thanos-community/galaxycache

galaxycache is a caching and cache-filling library based on groupcache.

galaxycache

galaxycache is a caching and cache-filling library, adapted from groupcache, intended as a replacement for memcached in many cases.

For API docs and examples, see http://godoc.org/github.com/vimeo/galaxycache

Quick Start

Initializing a peer

// Generate the protocol for this peer to Fetch from others with (package includes HTTP and gRPC)
grpcProto := NewGRPCFetchProtocol(grpc.WithInsecure())

// HTTP protocol as an alternative (passing the nil argument ensures use of the default basepath
// and opencensus Transport as an http.RoundTripper)
httpProto := NewHTTPFetchProtocol(nil)

// Create a new Universe with the chosen peer connection protocol and the URL of this process
u := NewUniverse(grpcProto, "my-url")

// Set the Universe's list of peer addresses for the distributed cache
u.Set("peer1-url", "peer2-url", "peer3-url")

// Define a BackendGetter (here as a function) for retrieving data
getter := GetterFunc(func(ctx context.Context, key string, dest Codec) error {
   // Define your method for retrieving non-cached data here, i.e. from a database
})

// Create a new Galaxy within the Universe with a name, the max capacity of cache space you would
// like to allocate, and your BackendGetter
g := u.NewGalaxy("galaxy-1", 1 << 20, getter)

// In order to receive Fetch requests from peers over HTTP or gRPC, we must register this universe
// to handle those requests

// gRPC Server registration (note: you must create the server with an ocgrpc.ServerHandler for
// opencensus metrics to propogate properly)
grpcServer := grpc.NewServer(grpc.StatsHandler(&ocgrpc.ServerHandler{}))
RegisterGRPCServer(u, grpcServer)

// HTTP Handler registration (passing nil for the second argument will ensure use of the default 
// basepath, passing nil for the third argument will ensure use of the DefaultServeMux wrapped 
// by opencensus)
RegisterHTTPHandler(u, nil, nil)

// Refer to the http/grpc godocs for information on how to serve using the registered HTTP handler
// or gRPC server, respectively:
// HTTP: https://golang.org/pkg/net/http/#Handler
// gRPC: https://godoc.org/google.golang.org/grpc#Server

Getting a value

// Create a Codec for unmarshaling data into your format of choice - the package includes 
// implementations for []byte and string formats, and the protocodec subpackage includes the 
// protobuf adapter
sCodec := StringCodec{}

// Call Get on the Galaxy to retrieve data and unmarshal it into your Codec
ctx := context.Background()
err := g.Get(ctx, "my-key", &sCodec)
if err != nil {
   // handle if Get returns an error
}

// Shutdown all open connections between peers before killing the process
u.Shutdown()

Concepts and Glossary

Consistent hash determines authority

A consistent hashing algorithm determines the sharding of keys across peers in galaxycache. Further reading can be found here and here.

Universe

To keep galaxycache instances non-global (i.e. for multithreaded testing), a Universe object contains all of the moving parts of the cache, including the logic for connecting to peers, consistent hashing, and maintaining the set of galaxies.

Galaxy

A Galaxy is a grouping of keys based on a category determined by the user. For example, you might have a galaxy for Users and a galaxy for Video Metadata; those data types may require different fetching protocols on the backend -- separating them into different Galaxies allows for this flexibility.

Each Galaxy contains its own cache space. The cache is immutable; all cache population and eviction is handled by internal logic.

Maincache vs Hotcache

The cache within each galaxy is divided into a "maincache" and a "hotcache".

The "maincache" contains data that the local process is authoritative over. The maincache is always populated whenever data is fetched from the backend (with a LRU eviction policy).

In order to eliminate network hops, a portion of the cache space in each process is reserved for especially popular keys that the local process is not authoritative over. By default, this "hotcache" is populated by a key and its associated data by means of a requests-per-second metric. The logic for hotcache promotion can be configured by implementing a custom solution with the ShouldPromote.Interface.

Step-by-Step Breakdown of a Get()

When Get is called for a key in a Galaxy in some process called Process_A:

1. The local cache (both maincache and hotcache) in Process_A is checked first
2. On a cache miss, the PeerPicker object delegates to the peer authoritative over the requested key
3. Depends on which peer is authoritative over this key...

• If the Process_A is the authority:
  • Process_A uses its BackendGetter to get the data, and populates its local maincache
• If Process_A is not the authority:
  • Process_A calls Fetch on the authoritative remote peer, Process_B (method determined by FetchProtocol)
  • Process_B then performs a Get to either find the data from its own local cache or use the specified BackendGetter to get the data from elsewhere, such as by querying a database
  • Process_B populates its maincache with the data before serving it back to Process_A
  • Process_A determines whether the key is hot enough to promote to the hotcache
    • If it is, then the hotcache for Process_A is populated with the key/data

4. The data is unmarshaled into the Codec passed into Get

Changes from groupcache

Our changes include the following:

• Overhauled API to improve usability and configurability
• Improvements to testing by removing global state
• Improvement to connection efficiency between peers with the addition of gRPC
• Added a Promoter.Interface for choosing which keys get hotcached
• Made some core functionality more generic (e.g. replaced the Sink object with a Codec marshaler interface, removed Byteview)

New architecture and API

• Renamed Group type to Galaxy, Getter to BackendGetter, Get to Fetch (for newly named RemoteFetcher interface, previously called ProtoGetter)
• Reworked PeerPicker interface into a struct; contains a FetchProtocol and RemoteFetchers (generalizing for HTTP and GRPC fetching implementations), a hash map of other peer addresses, and a self URL

No more global state

• Removed all global variables to allow for multithreaded testing by implementing a Universe container that holds the [Galaxies] (previously a global groups map) and PeerPicker (part of what used to be HTTPPool)
• Added methods to Universe to allow for simpler handling of most galaxycache operations (setting Peers, instantiating a Picker, etc)

New structure for fetching from peers (with gRPC support)

• Added an HTTPHandler and associated registration function for serving HTTP requests by reaching into an associated Universe (deals with the other function of the deprecated HTTPPool)
• Added gRPC support for peer communication, with a function for gRPC server registration
• Reworked tests to fit new architecture
• Renamed files to match new type names

A smarter Hotcache with configurable promotion logic

• New default promotion logic uses key access statistics tied to every key to make decisions about populating the hotcache
• Promoter package provides a ShouldPromote.Interface for creating your own method to determine whether a key should be added to the hotcache
• Newly added candidate cache keeps track of peer-owned keys (without associated data) that have not yet been promoted to the hotcache
• Provided variadic options for Galaxy construction to override default promotion logic (with your promoter, max number of candidates, and relative hotcache size to maincache)

Comparison to memcached

See: https://github.com/golang/groupcache/blob/master/README.md

Help

Use the golang-nuts mailing list for any discussion or questions.