Benchmarks

The Rust, Python and JavaScript bindings are timed on the same three cases, against businesstimedelta 1.0.1 and against the loop you would write with each language's standard library. The Python suite also times workclock.compat. The cases and the commands come from benchmarks/README.md. The numbers in benchmarks/RESULTS.md come from the Benchmarks workflow on a GitHub-hosted Ubuntu runner, and RESULTS.md links the run, so anyone can open it.

Runner CPUs vary from run to run. On 2026-10-26, three runs on AMD EPYC 7763 and Intel Xeon 8573C runners moved the times by up to about 20%, while the speed-ups over businesstimedelta, measured in the same run, stayed within about 6%. Compare the "vs businesstimedelta" and "vs stdlib" columns rather than raw times.

CI also counts instructions under Valgrind for Rust and Python, and fails a pull request whose single or year case executes more than 3% more instructions. Instruction counts has the details.

CasesCopy Anchor Link

The Rust, Python and JavaScript suites measure the same three cases, all with one schedule: Europe/London, Monday to Friday 09:00–17:00, a 12:00–13:00 break, and the bank holidays of England in 2026 and 2027. The 2027 ones are there because businesstimedelta 1.0.1 raises TypeError for instants after its last holiday (upstream issue 8).

Building the schedule and the pairs is not timed.

Standard-library baselinesCopy Anchor Link

Neither Python nor JavaScript can compute working time out of the box, so each suite also times stdlib: the loop a developer would write for this one schedule with the standard library alone. It walks the local dates from the first instant to the second, skips weekends and holidays, and adds the overlap with 09:00–12:00 and 13:00–17:00 London time.

Tests check both against workclock on the single and year cases, both daylight saving changes, Easter, a reversed pair and 10,000 batch pairs, so every speed-up compares code that gives the same answers. RESULTS.md divides each language's stdlib median by every row's median in its "vs stdlib" column. Rust's standard library has no time zones, so Rust has no baseline.

RunningCopy Anchor Link

cargo bench -p workclock --bench workclock
uv run pytest benchmarks/python
cd js && npm run build:node && npm run bench

The JavaScript suite needs Node.js 26, for Temporal.

The Python suite also measures workclock's compat API, between_many on lists and on NumPy arrays, and businesstimedelta 1.0.1. It runs the single case on compat and businesstimedelta a second time with a python-holidays object for England in place of the list; compat fills it in for 1970 to 2100. One round of the batch case takes businesstimedelta about half an hour; add -k "not (test_batch and businesstimedelta)" to skip it.

The JavaScript suite runs the batch case three ways: between in a loop, and betweenMany on arrays of Dates and on Float64Arrays.

ResultsCopy Anchor Link

Cases and commands are in README.md.

singleCopy Anchor Link

yearCopy Anchor Link

batchCopy Anchor Link