A dry run marks a step `unknown` behind a cached prerequisite that restores outputs, though the restore gives it the exact files it recorded

Offen
  • Stimmt etwas nicht?
#938 0 Kommentare 0 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen

Maintainer antworten meist innerhalb von 1 Tag

Dieses Issue hat noch niemand übernommen.

Bewertung

Schwierigkeit
5/5
Geschätzter Aufwand
Über eine Woche
Anfängerfreundlichkeit
35/100
Issue-Typ
Feature
Klarheit
Klar beschrieben
Aktivitätsstatus
Aktiv
Tech-Stack
rust
Bereich
build-system, devtools

Rechercherichtung

Start with CacheEntryValue in crates/vt/src/session/cache/mod.rs:104-116 and the dry-run RFC in issue #933, including its linked discussion. Trace how cache entries record archives and how dry-run checks outputs and inputs; the issue calls for a restored view that includes archived paths and parent directories. Done when the cache records archive contents and steps waiting on cache hits are checked against the restored view, with the stated unknown cases preserved.

Vom Indexierungsmodell aus dem Issue-Text verfasst.

Beschreibung

Written by an AI agent (Claude) on behalf of @jasonkuhrt.

The dry-run RFC in #933 marks a step unknown when a step it waits on would run, or would be a hit that restores outputs. Its Future section names the next step, and the discussion on #933 moved it to an RFC of its own: if cache entries recorded each output's path and hash, a dry run could check those steps against the files the hit will restore. This is that ask.

What happens

The RFC's example on a fresh checkout: both builds cached from one successful run, nothing else changed, NODE_ENV included.

// vp run --dry-run --json -r build, as the RFC specifies it
"packages/ui#build": { /* … */ "steps": [{ /* … */ "cache": { /* … */ "status": "hit" } }] },
"apps/web#build": {
  /* … */
  "dependsOn": ["packages/ui#build"],
  "steps": [{ /* … */ "cache": { /* … */ "status": "unknown" } }]
}

The real run replays @acme/ui#build, which restores packages/ui/dist/index.js, and @acme/web#build then hits. The dry run can't say so: an entry's value holds its archive's file name but not what the archive holds (CacheEntryValue, crates/vt/src/session/cache/mod.rs:104-116 at c4b2aacf).

On a fresh checkout every generated file is absent until its producer runs or replays, so a cached producer is a hit that restores, and every step behind it that would otherwise be hit or changed reads unknown. A consumer that plans from the dry run, such as CI picking a pull request's jobs (vite-plus#2848), has to treat each of them as a step that may run.

The ask

Record what each entry's archive holds, and run today's check of a step that waits on hits against the files as those hits will leave them:

sequenceDiagram
    participant DR as vp run --dry-run
    participant UI as @acme/ui#35;build entry
    participant D as files on disk
    DR->>UI: check against the disk
    UI-->>DR: hit, will restore dist/index.js (hash h1)
    Note over DR: restored view = disk + what the hit restores
    DR->>D: read @acme/web#35;build's other recorded paths
    Note over DR: check @acme/web#35;build against the view:<br/>dist/index.js is h1 as recorded, the rest unchanged
    Note over DR: @acme/web#35;build is hit
What the check runs against

The restored view: the files as the hits a step waits on, directly or through other steps, will leave them.

  • Each archived path brings what the check compares for it: a file's hash, a symlink's target, or a directory.
  • The parent directories a restore creates are in the view too.
  • Every other path is as it is on disk.

The entry records what its archive holds when it archives its outputs.

What it checks

Today's check, unchanged, run against the view instead of the disk:

  • recorded files
  • missing paths
  • directories checked for existence
  • directory listings
  • cache.input globs, which can match a file only a hit restores
When a step stays unknown
  • It waits on a step that would run, or on one that is itself unknown.
  • It reads a path whose content depends on which of two unordered hits restores it last, and no hit ordered after both restores it again.

The RFC's other rules hold: only a step that would otherwise be hit or changed can be unknown, and noBaseline, error and disabled stay as they are.

Cost and compatibility
  • A remote hit costs no download. The dry run already reads the entry's value where a real run would, and the value records what the archive holds.
  • -vv: baseline.outputs gives each output with what the check compares for it, so a consumer can see what a hit will restore.
  • The cache schema version bumps, as it did for #913, so existing caches fill again.
Our case

We'd like CI to pick a pull request's jobs from the dry run, checked against the entries our default branch uploads. We never commit generated source: in one local cache of our workspace, 104 of 601 step entries recorded at least one generated-source file, and each is absent on a fresh checkout until its producer replays. With this, those steps read as hit where the real run would hit.

Vorherrschende Sprache
Rust
Sterne
485
Forks
51
Ø Merge
1 T. 3 Std.
Gemergte PRs (30 T.)
70

Entwicklungsumgebung

In Codespaces öffnen

Startet den Dev-Container des Projekts im Browser, mit Ihrem eigenen GitHub-Konto.

Erste Schritte

  1. Lesen Sie das ganze Issue und danach den Beitragsleitfaden des Projekts.
  2. Schreiben Sie ins Issue, dass Sie es übernehmen — das erspart doppelte Arbeit.
  3. Forken Sie das Repository und arbeiten Sie in einem Branch.
  4. Öffnen Sie einen Pull Request, der die Issue-Nummer nennt.

Mehr aus voidzero-dev/vite-task

Alle Issues in voidzero-dev/vite-task

Ähnliche Issues

Weitere Issues zu Rust

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.