I'm using haproxy as a bastion server / cluster gateway, as only some of the nodes in my network have direct access to the external network. My internal nodes are part of a kubernetes cluster, and need to be able to pull images from a private registry external to my cluster which requires certificate identities.
k8s cluster internal node -> haproxy on edge node -> docker registry
I'm trying to configure my back_end in my haproxy.cfg to route to the docker registry, and update the request with the certificate identity of the edge node. (Internal node does not have certs acceptable to the docker registry, and I've not been allowed to host the external node's certs on the internal node.) What I have right now looks like the below...
frontend ft_ssl
bind <boxIP>:443
mode http
default_backend bk_global_import_registry_certs
backend bk_global_import_registry_certs
mode http
balance roundrobin
server registry_alias-0 <registryIP>:443 ssl ca-file fullyqualified/ca.crt crt fullyqualified/file.pem check
server registry_alias-1 <registryIP2>:443 ssl ca-file fullyqualified/ca.crt crt fullyqualified/file.pem checkI currently have the HTTPS_PROXY setting in /etc/systemd/system/docker.service.d/http-proxy.conf and am getting a 400 Bad Request. Scrubbed log message below, with only changes as removal of IPs or typos.
InternalIP:randomPort [09/Jul/2019:13:28:08.659] ft_ssl bk_global_import_registry_certs 0/0/10/5/15 400 350 - - ---- 1/1/0/0/0 0/0 {} "CONNECT externalFQDN:443 HTTP/1.1"
For those looking at this via the kubernetes or docker tags, I also considered setting up a pull-through cache, but realized this only works with Docker's public registry - see open Docker GitHub issue 1431 for other folks trying to find ways to get past that, as well.
I might be mistaking but as far as I know Haproxy cannot be used as a direct HTTP proxy, it only acts as a reverse HTTP proxy. You should use Squid or something like that to be a direct HTTP proxy.
If you want Haproxy to terminate SSL for you like you do, then you will need to change hostname part of your Docker image to be the hostname of your Haproxy node. You should also make sure your Docker daemons trust the Haproxy certificate or add Haproxy to the list of INSECURE_REGISTRIES on all Kube nodes.
Posting the answer that resolved the situation for us, in case it helps others...
All traffic from the internal node now routes through to the HAProxy.
I'm no longer using the HTTPS_PROXY setting in /etc/systemd/system/docker.service.d/http-proxy.conf. Proxying through was not appropriate, since I could not use my internal node's certificate to authenticate against the docker registry.
From the internal node, I'm now treating the docker registry as an insecure registry by adding its address to /etc/docker/daemon.json in insecure-registries: [].
We can now access our internal private registry using the certificates passed along on the forward on the HAProxy backend. We do verify the certificate on the front-end, but because we have the registry listed in insecure-registries, Docker accepts the cert name mismatch.
One side effect noticed: pulling an image from "default" docker registry without specifying a prefix does not succeed using our solution. To pull from Docker, for example, you'd pull from registry-1.docker.io/imageName:imageVersion, rather than just imageName:imageVersion.