Skip to content
New here? Rent a GPU by the second
Kracht
Sign inSign up

Studio Observability

See what a run
actually did.

Spans nested as traces, where the time went, and every span flat and sortable when the tree is not what you need.

Open the console
Studio
Traces
one per run
Spans
nested as they ran
Share
of the trace
Flat
sortable span list

The screens are real. The runs in them are not yours

example data

Studio saves a workflow but does not execute one yet, so there is nothing of yours to trace. Every console page here carries the same line it always has: this is the shape, not your account.

Traces

A run, nested the way it happened

A trace is the run; spans are what it did. Steps that ran at once sit side by side rather than one after another, because the shape of the tree is the first thing that tells you where to look.

  • Read a duration as first span to last, not the sum of them
  • See parallel steps sitting side by side, because they ran that way
  • Open any span for what it took and what it held
Open a trace
run4.2 sfetch0.3 sembed1.1 sanswer2.8 s

Waterfall

Where the time went

Share of the trace, per span, so the expensive step is the wide one rather than something you work out from timestamps. The answer to 'why was that slow' is usually one bar.

  • Compare share of the trace rather than a bare duration
  • Sort by what took longest instead of by what ran first
  • Read the percentile: one run in twenty was slower than this
Every step this has run
A waterfall of one trace lasting 4.2 seconds. Fetch takes 0.3 seconds, 7% of the trace. Embed takes 1.1 seconds, 26%. Answer takes 2.8 seconds, 67%, and is the step worth looking at.

Spans

Flat and sortable, when a tree is not what you need

A tree is right for one run and wrong for a hundred. The same spans render as a flat list you can sort and search by step, kind or trace, which is how you find the step that is slow every time rather than the run that was slow once.

  • Search by step, kind or trace
  • Trust the same rows as the tree, so the two cannot disagree
  • Expect ‘nothing measured’ rather than a zero, because a zero is a measurement
The span list
The same spans as a flat list, sorted by what took longest: answer 2.8 seconds at the 95th percentile, embed 1.1 seconds, then two fetches at 0.3 seconds each.

Resources

Where to go next

Documentation

How it works, endpoint by endpoint

Observability

Traces, waterfall and the span list

Evaluation

Score a change before you ship it

Frequently asked questions

Not yet. Studio saves a workflow but does not execute one, so every trace in the console today is a worked example. The screens are real; the data in them is not yours.

Want to see where a run spends its time?

Open the console
Two sides, one marketplace

Spin up a GPU, or earn from your own

Rent on-demand compute by the second, or list idle hardware and let it pay for itself. Same marketplace, both directions.

Start renting
List your hardware

No card required

85%
revenue to providers
Per-second
billing, live
EUR
metered to the microeuro