Notesfoundations🌱 seedling
Cache Invalidation
Also: Invalidation · TTL · Time to Live · Cache Lease
Invalidation is how a cache learns its copy is out of date: expire it after a TTL, delete it when the source changes, or derive deletes from the source's change log. Races mean you need a TTL as the backstop.
Why it exists
A cache is a copy, and copies go stale. Without invalidation, users read old data indefinitely.
How it works
- TTL: entries expire after a fixed lifetime. Self-healing, with staleness bounded by the TTL.
- Explicit: the writer deletes or updates the key after writing the source. Fresh, but every write path must comply.
- Change-driven: tail the database's commit log (Change Data Capture) and fan out deletes. Durable, replayable, and covers writes from any client.
- Leases: on a miss, the cache issues a token for the key; a delete voids it; a later fill with a voided token is rejected. This stops a slow reader writing back a value that's already been superseded.
Trade-offs
| Approach | Gains | Costs |
|---|---|---|
| TTL only | trivial, self-healing | up to one TTL of staleness; misses on expiry |
| explicit delete | fresh | easy to miss a write path; races |
| log-driven | complete, replayable | pipeline lag; more infrastructure |
In practice
- Facebook:
mcsquealdaemons read the MySQL commit log and broadcast deletes to every cache cluster. Leases (64-bit tokens per key) prevent stale sets.
Pitfalls
- No TTL: one lost invalidation or race leaves a stale entry forever.
- Invalidate-then-write order: deleting before the database commit lets a concurrent read refill the old value. Delete after the commit.
Connections
- Part of
- Caching
Recall
These cards are also in your review deck, scheduled with spaced repetition.