Struct LocalDiskStore 

pub struct LocalDiskStore { /* private fields */ }

Local filesystem CAS: one file per content hash under root.

Implementations

impl LocalDiskStore

pub fn new(root: impl Into<PathBuf>) -> LocalDiskStore

Trait Implementations

impl ContentStore for LocalDiskStore

fn put(&self, key: &InputHash, bytes: &[u8]) -> Result<(), Error>

Store bytes under key, atomically.

#2 — Model R concurrent-writer safety

Before Model R, daemons were single-writer per process and PID-scoped their CAS dir, so an in-place fs::write was adequate. Model R’s whole point is a shared --cas-dir across a fleet of daemons: N writers, multiple processes, one directory. An in-place truncate-then-write is not safe there — a concurrent reader doing contains()→get() can observe a half-written file. Content-addressing guarantees the final bytes converge (same key ⇒ identical bytes); it does not make an interleaved read whole.

Fix: write to a unique sibling temp file, fsync it, then rename it onto the final path. POSIX rename(2) within one filesystem is atomic — a reader sees either the old state (absent) or the complete new file, never a torn one. If another writer wins the race the rename simply replaces an identical-content file (content-addressed), so the result is still correct and the put stays idempotent (the AC#5 cache-hit invariant). The exists() fast-path is kept so a cache hit costs one stat, not a rewrite.

(Windows non-atomic-replace is out of scope — CLAUDE.md parks Windows in v1; macOS + Linux are the supported fleet hosts.)

fn get(&self, key: &InputHash) -> Result<Option<Vec<u8>>, Error>

Fetch the bytes for key, or None if absent.

fn contains(&self, key: &InputHash) -> Result<bool, Error>

Whether key is already present — the build pipeline checks this to skip a build entirely on a cache hit.