First, the key terms.

Metric
An aggregated measurement over time.
Trace
A linked view of operations contributing to a request or job.

The idea, explained

A latency metric can reveal a slowdown without explaining which dependency caused it. A distributed trace connects related spans so an engineer can inspect the path of a request. Logs add context but can leak personal data or tokens if collected carelessly. Sampling, cardinality and retention affect both cost and diagnostic value. Design telemetry around concrete questions such as whether users can complete an action, not just whether servers are running.

STEP 1

Start with a diagnostic question

For missing newsletters, ask whether eligible readers were selected, jobs processed and messages accepted. CPU utilization cannot answer those questions. Measure state transitions and failure counts at meaningful boundaries.

STEP 2

Connect events safely

A request or job identifier can connect logs and traces without storing every email body or token. Control access and retention. High-cardinality labels and indiscriminate payload capture can add cost and privacy exposure.

Follow a worked example.

Email jobs appear healthy by CPU usage, but users report missing newsletters. Queue age, delivery-state metrics and a trace through database selection and provider response reveal that a downstream provider is rejecting requests.

Walk it through, one step at a time

  1. A reader reports no digest. The selection count is correct, but provider-acceptance failures rise.
  2. A trace shows successful database access followed by provider rejection.
  3. Investigate the returned failure category and configuration. Do not resend to everyone before understanding which deliveries actually failed.

Where people get stuck

Collecting every payload is not necessary for good observability and can create privacy and security problems.

Now try it yourself.

Pause before opening each answer. Explain your reasoning in your own words.

1. Check the core idea

Why is a server uptime chart insufficient to establish newsletter reliability?

Reveal the answer & reasoning

The server can be running while selection, queuing or delivery fails. Measure the user-visible outcome as well as infrastructure health.

2. Apply what you learned

Why can detailed logs still be poor observability?

Reveal the answer & reasoning

They may be unstructured, missing the key boundary or too noisy to connect. More collected data is not equivalent to better answers, and can create avoidable risk.