Invalidation Strategies

Ways to make caches forget

Delete on write: after committing a change, delete the affected keys; the next read reloads. Prefer deleting to updating the cached value, which risks writing an older value over a newer one. Versioned keys: include a version number or content hash in the key (profile:42:v17); a change increments the version stored in the database, readers build keys from the current version, and old entries simply age out. This is also how fingerprinted static assets work (app.3f9a1c.js). Tag-based invalidation: label entries with the entities they depend on and invalidate by tag; CDNs offer this as cache tags or surrogate keys. Event-driven invalidation: services publish change events and cache owners invalidate on receipt. Change data capture (CDC): a process reads the database's change log (for example with Debezium) and invalidates keys for every committed change, catching even writes that bypass application code.

Invalidation from a change data capture stream

Every committed row change triggers invalidation, whichever code path made it.

def on_change_event(event):
    """Called for each committed row change read from the database's change log."""
    table = event["source"]["table"]
    row = event["after"] or event["before"]           # after is empty for deletes

    if table == "products":
        redis.delete(f"product:v2:{row['id']}")
        cdn.purge_tags([f"product-{row['id']}", f"category-{row['category_id']}"])
    elif table == "prices":
        redis.delete(f"product:v2:{row['product_id']}")
        cdn.purge_tags([f"product-{row['product_id']}"])

A library recall slip

When a book edition is corrected, the library sends recall slips to every branch holding a copy (invalidation by tag). Printing a new edition number on the cover (versioned keys) means branches can simply stop lending the old edition.

Quick check: Why is deleting a cache key on write usually safer than writing the new value into the cache?

• Deleting is faster in every cache
• Writing to caches is not allowed
• Deleting also updates the database
• Concurrent writers could otherwise leave an older value in the cache last