uds-proxy provides a UNIX domain socket and forwards traffic to HTTP(S) remotes through a customizable connection pool (i.e. using persistent connections).
Interacting with microservices often involves communication overhead: Every contact with another service may involve DNS lookups and establishment of a TCP connection plus, most likely, a HTTPS handshake.
This overhead can be costly and especially hard to circumvent for legacy applications -- thus uds-proxy.
uds-proxy creates a UNIX domain socket and forwards communication to one or more remote web servers. In a way, uds-proxy aims a bit at reducing application/API complexity by providing a generic and simple solution for connection pooling.
uds-proxy is implemented in Go, so it runs as native application on any OS supporting Go and UNIX domain sockets (i.e. not on Windows). Critical performance metrics of uds-proxy (request latencies, response codes...) and Go process statistics are exposed through Prometheus client library.
Building requires a local Go 1.11+ installation:
go get -v github.com/schnoddelbotz/uds-proxy/cmd/uds-proxy
... or just grab a uds-proxy binary release.
See usage-example-for-an-https-endpoint for Docker usage.
To start uds-proxy at system boot, create e.g. a systemd unit. Don't try to run uds-proxy as root. It won't start.
Usage of ./uds-proxy:
-client-timeout int
http client connection timeout [ms] for proxy requests (default 5000)
-idle-timeout int
connection timeout [ms] for idle backend connections (default 90000)
-max-conns-per-host int
maximum number of connections per backend host (default 20)
-max-idle-conns int
maximum number of idle HTTP(S) connections (default 100)
-max-idle-conns-per-host int
maximum number of idle conns per backend (default 25)
-no-access-log
disable proxy access logging
-no-log-timestamps
disable timestamps in log messages
-pid-file string
pid file to use, none if empty
-prometheus-port string
Prometheus monitoring port, e.g. :18080
-remote-https
remote uses https://
-socket string
path of socket to create
-socket-read-timeout int
read timeout [ms] for -socket (default 5000)
-socket-write-timeout int
write timeout [ms] for -socket (default 5000)
-version
print uds-proxy version
Clone this repository and check the Makefile targets.
Most relevant make
targets:
make monitoring_test
spins up Prometheus, grafana and uds-proxy using Docker and starts another uds-proxy instance locally (outside Docker, on Mac only). The uds-proxy instances will be scraped by dockerized Prometheus and Grafana will provide dashboards. See monitoring/README.md for details.make run_proxy
starts a local uds-proxy instance for testing purposes.TEST_SOCKET
environment variable controls socket location, defaults touds-proxy-test.socket
.make test
runs unit and functional tests from proxy_test directory.make coverage
generates code test coverage statistics.make test_integration
starts a local uds-proxy and runs some proxied-vs-non-proxied perf tests.make realclean
removes leftovers from tests or builds.
Start the proxy:
uds-proxy -socket /tmp/proxied-svc.sock -prometheus-port :28080 -remote-https
Docker users:
mkdir -p /tmp/mysock_dir
docker run --rm -it -p28080:28080 -v/tmp/mysock_dir:/tmp schnoddelbotz/uds-proxy
For both cases, metrics should be available at http://localhost:28080/metrics while uds-proxy is running.
# without uds-proxy, you would...
time curl -I https://www.google.com/
# with uds-proxy, always ...
# a) talk through socket and
# b) use http:// and let `-remote-https` ensure https is used to connect to remote hosts
time curl -I --unix-socket /tmp/proxied-svc.sock http://www.google.com/
# ... or using socket provided by dockerized uds-proxy:
time curl -I --unix-socket /tmp/mysock_dir/uds-proxy-docker.sock http://www.google.com/
<?php
// without uds-proxy
$ch = curl_init();
curl_setopt($ch, CURLOPT_URL, "https://www.google.com/");
curl_exec($ch);
// with uds-proxy
$ch = curl_init();
curl_setopt($ch, CURLOPT_URL, "http://www.google.com/");
curl_setopt($ch, CURLOPT_UNIX_SOCKET_PATH, "/tmp/proxied-svc.sock");
curl_exec($ch);
Mac's (i.e. BSD's) netcat allows to talk to unix domain sockets.
It can be used to e.g. ensure correct behaviour of uds-proxy's
-socket-(read|write)-timeout
options. Try nc -U /path/to/uds-proxy.sock
.
- fix/drop sudo nobody for dockerized tests
- fixme: add option -dont-follow-redirects
- for http/s client:
- wrap in circuit breaker?
- wrap in retry /w exponential backoff? consider api consumer constraints (i.e. timeout - worth it?)
- travis-ci + github release push
- example systemd unit
- sock umask / cli opt
- support magic uds request headers...?
- X-udsproxy-timeout: 250ms
- X-udsproxy-debug: true
- https://godoc.org/gotest.tools/assert
- https://golang.org/pkg/net/#hdr-Name_Resolution
- https://stackoverflow.com/questions/17948827/reusing-http-connections-in-golang
- https://medium.com/@povilasve/go-advanced-tips-tricks-a872503ac859
- https://github.com/bouk/monkey/blob/master/monkey_test.go
- https://github.com/prometheus/client_golang/blob/master/prometheus/examples_test.go
- https://github.com/prometheus/client_golang/blob/master/prometheus/promhttp/instrument_server.go
Obviously, uds-proxy is a kludge. Simply use connection pooling if available!
- for Python and HTTP, simply reuse requests library's session objects and you're set
- for Python and Redis, use a redis.py connection pool
- for Redis and PHP, phpredis supports connection pooling since v4.2.1
- a potentially more sophisticated solution can be found in this TCP vs UDS speed comparison stackoverflow thread: Speedus intercepts relevant system calls, which avoids need for any code changes. However, if I understood correctly, Speedus only helps if services actually sit on the same host system (?).
Maybe look at phantom?
You can also use NGINX to create a UDS HTTP/S pooling forward proxy like uds-proxy. It seems that neither Apache nor Squid (?) are able to do that.
MIT