Local Cache

Every time Crab downloads chunks from cloud storage (during hydration or fetch), it stores them in a local cache. The next time you need those chunks — re-hydrating the same file, switching branches, or a teammate hydrating shared content on a shared machine — the data is served from local disk instead of the network.

How the Cache Works

The cache is a content-addressed store — chunks are identified by their blake3 hash. This means:

• Cross-file sharing — If two files share chunks (common in dataset versions), the cache serves both from a single copy.
• Cross-branch sharing — Switching branches and re-hydrating reuses cached chunks from the previous branch.
• Idempotent — Fetching the same chunk twice is a no-op; it's already cached.

Cache Location

To use a fast NVMe drive for the cache:

export CRAB_CACHE_DIR=/mnt/nvme/crab-cache

Cache Size Management

The cache grows as you hydrate files. You can set a maximum size:

# .crab/config.toml
[cache]
max_size = "50GB"

When the cache exceeds this limit, least-recently-used chunks are evicted. Eviction is lazy — it happens during the next cache write, not in the background.

Inspecting the Cache

crab cache stats

Shows total size, number of objects, hit rate, and cache directory path.

Cleaning the Cache

# Remove everything (re-download on next hydrate)
crab cache clean

# Remove only unreferenced objects
crab prune

crab prune is usually preferred — it keeps chunks that are still referenced by your current working tree, only removing orphaned data.

Pre-warming the Cache

If you know you'll need certain files later (going offline, preparing for a demo), pre-warm the cache:

crab fetch --include '*.safetensors'

This downloads chunks without hydrating files — the data sits in cache ready for instant hydration later.

When to Clean vs. Prune

Remote Cache Service

The local cache is per-machine — each developer and CI runner maintains their own. For teams, Crab offers an optional remote cache service (crab-cache-server) that shares cached objects across all machines in your organization.

When configured, the lookup order is: local cache → remote cache → cloud storage. The first fetch from any team member warms the remote cache, and subsequent fetches from anyone else are served at network-local speed instead of cloud latency.

For deployment and configuration details, see the Cache Service guide.

CLI Reference

For complete command syntax, see the crab cache reference.