Episodes
Watch the series on YouTube- EP 01
Claude Code
Tokens, cost, tool decisions and every LLM request, exported with one settings file.
tracesmetricslogs - EP 02
Docker Buildx
Span names are your Dockerfile steps, and one span name tells you exactly where your cache breaks.
tracesmetrics - EP 03
Keycloak
One login is 28 spans, and the database is 0.7% of it. The trace shows where the other 99% goes.
tracesmetricslogs - EP 04
OpenTofu
26 seconds of resource work inside a 10-second apply, and the trace names the resources that set the floor.
traces - EP 05
eve
Every model call reads ~7,000 tokens and writes ~200, a 30:1 ratio on every step of every turn.
traces - EP 06
Ansible
Three hosts finish every task in under a second; the fourth takes the whole 13-second play, and the trace names it.
traces - EP 07
OpenAI Codex CLI
The same job asked two ways: 104.2k tokens one time, 87.1k the other, and the difference is naming the file.
logs - EP 08
Apollo Router
The product page takes 261 ms, and 256 ms of it is one subgraph the router could not call until the catalog had answered.
tracesmetrics - EP 09
OpenRouter
17 of 75 first attempts were refused with a 429. Every one was retried on another provider, and the agent never said so.
traces - EP 10
Jenkins
The pipeline with the least work is the slowest build: 26 of its 35 seconds are one span, waiting for an executor.
tracesmetricslogs - EP 11
Kubernetes API server
A deploy succeeds after 14 seconds with no error, and the trace puts 13.8 of them on one Kyverno webhook call.
traces - EP 13
Claude Platform
The answer looked fine. The trace showed the agent's first command had failed.
traces - EP 14
n8n
292 of 300 HTTP calls in one workflow run were waiting for a database connection, up to 4.7 s.
traces - NEXT EPISODE
Your tool?
Know a tool with observability on board that we haven't covered? Tell us which one and where its OpenTelemetry docs live.
Request an episode