DEV Community

Alex Georghiou
Alex Georghiou

Posted on

Correlation IDs Are Not Observability

Correlation IDs Are Not Observability

Adding a correlation ID to logs is a useful first step. It is not an observability strategy.

When latency spikes, you need to know where the time went. Request logs tell you when an operation started and finished. Traces reveal which downstream call or query consumed the budget. Metrics tell you whether the problem affects a handful of requests or the entire fleet.

I prefer to define a few service-level questions before choosing tooling: What percentage of requests succeed? What does tail latency look like? Which dependency causes retries? Are queue consumers falling behind? How long do failed jobs sit without resolution?

High-cardinality labels can destroy a metrics system. User IDs, arbitrary URLs, and raw exception messages belong in carefully controlled traces or logs, not as metric dimensions. Logging sensitive payloads because "it helps debugging" creates a different class of incident.

A good trace should follow work across a synchronous request and, where possible, into asynchronous processing. A good dashboard should help someone decide whether to roll back, shed load, or investigate a particular dependency.

Observability is the ability to answer unfamiliar production questions with evidence. More log lines are not automatically more evidence.

Top comments (0)