Darkplane Automerge would approve this pull request. It scored 77.79 of 100, clearing the threshold of 75. Dry-run mode is on, so nothing was approved or merged.

• 🟡 Regression coverage scored 38.
  The new ENV_VARS_FILE_CACHE in src/otlp/resources.rs has no test that pins caching: nothing checks that a second read skips the file, or what happens when the file is absent or unreadable on first call (the None is cached for the process lifetime). The existing src/otlp/resources.rs tests (~lines 158–410, run by cargo test, CI green) still cover value lookup, but the PR adds a reset call to each so they bypass the cache. A regression such as a stale or permanently missing resource attribute would pass CI.

The other 3 criteria

• 🟢 Customer-function safety scored 88.
  The change in src/otlp/resources.rs memoizes the /tmp/dash0_env_vars read behind a parking_lot mutex. It adds no unwrap or panic, no new await, and no I/O on the proxied /invocation/next, /response or /error path (verified by reading). The one residual risk is a read that fails before write_env_vars runs in the wrapper (e.g. a read during INIT): that None would be cached for the process lifetime and drop wrapper-added resource attributes and the service.name fallback. That would only degrade telemetry, and the callers, exporter.rs and metrics_creation.rs among them, appear to run at invocation time, after the wrapper has written the file. I read the callers and did not trace every call path.
• 🟢 Telemetry & config contract scored 82.
  The change in src/otlp/resources.rs only caches the /tmp/dash0_env_vars read, so precedence for OTEL_RESOURCE_ATTRIBUTES and OTEL_SERVICE_NAME and the emitted attribute keys stay the same, and no config or masking variable changes. Residual risk: a failed or missing read is cached as None for the whole process, so if anything built resource attributes before write_env_vars in opt/shared.sh wrote the file, faas.* and service.name would fall back to unknown_service. I couldn't confirm whether any such early call exists, so I scored it down a little.
• 🟢 Cold start & footprint scored 97.
  src/otlp/resources.rs now caches /tmp/dash0_env_vars after the first read, which removes repeated file reads and JSON parses. The PR reports 9.4us per call uncached versus 35ns cached. once_cell and parking_lot are already in Cargo.toml, so there are no new dependencies. The one-time read still happens on first use, and the cached JSON stays in memory for the process lifetime, which is negligible.

_{An AI evaluator produced these scores from the diff, the repository history, and the pull request text. The pull request text is author-controlled, so treat the scores as signals rather than proof. Reviewed for commit 6882ae3.}