A year of LLM traffic

Model family

Loading the trace…

Starting DuckDB in your browser…

Technical notes

Data source

The Chutes production LLM trace published with arXiv:2608.13573 (github.com/HarvardMadSys/chutes_workload). This page reads traffic.parquet, a rollup of that trace to rows: one per trace day × trace hour × model family, from .

Aggregation

Requests and tokens are sums. Tokens = input + output tokens.
Averages are ratios of sums: output tokens per request = output tokens ÷ requests with an output count; time to first token = summed seconds ÷ requests with a measured first token.
Month views show requests per day of the month, so the partial months () compare fairly with full ones; they are labelled partial wherever they appear.
Colors: the five largest named families by requests have their own hue; every other family, including the catch-all labels, is drawn as in muted grey. The family menu lists all .
Shares are of the requests (or tokens) in view: with a family filter they are shares among the chosen families.

What the filters reach

The family menu filters everything. A click on a family mark filters the numbers, the hour views and the time strip, and highlights the family in the chart you clicked.
Dates (brush or Play): the KPIs, the share views (act I) and the hour views (act III) follow the window. The month views (acts II and IV) keep the whole year as context and dim the months outside the window.
Scene titles describe all the data; when a filter is on, the chart and the numbers follow it, and a note under the title says so.

Limitations

Dates are reconstructed from the paper's stated span: trace day 0 is taken as 11 April 2025.
Hours are on the trace clock; its timezone is not stated.
Time to first token exists only for streamed calls: of requests have a measured first token.
Token counts are missing for of requests; token totals cover the rest, and output tokens per request is divided only by requests with a count.
Quietest day in the trace: . Slowest day by time to first token: .
No cost, error or user data: the trace can't say why families rose or fell.

Built with

Mosaic 0.31.0 and DuckDB-WASM (every number is a query in your browser), D3 7.9.0 polygon morphs, d3-3d 1.0.0 for the 3D boxes, Tabler icons 3.35.0.