gRPC pool for groupcache

CEll, Wed 20 July 2016, Programming

Go, gRPC, Systems Programming

Groupcache is a system first written at Google to provide a caching layer for the dl.google.com service. It uses consistent hashing to distribute the key space over a group of machines. Each node is able to answer queries by first checking its' own cache or by forwarding the request to a peer that could have it cached. To avoid the thundering herd problem when there's a cache miss, only one request is allowed to read-through to the backend service. Any other requests for the same key would be blocked until the first request returns and the cache is populated.

The default implementation of groupcache uses HTTP to communicate between peers. Whilst this is a perfectly good solution, if you're attemmpting to incorporate groupcache into a gRPC based service, it might be more desirable to use gRPC to communicate between the peers as well. gcgrpcpool is a drop-in replacement for the default HTTPPool which implements peer communications using gRPC.

server := grpc.NewServer()

p := NewGRPCPool("127.0.0.1:5000", server)
p.Set(peerAddrs...)

getter := groupcache.GetterFunc(func(ctx groupcache.Context, key string, dest groupcache.Sink) error {
    dest.SetString(...)
    return nil
})

groupcache.NewGroup("grpcPool", 1<<20, getter)
lis, err := net.Listen("tcp", "127.0.0.1:5000")
if err != nil {
    log.Fatalf("Failed to start server")
}

server.Serve(lis)

Because gRPC connections are long-lived, gcgrpcpool attempts to keep unchanged peer connections intact during a Set method call. A future improvement might be to use a custom gRPC Balancer implementation to maintain the connections.