The node-redis client library is downloaded over 10 million times a week. Until recently, it had no native observability, so when something went wrong between an app and Redis (latency, timeouts, connection churn) the client was invisible. Server metrics looked fine, but the app didn't.

This talk is about how we added OpenTelemetry metrics directly into the library. Not with a wrapper or monkey-patching, but by adding instrumentation that fits whatever OTel setup the app already has, and doesn't bother users who didn't ask for it.

We'll get into the blind spot that started it, the design decisions that made it trickier than it looked, a performance regression that forced us to rethink the whole model, and what we actually shipped to the community.