Public | Automated Build

Last pushed: 5 hours ago
Short Description
Docker container with HAProxy, transparent proxying enabled
Full Description

HAProxy container - with transparent TCP proxying

Docker container for HAProxy with transparent proxying and zero-downtime reloads

Zero-Downtime Reloads

When HAProxy reloads using its 'graceful reload' feature, there's a tiny amount of time where a number of packets might be dropped in the process. This is well documented elsewhere around the internet. This container uses the 'drop syn packets' technique to mitigate that. There are more sophisticated techniques available which lead to lower delays on a restart. If you'd like to implement one of those (for example, a variation of the Yelp qdisc technique that works for incoming traffic or the unbounce IP tableflip technique) in this container, feel free to fork. Pull requests gratefully received.

By default, the container assumes HAProxy will be listening on ports 80 & 443. If you're using HAProxy on different ports, you can alter the reload process using this environment variable when you run the container:

docker run [your other options] -e "HAPROXY_PORTS=1234,5678" [capabilities options defined below] tombull/haproxy

Transparent proxying

You can normally proxy HTTP and HTTPS transparently (appearing as if requests come from the original requesting IP address) ONLY if you terminate the request with HAProxy and then pass the request on with extra proxy headers. What about transparent TCP proxying? Where you just pass the request straight through to your web (or other TCP) server? Well HAProxy can do that. But it's a bit difficult with docker - you need some extra security capabilities.

Run the container with extra capabilities and host networking

I've tried very hard to minimise the extra capabilities this container needs to function - but these are the absolute minimum unless you want to set some of the extra host settings outside the container.

docker run [your other options] --volume=/proc/sys/net/ipv4/ip_nonlocal_bind:/var/proc/sys/net/ipv4/ip_nonlocal_bind --net=host --cap-add=SYS_MODULE --cap-add=NET_ADMIN --cap-add=NET_RAW tombull/haproxy

You can remove the --volume directive if you set the following sysctl settings on the host:

net.ipv4.ip_nonlocal_bind = 1

You can remove the --cap-add=SYS_MODULE if you ensure the following kernel module is enabled on the host:

xt_TPROXY

Running the container with --net=host and the capabilities NET_ADMIN and NET_RAW is completely unavoidable.

Final Note: If you don't want transparent TCP proxying, and you just want graceful reloading, in theory you can just use the NET_ADMIN capability and nothing else. At best, you will get some odd but harmless iptables rules cropping up somewhere, at worst it's just going to break in weird and horrible ways. I've not tested this, so I don't know what exactly will happen!

When stopping the container

Always give the container enough time to shutdown. Otherwise you're going to be left with iptables rules and other stuff on your host system that you don't necessarily want.

docker stop --time=300 your_container_name

The time is in seconds and should be a little longer than your haproxy timeouts as set in your haproxy.cnf. Suggested: 300 seconds because there's a hard-coded timeout of 295 seconds in the container. The container will stop and remove iptables rules, etc before the end of the 300 seconds, guaranteed.

Docker Pull Command
Owner
tombull
Source Repository