There's a lot of optimization opportunities with the formatter with a hot cache (the example is apache airflow with rayon removed):

• Sorting the result takes a long time so does building the result vec, even though we know its approximate size
• Don't read source files in formatter when there is a cache hit #8132
• Deserializing the cache takes noticeable time even if we just need 4k paths plus timestamps
• We first iterate over each directory only to check the timestamps later, could those be combined?

Instructions for generating the flamegraph:

• Install perf and cargo install flamegraph
• git clone apache airflow
• Apply https://gist.github.com/konstin/2e8c45c7f1fa53dc003833870226bab8
• flamegraph -F 9999 -- target/release-debug/ruff format ../airflow/ (Set your cpu to powersave for more samples)