MapColonies / shigola / bf16266bf

57%

master: 44%

Build Type

push

github

Committed by web-flow

Commit Message

feat(tracing): add OpenTelemetry tracing, exporting OTLP to Tempo (MAPCO-11497) (#21)

* build(deps): vendor the OTLP trace exporters and the OTel trace SDK

The OpenTelemetry API, SDK and otelhttp were already in the module graph as
indirect dependencies, pulled in by the GCS cache's transitive closure, and the
trace SDK and OTLP exporters were not there at all. Promote the six that
shigola is about to import directly and add
go.opentelemetry.io/otel/exporters/otlp/otlptrace, which brings the OTLP proto
package, cenkalti/backoff and grpc-gateway with it.

gRPC and protobuf are already vendored, so the incremental link cost is the
exporter and its proto definitions rather than a new transport.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* feat(tracing): add a tracing seam with a no-op default

Its own package, not part of observability. That one is a metrics abstraction
which Mimir consumes unchanged, so OTEL metrics are never wanted: nothing here
registers a prometheus collector, and the HTTP decorator pins otelhttp's meter
provider to a no-op explicitly rather than relying on nobody ever installing a
global one.

Off is a distinct implementation rather than a flag read inside each method.
Null returns its argument from every Instrumented* method, so a process with
tracing disabled carries no decorator at all — not one that starts a discarded
span per cache read. That makes "no measurable overhead when disabled" a
property of the wiring instead of a claim about how cheap a dropped span is.

The cache decorator is transparent by contract: a read failure is still a miss,
never an error, and the returned tuple is the wrapped cache's. It records the
error on the span either way but sets the span's *status* to error only when
the failure is the cache's own, which is the same line the prometheus
observer's countReadError draws — a trace saying a tier failed while
shigola_cache_tier_errors_total says it did not would be worse than e... (continued)

Coverage Stats

417 of 599 new or added lines in 21 files covered. (69.62%)

2 existing lines in 2 files now uncovered.

5714 of 10326 relevant lines covered (55.34%)

124.11 hits per line