InternalURLs and ExternalURL

The Arvados configuration is stored at /etc/arvados/config.yml. See the Configuration reference for more detail.

The Services section lists a number of Arvados services, each with an InternalURLs and/or ExternalURL configuration key. This document explains the precise meaning of these configuration keys, and how they are used by the Arvados services.

The ExternalURL is the address where the service should be reachable by clients, both from inside and from outside the Arvados cluster. Some services do not expose an Arvados API, only Prometheus metrics. In that case, ExternalURL is not used.

The keys under InternalURLs are addresses that are used by the reverse proxy (e.g. Nginx) that fronts Arvados services. The exception is the Keepstore service, where clients connect directly to the addresses listed under InternalURLs. If a service is not fronted by a reverse proxy, e.g. when its endpoint only exposes Prometheus metrics, the intention is that metrics are collected directly from the endpoints defined in InternalURLs.

InternalURLs are also used by the service itself to figure out which address/port to listen on.

If the Arvados service lives behind a reverse proxy (e.g. Nginx), configuring the reverse proxy and the InternalURLs and ExternalURL values must be done in concert.

Overview

¹ If Controller runs on a different host than RailsAPI, the InternalURLs will need to be reachable from the host that runs Controller.
² If the reverse proxy (e.g. Nginx) does not run on the same host as the Arvados service it fronts, the InternalURLs will need to be reachable from the host that runs the reverse proxy.
³ If the Prometheus metrics are not collected from the same machine that runs the service, the InternalURLs will need to be reachable from the host that collects the metrics.

When InternalURLs do not need to be reachable from other nodes, it is most secure to use loopback addresses as InternalURLs, e.g. http://127.0.0.1:9005.

It is recommended to use a split-horizon DNS setup where the hostnames specified in ExternalURL resolve to an internal IP address from inside the Arvados cluster, and a publicly routed external IP address when resolved from outside the cluster. This simplifies firewalling and provides optimally efficient traffic routing. In a cloud environment where traffic that flows via public IP addresses is charged, using split horizon DNS can also avoid unnecessary expense.

Examples

The remainder of this document walks through a number of examples to provide more detail.

Keep-balance

Consider this section for the Keep-balance service:

      Keepbalance:
        InternalURLs:
          "http://ip-10-0-1-233.internal:9005/": {}

Keep-balance has an API endpoint, but it is only used to expose Prometheus metrics.

There is no ExternalURL key because Keep-balance does not have an Arvados API, no Arvados services need to connect to Keep-balance.

The value for InternalURLs tells the Keep-balance service to start up and listen on port 9005, if it is started on a host where ip-10-0-1-233.internal resolves to a local IP address. If Keep-balance is started on a machine where the ip-10-0-1-233.internal hostname does not resolve to a local IP address, it would refuse to start up, because it would not be able to find a local IP address to listen on.

It is also possible to use IP addresses in InternalURLs, for example:

      Keepbalance:
        InternalURLs:
          "http://127.0.0.1:9005/": {}

In this example, Keep-balance would start up and listen on port 9005 at the 127.0.0.1 IP address. Prometheus would only be able to access the Keep-balance metrics if it could reach that IP and port, e.g. if it runs on the same machine.

Finally, it is also possible to listen on all interfaces, for example:

      Keepbalance:
        InternalURLs:
          "http://0.0.0.0:9005/": {}

In this case, Keep-balance will listen on port 9005 on all IP addresses local to the machine.

Keepstore

Consider this section for the Keepstore service:

      Keepstore:
        InternalURLs:
          "http://keep0.ClusterID.example.com:25107": {}
          "http://keep1.ClusterID.example.com:25107": {}

There is no ExternalURL key because Keepstore is only accessed from inside the Arvados cluster. For access from outside, all traffic goes via Keepproxy.

When Keepstore is installed on the host where keep0.ClusterID.example.com resolves to a local IP address, it will listen on port 25107 on that IP address. Likewise on the keep1.ClusterID.example.com host. On all other systems, Keepstore will refuse to start.

Keepproxy

Consider this section for the Keepproxy service:

      Keepproxy:
        ExternalURL: https://keep.ClusterID.example.com
        InternalURLs:
          "http://localhost:25107": {}

The ExternalURL advertised is https://keep.ClusterID.example.com. The Keepproxy service will start up on localhost port 25107, however. This is possible because we also configure Nginx to terminate SSL and sit in front of the Keepproxy service:

upstream keepproxy {
  server                127.0.0.1:25107;
}

server {
  listen                  443 ssl;
  server_name             keep.ClusterID.example.com;

  proxy_connect_timeout   90s;
  proxy_read_timeout      300s;
  proxy_set_header        X-Real-IP $remote_addr;
  proxy_http_version      1.1;
  proxy_request_buffering off;
  proxy_max_temp_file_size 0;

  ssl_certificate     /YOUR/PATH/TO/cert.pem;
  ssl_certificate_key /YOUR/PATH/TO/cert.key;

  # Clients need to be able to upload blocks of data up to 64MiB in size.
  client_max_body_size    64m;

  location / {
    proxy_pass            http://keepproxy;
  }
}

If a client connects to the Keepproxy service, it will talk to Nginx which will reverse proxy the traffic to the Keepproxy service.

Workbench

Consider this section for the Workbench service:

  Workbench1:
    ExternalURL: "https://workbench.ClusterID.example.com"

The ExternalURL advertised is https://workbench.ClusterID.example.com. There is no value for InternalURLs because Workbench1 is a Rails application served by Passenger. The only client connecting to the Passenger process is the reverse proxy (e.g. Nginx), and the listening host/post is configured in its configuration:

server {
  listen       443 ssl;
  server_name  workbench.ClusterID.example.com;

  ssl_certificate     /YOUR/PATH/TO/cert.pem;
  ssl_certificate_key /YOUR/PATH/TO/cert.key;

  root /var/www/arvados-workbench/current/public;
  index  index.html;

  passenger_enabled on;
  # If you're using RVM, uncomment the line below.
  #passenger_ruby /usr/local/rvm/wrappers/default/ruby;

  # `client_max_body_size` should match the corresponding setting in
  # the API.MaxRequestSize and Controller's server's Nginx configuration.
  client_max_body_size 128m;
}

API server

Consider this section for the RailsAPI service:

      RailsAPI:
        InternalURLs:
          "http://localhost:8004": {}

There is no ExternalURL defined because the RailsAPI is not directly accessible and does not need to advertise a URL: all traffic to it flows via Controller, which is the only client that talks to it.

The RailsAPI service is also a Rails application, and its listening host/port is defined in the Nginx configuration:

server {
  # This configures the Arvados API server.  It is written using Ruby
  # on Rails and uses the Passenger application server.

  listen localhost:8004;
  server_name localhost-api;

  root /var/www/arvados-api/current/public;
  index  index.html index.htm index.php;

  passenger_enabled on;

  # If you are using RVM, uncomment the line below.
  # If you're using system ruby, leave it commented out.
  #passenger_ruby /usr/local/rvm/wrappers/default/ruby;

  # This value effectively limits the size of API objects users can
  # create, especially collections.  If you change this, you should
  # also ensure the following settings match it:
  # * `client_max_body_size` in the previous server section
  # * `API.MaxRequestSize` in config.yml
  client_max_body_size 128m;
}

So then, why is there a need to specify InternalURLs for the RailsAPI service? It is there because this is how the Controller service locates the RailsAPI service it should talk to. Since this connection is internal to the Arvados cluster, Controller uses InternalURLs to find the RailsAPI endpoint.

Controller

Consider this section for the Controller service:

  Controller:
    InternalURLs:
      "http://localhost:8003": {}
    ExternalURL: "https://ClusterID.example.com"

The ExternalURL advertised is https://ClusterID.example.com. The Controller service will start up on localhost port 8003. Nginx is configured to sit in front of the Controller service and terminates SSL:

# This is the port where nginx expects to contact arvados-controller.
upstream controller {
  server     localhost:8003  fail_timeout=10s;
}

server {
  # This configures the public https port that clients will actually connect to,
  # the request is reverse proxied to the upstream 'controller'

  listen       443 ssl;
  server_name  ClusterID.example.com;

  ssl_certificate     /YOUR/PATH/TO/cert.pem;
  ssl_certificate_key /YOUR/PATH/TO/cert.key;

  # Refer to the comment about this setting in the passenger (arvados
  # api server) section of your Nginx configuration.
  client_max_body_size 128m;

  location / {
    proxy_pass               http://controller;
    proxy_redirect           off;
    proxy_connect_timeout    90s;
    proxy_read_timeout       300s;
    proxy_max_temp_file_size 0;
    proxy_request_buffering  off;
    proxy_buffering          off;
    proxy_http_version       1.1;

    proxy_set_header      Host              $http_host;
    proxy_set_header      Upgrade           $http_upgrade;
    proxy_set_header      Connection        "upgrade";
    proxy_set_header      X-External-Client $external_client;
    proxy_set_header      X-Forwarded-For   $proxy_add_x_forwarded_for;
    proxy_set_header      X-Forwarded-Proto https;
    proxy_set_header      X-Real-IP         $remote_addr;
  }
}