How would you combine an in-process L1 cache with Redis as L2, and keep the L1 copies consistent?

Read through the layers in order, and use Redis Pub/Sub to tell every instance to delete changed keys. A short TTL on L1 is the safety net.

Reading:

• Check the local L1 cache first. L1 is a cache inside the app process's own memory.
• On a miss, check Redis (the L2 cache, shared by all instances).
• On a miss, read the database.
• On the way back, fill each layer with the value.

Give L1 a short TTL (time to live), in seconds rather than minutes. Also give it a size limit. This keeps stale data and heap (process memory) use small.

Invalidation (removing old copies):

• A service updates the data.
• It deletes the Redis key.
• It publishes the key name on a Pub/Sub channel.
• Every instance subscribes to that channel and evicts the key from its L1.

Pub/Sub can lose messages while a connection is down. So the short L1 TTL stays as the safety net. An instance can also clear its whole L1 after it reconnects.

Redis 6 client-side caching: the CLIENT TRACKING command lets the server help. The server remembers which keys a connection read. In broadcast mode, it tracks key prefixes instead. When those keys change, the server pushes invalidation messages to the client. Client library support varies, so check your library.