Performance

What a record costs your thread

Published measurements from the rlg repository, with the machine they came from and what they do not show.

What is measured?

The competitive_bench suite times the cost to the calling thread of emitting one record. The three contenders do different work:

  • rlg fire() checks the level and pushes the record into the ring buffer. Formatting and I/O happen later on the background thread and are not in the timed path.
  • tracing::info! goes through a tracing_subscriber::fmt subscriber that formats the event on the calling thread and writes it to a discarding writer.
  • log::info! goes to a logger that does nothing: the cost of the facade alone.

The numbers

From the repository's docs/BENCHMARKS.md: one run on a GitHub-hosted ubuntu-latest runner, stable Rust, release profile, on the 0.0.13 release branch (run 36802323103). Typical time per iteration, with Criterion's 95% confidence interval.

Scenario rlg fire() tracing::info! log::info! (no-op)
One record 598 ns (582 to 615) 591 ns (589 to 593) 2.7 ns
One record, 3 attributes 947 ns (923 to 969) 1,117 ns (1,115 to 1,120) 3.0 ns
Burst of 10,000 records 6.86 ms (6.57 to 7.20) 6.97 ms (6.95 to 7.00) 0.03 ms
Latency distribution 649 ns (639 to 659) 587 ns (585 to 590) 2.7 ns

How should I read them?

On the calling thread, rlg costs about what tracing costs, and less with attributes, while tracing has already formatted the event and rlg has not. What rlg buys is that the caller never waits on the sink's I/O, which this suite's discarding writer does not exercise. Runner speed varies between runs: compare within one run, not across runs.

What about many threads?

Version 0.0.14 cut contention between producers. On the maintainer's development machine, with eight threads logging at once, fire() went from 3.9 to 1.8 µs per call, and from 3.0 to 0.95 µs with the fast-queue feature. These come from the CHANGELOG, not from a published run.

How do I reproduce them?

From a clone of the repository:

cargo bench -p rlg --bench competitive_bench

Each release tag also runs every workspace benchmark on a GitHub-hosted runner and keeps the Criterion report as a workflow artifact for 90 days.