Retries change load.
One browser request can become several downstream calls, even when every call returns nothing useful.
Lab tool / simulated request path
Request X-Ray
One request, seven layers, and a handful of switches. Change retries, deadlines, cache behaviour, rate limits, and service errors, then watch the path explain itself.
01 / Simulator
Break the request, then read the route.
The main instrument stays in one working surface: scenario tabs first, live trace in the middle, switches underneath, and the run summary close enough to compare without hunting.
Scenarios
Latency budget
0ms of 500ms
- Status
- idle
- Total latency
- 0ms
- Downstream calls
- 0
- Attempts
- 0
- Terminated at
- -
- Pattern
- -
edge
Seeded extra latency across the path.
router
Refuse before downstream work.
api
Deadline is per attempt; backoff is visible on the API hop.
service
Return 503 before reading anything.
cache
Fast answers can still be wrong answers.
02 / What this proves
Reliability failures have shape.
A failed request is not a single red light. It is a path through owned layers: where the deadline is spent, where retries multiply load, where cache hides truth, and where a refusal prevents damage.
Timeouts have owners.
A per-attempt deadline can keep individual attempts bounded while the whole request still drifts over budget.
Fast is not always healthy.
A stale cache hit looks cheap on latency and expensive on truth. That distinction matters.
Layer detailbrowser
Configured behaviour
Last run outcome
Hop tableAccessible trace
The same trace without the animation. This is the accessible equivalent of the diagram and the fastest way to audit a run.
| # | Layer | Attempt | Outcome | Latency | Note |
|---|
ComparePin a run
Pin the current configuration, change the switches, and read the two runs side by side.
Nothing pinned yet.
Model notesSimulation boundary
This page does not call a real service. The latency model is deliberate, repeatable, and documented so a permalink reproduces the same experiment instead of rolling new data.
Read the latency model