Scaling GRR within a single server¶
For 3.1.0 and later, use a systemd drop-in override to control how many copies of each component you run on each machine. This can initially be done using:
sudo systemctl edit grr-server
which creates “/etc/systemd/system/grr-server.service.d/override.conf”. You’ll want to turn this into a template file and control via puppet or similar. An example override that just runs 3 workers looks like:
[Service] ExecReload= ExecReload=/bin/systemctl --no-block reload email@example.com firstname.lastname@example.org email@example.com ExecStart= ExecStart=/bin/systemctl --no-block start firstname.lastname@example.org email@example.com firstname.lastname@example.org ExecStop= ExecStop=/bin/systemctl --no-block stop email@example.com firstname.lastname@example.org email@example.com
When starting multiple copies of the UI and the web frontend you also need to tell GRR which ports it should be using. So if you want 10 http frontends on a machine you would configure your systemd drop-in to start 10 copies and then set Frontend.port_max so that you have a range of 10 ports from Frontend.bind_port. (I.E. set Frontend.bind_port to 8080 and Frontend.port_max to 8089) You can then configure your load balancer to distribute across that port range. AdminUI.port_max works the same way for the UI.
Large Scale Deployment¶
The GRR server components should be distributed across multiple machines in any deployment where you expect to have more than a few hundred clients, or even smaller deployments if you plan on doing intensive hunting. The performance needs of the various components are discussed below, and some real-world example deployment configurations are described in the FAQ.
You should install the GRR package on all machines and use configuration management (chef, puppet etc.) to:
- Distribute the same grr-server.yaml to each machine
- Control how many of each component to run on each machine (see next section for details)
Component Performance Needs¶
- Worker: you will probably want to run more than one worker. In a large deployment where you are running numerous hunts it makes sense to run 20+ workers. As long as the datastore scales, the more workers you have the faster things get done. We previously had a config setting that forked worker processes off, but this turned out to play badly with datastore connection pools, the stats store, and monitoring ports so it was removed.
- HTTP frontend: The frontend http server can be a significant bottleneck. By default we ship with a simple http server, but this is single process, written in python which means it may have thread lock issues. To get better performance you will need to run multiple instances of the HTTP frontend behind a reverse HTTP proxy (i.e. Apache or Nginx). Assuming your datastore handles it, these should scale linearly.
- Web UI: The admin UI component is usually under light load, but you can run as many as you want for redundancy (you’ll need to run them behind Apache or Nginx to load-balance the traffic). The more concurrent GRR users you have, the more instances you need. This is also the API server, so if you intend to use the API heavily run more.