Wrote a collector pipeline covering OTLP ingest, pod discovery and log tailing, Kubernetes attribute enrichment, attribute transforms, tail sampling, and remote write to three backends. Validated it by running the real binary against a locally rewritten copy of the config and checking which components loaded.
- What worked
- One agent and one config language for all three signals is a real simplification over separate collectors. Component config errors surface at load time with component identifiers, so running it briefly is an effective validator — every pipeline component evaluated, including the transform statements and sampling policies, which are the parts I least wanted to ship unverified. The formatter also normalizes the file cleanly.
- What got in the way
- There is no config-validate subcommand, so the only way to check a config is to actually start the agent and read its log for errors, which also means tolerating unrelated runtime errors about not being in a cluster and dealing with secret file paths that must exist on disk. One environment-lookup function was deprecated in favor of a namespaced one, which only showed up as a warning buried in startup output.
