Skip to content
Mustry Solutionsiglint

Performance and caching

iglint is designed to be fast without skipping work: each file is read and parsed once, files are processed in parallel, no subprocesses are started (except pylint for SCR900, which is off by default), and results are cached between runs.

Benchmark results

Section titled “Benchmark results”

The numbers below are from bench/README.md, measured on 2026-10-24 on a development machine with 16 CPUs (8 cores, 2 threads each) and 23 GB of RAM, release build. The corpora are synthetic Ignition repositories made by the deterministic generator in bench/gen, with distributions calibrated on the aggregate statistics of a production project (no names or content were used). Times are minimum wall times; x is ign-lint's best working configuration (jobs=1) divided by iglint's time.

small (50 views) medium (500) large (5,000) xl (20,000) ¹
ign-lint jobs=1 (best working) 1.64 s 16.4 s 177 s 713 s
iglint ign-lint, with pylint (equal work) 236 ms (6.9x) 587 ms (28x) 4.25 s (42x) 16.5 s (43x)
ign-lint without pylint 226 ms 1.60 s 15.4 s 65.6 s
iglint ign-lint, without pylint (equal work) 21 ms (11x) 103 ms (16x) 614 ms (25x) 1.24 s (53x)
iglint check, default rules, cold 29 ms 164 ms 878 ms 1.89 s
iglint check, warm cache 23 ms 126 ms 658 ms 1.51 s
peak memory: ign-lint jobs=1 / iglint check 66 / 174 MB 193 / 304 MB 1,427 / 1,115 MB 5,560 / 3,195 MB

¹ The xl column is from an earlier run on an idle machine; the others were re-measured while the machine also ran other workloads.

How to read this honestly:

  • Equal work means the same files, the same rules and, with pylint, the same pylint version over the same scripts: iglint ign-lint against ign-lint. iglint check does more work than ign-lint (it parses every resource of the repository and runs 88 rules), so its ratios are not like-for-like.
  • With pylint, iglint's time is mostly pylint's own: about 0.2 s to start, then pylint's cost per module. That is why the 50-view corpus stays below 10x (6.9x).
  • Without pylint, the speedup is 10x or more from about 300 views on. On the 50-view corpus it is 8x to 11x depending on the run, because iglint takes about 20 ms there and a few milliseconds of noise move the ratio.
  • On fewer CPUs the ratios are smaller: pinned to one CPU, iglint ign-lint is 2.6x to 3.8x faster with pylint and 2.7x to 12x without. On 2 CPUs (a typical hosted CI agent), it is 10x or more faster than ign-lint as shipped (bundled rcfile, jobs=4) on every corpus measured, and 4x to 7x faster than ign-lint's best configuration with pylint.

On a production customer project (287 views, with its own configuration and whitelist; aggregate numbers only), iglint ign-lint took 436 ms with pylint (14x faster than ign-lint's best configuration) and 83 ms without (23x), and iglint check over the whole repository took 124 ms.

There is no fixed start-up cost worth removing: starting the process and reading the configuration take about 2 ms. On a small project, most of the time is spent reading and indexing the rest of the repository (script library, gateway configuration, tags), which iglint needs to resolve what views refer to.

The result cache

Section titled “The result cache”

iglint check caches results in .iglint_cache in the project root. The directory contains a .gitignore that ignores everything in it and a CACHEDIR.TAG, so it stays out of Git and backups without configuration.

  • A file whose content and context have not changed is not linted again.
  • The cache context is the layout, the resource manifests and a hash of every resource summary. An edit inside one view or script that does not change those summaries keeps every other file's results.
  • A test checks every one-file change of the fixture workspace (and a dependent view) against an uncached run, so the cache cannot hide findings.

Options:
--no-cache do not read or write the cache
--cache-dir DIR or cache-dir = "..." put the cache somewhere else

Threads

Section titled “Threads”

iglint uses one worker thread per CPU. --threads N limits that, for example on a shared build agent.

Measuring

Section titled “Measuring”

--timing prints the time spent in each phase to standard error. To reproduce the benchmarks, see bench/run.sh and bench/README.md in the repository: it generates the corpora, installs ign-lint 0.6.4 in a virtual environment and runs hyperfine over every configuration. bench/check-regression.sh fails on a slowdown of more than 10% on the medium corpus, compared with a recorded, machine-specific baseline.