Amplification and goodput
How many calls one user request causes, and how much of a callee's running time the caller actually used.
// zombie-work x-ray for OpenTelemetry traces
Is your backend doing work nobody is waiting for?
Drop in OpenTelemetry traces. See how much work was wasted on retries and on callers that had already given up, and which fix is missing.
No traces handy?
or run locally
curl --data-binary @traces.json <this site>/api/analyze
- No install
- No account
- Analyzed in memory, never stored or logged
// how it works
It counts the work nobody used.
xray reads the spans of each trace, matches every call to the work it caused downstream, and checks who was still waiting for the answer.
Zombie work
Work that kept running after its caller gave up, and for how long, all the way down the chain.
The missing guard
Every finding names its fix: a retry budget, backoff, or sending the caller's remaining time downstream.
The bundled retry-storm sample, drawn to scale
// results
Does the fix work? A load test.
The same system under rising load, three ways: no deadline sent, cancel at the deadline, and cancel plus refuse jobs that cannot finish in time.
Chart
// private by design
Your traces are looked at, not kept.
Parsed, analyzed, dropped
Nothing you upload is written anywhere. There is no account to create and nothing to sign in to.
Errors never repeat your data
Messages name the file, the line and the kind of problem, never what was inside the file.
Hard limits
The server enforces .
// bring your own traces
Export them with the Collector.
Every service needs OpenTelemetry tracing with context propagation, so a call from one service and the work it causes in the next end up in the same trace. Then write the traces to a file with the OpenTelemetry Collector's file exporter.
Supported: OTLP/JSON (one export request per line, or one document), the OpenTelemetry Python SDK's span JSON, Zipkin v2 JSON (an array of spans), and Jaeger Trace JSON (save the page the Jaeger UI's Trace JSON option opens, or fetch it from the query API at /api/traces). Optional: if a service queues requests before running them, set the span attribute app.queue_wait_ms on its server span, so waiting is not counted as work.
exporters:
file:
path: /data/otlp-traces.json
service:
pipelines:
traces:
receivers: [otlp]
exporters: [file]// saved runs