Measure the cost

The bench example measures what a tool and a reader pay per call, with the files in the job folder it picks for itself:

cargo run --release --example bench

One run on an AMD Ryzen 9 9950X, Linux, Rust 1.97.1, on a busy machine (load average about 35), printed:

Job::progress                          235.2 ns
Handle::frame                           26.0 ns
Writer::event, one line                237.3 ns
Line::parse, one line                  777.4 ns
read_dir, per record                  3688.9 ns
job record on disk                       453 B
Job::progress, PITO_JOBS=off           138.3 ns

Numbers move from run to run with the machine’s load. The chart and table below are the run of 2026-10-14 on the same machine at a load of about 70, which the overview quotes:

Call	Cost	How
Handle::frame	26.5 ns	5,000,000 frames into a live heartbeat
Job::progress	292.2 ns	5,000,000 calls on a live job, its file rewritten once a second
Job::progress, PITO_JOBS=off	176.0 ns	the same with every write off
Writer::event	231.7 ns	1,000,000 progress lines into io::sink
Line::parse	670.7 ns	1,000,000 parses of one progress line
read_dir	4.2 µs a record	2,000 reads of a folder of 32 job records
A job record on disk	454 bytes	the benchmark’s own record

At 60 fps a frame has 16.7 ms, so Handle::frame costs 0.00016% of it, and a progress call per frame is under 0.002%.