Merge pull request #2584 from ludovic111/perf/parity-manifest-aware
🤖 Agent: crate splits' workspace-only manifest changes no longer select every harness (354 to 209 on #2160 and #2215); other PRs select as before.
This commit is contained in:
@@ -64,7 +64,7 @@ A harness can declare, at the top level of its script:
|
||||
|
||||
### Which harnesses a change affects
|
||||
|
||||
`--changed` and `--changed-files` narrow each harness's Rust sources to those its probes can reach (`tools/parity_reach.py`), so a change to one module of `horizon-client` no longer reruns every harness whose probe links the crate. The narrowing works on items, not files: from the probe modules a harness ran (or the binary it built), a free item (fn, type, const, static) is reached when a reached item names it in its own module, through a path or an import of a module that is named, or through a glob import; an inherent impl's methods count one by one once their type is reached and their name is mentioned; a trait impl counts once its trait and its type are reached (std's traits and types always are); a macro counts once invoked. Names match by spelling, so the result only over-approximates. A reached file's parent modules count through their `mod` declarations of it, compared with the merge base; Rust files a change adds take part, a dependency that was deleted counts as changed, code under `#[cfg(test)]` never counts, and a harness whose probe can't be found keeps every source. Parses are cached by content in `reach-cache.json` of the shared parity folder (beside `passes/`), so a new worktree starts warm; a harness's search stops as soon as it reaches a changed file, and after `PARITY_REACH_BUDGET` seconds (default 180) the remaining harnesses count every source their probes link, so choosing never holds a slow machine for long. `--by-crate` turns the narrowing off. On the last 20 pull requests merged on 2026-10-25, the harnesses picked went from 2255 to 1061 (2384 to 1592 harness-minutes; game-binary harnesses, which link all of `game/`, set the floor), and every harness a CI parity run failed in the week before was still picked; every source file a sample of probes executed (LLVM `-C instrument-coverage`) is among those kept. Building the graph takes 15 s cold, under a second warm, and up to a minute and a half on a change that reaches many harnesses.
|
||||
`--changed` and `--changed-files` narrow each harness's Rust sources to those its probes can reach (`tools/parity_reach.py`), so a change to one module of `horizon-client` no longer reruns every harness whose probe links the crate. The narrowing works on items, not files: from the probe modules a harness ran (or the binary it built), a free item (fn, type, const, static) is reached when a reached item names it in its own module, through a path or an import of a module that is named, or through a glob import; an inherent impl's methods count one by one once their type is reached and their name is mentioned; a trait impl counts once its trait and its type are reached (std's traits and types always are); a macro counts once invoked. Names match by spelling, so the result only over-approximates. A reached file's parent modules count through their `mod` declarations of it, compared with the merge base; Rust files a change adds take part, a dependency that was deleted counts as changed, code under `#[cfg(test)]` never counts, and a harness whose probe can't be found keeps every source. Parses are cached by content in `reach-cache.json` of the shared parity folder (beside `passes/`), so a new worktree starts warm; a harness's search stops as soon as it reaches a changed file, and after `PARITY_REACH_BUDGET` seconds (default 180) the remaining harnesses count every source their probes link, so choosing never holds a slow machine for long. `--by-crate` turns the narrowing off. With `--changed`, `Cargo.lock` and `Cargo.toml` changes that only add, remove or move workspace crates (a crate split: no registry or git package, version, feature, dependency or build setting changed across them, compared with the merge base) don't count: the moved code counts through its sources. A workspace crate's feature that such a change defines or turns on is allowed only where every source naming it changed too (the new crate's own files). #2160's split went from 352 harnesses to 123 this way (its own changes to the runner still select them all). On the last 20 pull requests merged on 2026-10-25, the harnesses picked went from 2255 to 1061 (2384 to 1592 harness-minutes; game-binary harnesses, which link all of `game/`, set the floor), and every harness a CI parity run failed in the week before was still picked; every source file a sample of probes executed (LLVM `-C instrument-coverage`) is among those kept. Building the graph takes 15 s cold, under a second warm, and up to a minute and a half on a change that reaches many harnesses.
|
||||
|
||||
## The gate
|
||||
|
||||
|
||||
Reference in New Issue
Block a user