fastcache

fastcache - fast thread-safe inmemory cache for big number of entries in Go

Features

• Fast. Performance scales on multi-core CPUs. See benchmark results below.
• Thread-safe. Concurrent goroutines may read and write into a single cache instance.
• The fastcache is designed for storing big number of entries without GC overhead.
• Fastcache automatically evicts old entries when reaching the maximum cache size set on its creation.
• Simple API.
• Simple source code.
• Cache may be saved to file and loaded from file.
• Works on Google AppEngine.

Benchmarks

Fastcache performance is compared with BigCache, standard Go map and sync.Map.

…

MB/s column here actually means millions of operations per second. As you can see, fastcache is faster than the BigCache in all the cases. fastcache is faster than the standard Go map and sync.Map on workloads with inserts.

Limitations

• Keys and values must be byte slices. Other types must be marshaled before storing them in the cache.
• Big entries with sizes exceeding 64KB must be stored via distinct API.
• There is no cache expiration. Entries are evicted from the cache only on cache size overflow. Entry deadline may be stored inside the value in order to implement cache expiration.

Architecture details

The cache uses ideas from BigCache:

• The cache consists of many buckets, each with its own lock. This helps scaling the performance on multi-core CPUs, since multiple CPUs may concurrently access distinct buckets.
• Each bucket consists of a hash(key) -> (key, value) position map and 64KB-sized byte slices (chunks) holding encoded (key, value) entries. Each bucket contains only O(chunksCount) pointers. For instance, 64GB cache would contain ~1M pointers, while similarly-sized map[string][]byte would contain ~1B pointers for short keys and values. This would lead to huge GC overhead.

64KB-sized chunks reduce memory fragmentation and the total memory usage comparing to a single big chunk per bucket. Chunks are allocated off-heap if possible. This reduces total memory usage because GC collects unused memory more frequently without the need in GOGC tweaking.

Users

• Fastcache has been extracted from VictoriaMetrics sources. See this article for more info about VictoriaMetrics.

FAQ

What is the difference between fastcache and other similar caches like BigCache or FreeCache?

• Fastcache is faster. See benchmark results above.
• Fastcache uses less memory due to lower heap fragmentation. This allows saving many GBs of memory on multi-GB caches.
• Fastcache API is simpler. The API is designed to be used in zero-allocation mode.

Why fastcache doesn't support cache expiration?

Because we don't need cache expiration in VictoriaMetrics. Cached entries inside VictoriaMetrics never expire. They are automatically evicted on cache size overflow.

It is easy to implement cache expiration on top of fastcache by caching values with marshaled deadlines and verifying deadlines after reading these values from the cache.

Why fastcache doesn't support advanced features such as thundering herd protection or callbacks on entries' eviction?

Because these features would complicate the code and would make it slower. Fastcache source code is simple - just copy-paste it and implement the feature you want on top of it.