fluuz

A Spring CacheManager with distributed evictions.

Distributed caches solve the problem of consistency at the expense of performance. In general, distributed caches rely on serialization of objects, and synchronization of cache nodes.

An in-memory cache such as Guava or caffeine provides the best performance, at the expense of consistency. For high-read low-write situations it is tempting to use an in-memory cache, however there is the risk that caches can become stale and out of sync.

This project aims to provide in-memory caches with no stale data. Whenever a key/value is forcibly evicted, this eviction is broadcast to other nodes who then evict the key. Local evictions due to policy are not broadcast, as they do not represent a change of the state of data.

Limitations

A system using this strategy needs to be written with it in mind. Specifically, whenever data that might be cached is altered at the source (datastore), an eviction notice must be published for the affected keys.

Design

Fluuz provides a CacheManager implementation that wraps another CacheManager and accepts an EventManager implementation. The EventManager keeps track of forced evictions, notifying peers of local forced evictions, and propagating evictions from peers' notifications. In this way a cache might miss, but will not be out of date.

The underlying CacheManagers do not need to be configured identically or even be of the same implementation type between nodes. One node might use a CaffeineCacheManager with a LRU eviction policy and another might use a 'EhCacheCacheManager' with a LFU policy and write to disk. All peers will however need to use the same EventManager implementation, configured consistently.

Implementation

RedisEventManager is provided as an EventManager implementation that uses Redis PUB/SUB for communication between nodes. It requires a Redis server (or cluster) host and port in its constructor.