NCP is a tiny but very fast caching proxy written in Go. It uses Redis for data storage, which in combination with the speed of Go makes it incredibly fast. NCP is well-optimized and can be run on almost any platform, so if you have a Raspberry Pi, you can install NCP as your local cache there.
NCP is modular. Now it has only one database adapter which is Redis. If you need support for any other database, feel free to open an issue or implement it on your own and then open a pull request (bonus points).
NCP is very simple. It just proxies requests to an upstream registry, caches response and returns cached response for next requests to the same package. Cached data are stored in Redis with an original request URL as a key.
NCP binaries for different paltforms can be downloaded can be downloaded on the Releases page. Also, Docker image is provided on Docker Hub.
The quickies way to get started with NCP is to use Docker.
# run proxy inside of docker container in background
docker run -e REDIS_ADDRESS=host.docker.internal:6379 -p 8080:8080 -it -d pkgems/npm-cache-proxy
# configure npm to use caching proxy as registry
npm config set registry http://localhost:8080
NCP provides command line interface for interaction with a cached data.
Options
Options | Env | Default | Description |
---|---|---|---|
--listen <address> |
LISTEN_ADDRESS |
locahost:8080 |
Address to listen |
--upstream <address> |
UPSTREAM_ADDRESS |
https://registry.npmjs.org |
Upstream registry address |
--silent <address> |
SILENT |
0 |
Disable logs |
--cache-limit <count> |
CACHE_LIMIT |
- | Cached packages count limit |
--cache-ttl <timeout> |
CACHE_TTL |
3600 |
Cache expiration timeout in seconds |
--redis-address <address> |
REDIS_ADDRESS |
http://localhost:6379 |
Redis address |
--redis-database <database> |
REDIS_DATABASE |
0 |
Redis database |
--redis-password <password> |
REDIS_PASSWORD |
- | Redis password |
--redis-prefix <prefix> |
REDIS_PREFIX |
ncp- |
Redis keys prefix |
Start NCP server.
List cached url paths.
Purge cached url paths.
Benchmark is run on Macbook Pro 15″ 2017, Intel Core i7-7700HQ.
# GOMAXPROCS=1 ncp --silent
$ go-wrk -c 100 -d 6 http://localhost:8080/tiny-tarball
Running 6s test @ http://localhost:8080/tiny-tarball
100 goroutine(s) running concurrently
70755 requests in 5.998378587s, 91.16MB read
Requests/sec: 11795.69
Transfer/sec: 15.20MB
Avg Req Time: 8.477674ms
Fastest Request: 947.743µs
Slowest Request: 815.787409ms
Number of Errors: 0
# ncp --silent
$ go-wrk -c 100 -d 6 http://localhost:8080/tiny-tarball
Running 6s test @ http://localhost:8080/tiny-tarball
100 goroutine(s) running concurrently
115674 requests in 5.98485984s, 149.04MB read
Requests/sec: 19327.77
Transfer/sec: 24.90MB
Avg Req Time: 5.173902ms
Fastest Request: 273.015µs
Slowest Request: 34.777963ms
Number of Errors: 0
NCP provides proxy
go package that can be used programmatically. Docs are available on godoc.org.
package main
import (
"net/http"
"time"
npmproxy "github.com/pkgems/npm-cache-proxy/proxy"
"github.com/go-redis/redis"
)
func main() {
// create proxy
proxy := npmproxy.Proxy{
// use redis as database
Database: npmproxy.DatabaseRedis{
// see github.com/go-redis/redis
Client: redis.NewClient(&redis.Options{
Addr: "localhost:6379",
}),
},
// reuse connections
HttpClient: &http.Client{},
}
// create and start server
proxy.Server(npmproxy.ServerOptions{
ListenAddress: "localhost:8080",
// allow fetching options dynamically on each request
GetOptions: func() (npmproxy.Options, error) {
return npmproxy.Options{
DatabasePrefix: "ncp-",
DatabaseExpiration: 1 * time.Hour,
UpstreamAddress: "https://registry.npmjs.org",
}, nil
},
}).ListenAndServe()
}