mirror of
https://github.com/pnpm/pnpm.git
synced 2026-06-27 09:25:24 -04:00
1310ab53c4
* perf(cli): keep the repeat-install fast path off the lockfile parse and thread spawns The "Already up to date" short-circuit decides from manifest mtimes alone (mirroring upstream checkDepsStatus, which never reads the wanted lockfile), yet pacquet parsed pnpm-lock.yaml eagerly in State::init before the check ran — a multi-millisecond YAML parse on every no-op install, scaling with lockfile size (babylon's 720 KB lockfile dominated its repeat-install wall time). pnpm-lock.yaml now loads through a LazyLockfile (OnceLock-backed) that install forces only after the fast path has passed on the run; add / update / remove / outdated / the pnpr path force it up front, keeping their behavior unchanged. The repeat-install regenerate branch probes for the file's existence instead of its parsed contents, so the fast path stays mtime-cheap. The rayon global pool is likewise no longer built eagerly at startup: the worker count is published via RAYON_NUM_THREADS (set in fn main while the process is still single-threaded) so the pool spawns lazily on first parallel use — commands that never reach a parallel phase no longer pay 2×CPUs thread spawns. A corrupt lockfile now surfaces its parse error when the install actually reads the file; an up-to-date project with an unreadable lockfile reports "Already up to date" exactly as pnpm does. * perf(cli): finish up-to-date installs before building the async runtime A plain `pacquet install` that is already up to date now completes on the main thread, before the tokio runtime, the rayon pool, the HTTP client, or any install state exists. The new `install_already_up_to_date` twin of the repeat-install short-circuit reuses the exact same workspace discovery and `check_optimistic_repeat_install` inputs as `Install::run`, and the CLI invokes it from the (now synchronous) `main` after clap parsing. Gates mirror everything that would make the full path behave differently: `--frozen-lockfile` / `--lockfile-only`, a configured pnpr server (that path never runs the optimistic check), `--recursive` / `--filter`, config dependencies, and pnpmfile updateConfig hooks (both can mutate the config the check compares against). Any gate or error falls through to the full install path, which re-runs the check and reproduces failures with their established error shape; the "Already up to date" + summary emissions are byte-identical. Repeat-install instruction count on the vue fixture drops from ~203M to ~41M retired instructions — within a rounding error of the `pacquet --version` floor (~39M). * perf(resolving): renew metadata-mirror freshness on 304 Not Modified The minimumReleaseAge freshness shortcut treats a metadata mirror younger than the cutoff as authoritative and resolves without touching the network. But a 304 revalidation never rewrote the mirror file, so its mtime froze at the last 200 response: once a cached packument grew older than minimumReleaseAge (24h by default), every subsequent install re-validated every package against the registry, forever. A 304 proves the cached packument equals the registry's current document, so the validation clock legitimately restarts at the response: bump the mirror's mtime to now (fire-and-forget — a read-only cache dir only costs the next install another conditional request). Applied to both stacks: pnpm's pickPackage notModified branch and pacquet's fetch_full_metadata_cached 304 path. On a vue-fixture install with a stale cache, the second warm resolve drops from ~2s (520 conditional requests) to ~250ms (zero requests). * style(resolving): use clippy's preferred Duration units in the 304 mtime test CI clippy denies duration_suboptimal_units; from_hours / from_mins replace the hand-multiplied from_secs values. * style(package-manager): use clippy's preferred Duration unit in the sync fast-path test Same duration_suboptimal_units deny as the previous commit, one site CI's clippy surfaced after it stopped at the first failing crate. * fix(lockfile): address review — dir-addressed LazyLockfile, read-open 304 touch LazyLockfile resolved pnpm-lock.yaml against the process cwd while the CLI honours a canonicalized --dir without chdir, so the deferred load and the existence probe could consult a different lockfile than the rest of the install (which derives lockfile_path from the manifest's directory). The lazy handle now carries the manifest's directory and loads via load_wanted_from_dir; lockfile-disabled config gets an explicit disabled() constructor. The pre-runtime fast path builds its handle from the same directory, keeping verdict parity with Install::run. renew_mirror_freshness opened the mirror with append just to bump the mtime; set_modified only needs a file handle plus ownership (futimens semantics), so a plain read-open also covers mirrors whose mode dropped write permission. * test(integrated-benchmark): compare best-of-N samples in the slow-start proxy test The test raced two single wall-clock samples, and a loaded CI runner can inflate the ramped-vs-flat comparison in either direction (macOS runner measured flat 318ms vs ramped 304ms against a ~66ms model). Scheduler stalls only ever inflate a sample, while slow start's ramp overhead is structural and survives in every sample — so the minimum of several runs per side is the noise-resistant estimator. * chore(deps): bump esbuild to 0.28.1 to clear GHSA-gv7w-rqvm-qjhr The new advisory (install-module RCE via NPM_CONFIG_REGISTRY, patched in 0.28.1) fails the audit gate. 0.28.1 was published within the minimumReleaseAge window, so the patched version is excluded from the age gate — the same mechanism pnpm audit --fix uses — including its '@esbuild/*' platform packages, whose versions move in lockstep with the root package. * fix(lockfile): make the unloaded presence probe match the loader's absence rules The loader treats an empty file and an env-only combined document as an absent wanted lockfile (Ok(None)), but is_loaded_or_on_disk probed bare file existence, so the repeat-install path could skip restoring a semantically-missing pnpm-lock.yaml. The probe now reads the file and checks the main document is non-empty (Lockfile::wanted_exists_in_dir) — the loader's exact absence rules, still without the YAML parse. * fix(cli): build the rayon pool after the fast-path gate instead of injecting env Publishing the worker count through RAYON_NUM_THREADS leaked the variable into every child process the install spawns — lifecycle scripts, node probes, git — and pnpm exposes no such variable to scripts. Build the global pool with ThreadPoolBuilder again, but only once the repeat-install fast path has declined: real installs pay exactly the cost they always did, the no-op path still spawns no workers, and the process environment stays untouched (which also drops the unsafe set_var and its single-threaded contract). * fix(lockfile): treat only NotFound as absence in the presence probe A permission or I/O failure reading pnpm-lock.yaml reported the file as absent, which would send the repeat-install path into the regenerate-on-missing branch — overwriting an existing lockfile it merely could not read. Only NotFound counts as absent now; any other read failure reports presence, and the real load surfaces the underlying error when the contents are actually needed. * fix(resolving): open the mirror write-capable for the 304 touch, read-only as fallback Windows' set_modified requires write-attributes access on the handle, so the read-only open silently failed there (caught by the Windows CI run of a_304_renews_the_mirror_mtime). Append-mode open carries that access; the read-only fallback still covers Unix mirrors whose mode dropped write permission, where timestamp syscalls need ownership rather than write access. * fix(package-manager): never short-circuit partial installs as already up to date add and remove mutate the manifest in memory and persist it only after Install::run returns, so the on-disk mtimes the optimistic repeat-install check reads still describe the pre-mutation project. With a fresh workspace state, `pacquet add X` right after a clean install reported "Already up to date", skipped the entire install, and then saved a package.json declaring a dependency that was never resolved, lockfiled, or materialized (self-healing on the next run, which sees the newer manifest mtime). Gate the short-circuit on is_full_install, mirroring upstream installDeps calling checkDepsStatus only for the plain-install mutation, never for installSome / uninstallSome. The new partial_install_disables_optimistic_short_circuit test fails without the gate. The bug predates this PR (the KeepAll gate has carried add since the optimistic path landed) — surfaced by CodeRabbit review on pnpm/pnpm#12364.