Preparing your design workspace

Loading the problem and your latest saved attempt…

Large Model Artifact Cache

Design a distributed cache that avoids downloading a 40 GB model artifact on every container cold start.

Problem statement

For this exercise, assume immutable model artifacts contain a manifest and content-addressed chunks, with individual artifacts around 40 GB. GPU nodes may request the same artifact concurrently.

The system should make the second start fast through regional and node caches, coalesce duplicate downloads, verify integrity, evict safely under byte pressure, and prevent a running container from losing pinned data. Exclude model-registry authoring and object-store implementation.

What to build

  1. Resolve an immutable model version to a signed manifest and verified chunks.
  2. Coalesce concurrent misses and materialize an artifact on a target node with progress and cancellation.
  3. Pin active chunks, evict safe candidates, repair corruption, and expose cache health.

At scale

  • Artifacts average 40 GB and consist of 4 to 64 MB content-addressed chunks.
  • A region may launch 1,000 containers for one artifact within a minute.
  • Warm node-cache hits should add less than 500 ms before runtime loading begins.
  • Corrupted or partial chunks must never be presented as complete artifacts.