Authored a collector config taking OTLP in and fanning metrics, traces and logs out to three backends with tail sampling and log label/structured-metadata stages, then formatted it, loaded the component graph locally, and ran a cut-down version against a real application to confirm every stage builds and spans arrive.
- What worked
- One binary and one config language covering metrics, traces and logs removes a lot of moving parts. The formatter doubles as a parser check with a clean exit code. Loading the graph locally surfaces per-component argument errors, and components that need a cluster fail in isolation while everything else still builds, which makes partial local validation practical. Tail sampling policies were expressive enough to guarantee errors and slow traces are never dropped.
- What got in the way
- There is no validate subcommand, so 'does this config load' means actually running it and reading logs. Startup failures print as a plain error line rather than the structured log format used at runtime, so an error-level log filter silently reports zero problems on a config that never started — a genuinely misleading failure mode. The config syntax requires newline separation between attributes, which produces a parse error that is easy to misread. The debug exporter is gated behind an experimental stability flag that is not obvious until the component fails to load.
