[SizzleUnrlsd]

Profile

mitmproxy, -j 8, main: 501 components, each one a task.

Phase	Time
Preparing the project per task: plugins, dependencies, advisories, models	49 s of CPU
Decoding the transitive summaries each task received	32 s of CPU
Parent encoding and submitting those summaries, sequentially	32 s of wall time
Analysis itself	58 s of CPU

364 MB of summaries were shipped in total.

Change

• Each worker prepares the project once, through the pool's initializer (_Project, _start_worker).
• A task carries only its paths and, for each imported module, its import name and spool key.
• The parent stores finished modules in a spool after each wave: the --cache directory, or a temporary directory of the run, removed at the end. Each worker reads an entry at most once.
• A missing entry raises instead of analysing with missing summaries.
• encode_index and decode_index, now unused, are removed.
• The -j 1 path is unchanged.

Measurements

macOS, 8 cores (4 performance, 4 efficiency), Python 3.14, bundled plugins, cold start, low load. Reports are identical in every mode.

Project	-j 1	-j 8 on main	-j 8 with this change	User time of -j 8
healthchecks	15.3–15.5 s	15.6–15.9 s	7.9–9.2 s	52 s → 27 s
mitmproxy	51.6 s	68.5 s	45.6 s	178 s → 90 s

Runs under heavy external load point the same way (user time, -j 8: 286 s → 124 s).

What remains

mitmproxy gains little over -j 1. One 37-module import cycle is analysed by a single process, and each wave waits for its slowest component. docs/usage.md now says so, in place of the "saves a quarter" sentence, with the measured table.

Verification

• tests/test_jobs_workers.py, new. One test checks that a worker instantiates the plugins once (it fails on main: 8 instantiations for 2 workers). Two tests check that summaries reach the modules of later waves, with and without --cache, with the same report as -j 1. No existing test changed.
• Full suite green, apart from the local test_packaging metadata issue (fails on main too). ruff check and mypy clean.