[lbotos]

It's a lot easier, and dead simple to run a small single server install. When you start scaling, it gets a little harder, but our vision it to leverage K8s for that type of deployment.

(Source: work at GitLab, and I wouldn't If we had to manage the infra by hand vs. omnibus/k8s)

[meddlepal]

Any comment on how well the single server install scales before needing to invest in the more complex HA solution?

[lbotos]

Sure! This depends on your orgs "capacity for risk." We don't recommend more than ~300 on a single node just for the developer area. If you have 300 people relying on a literal SPOF (especially a dev shop) that's gonna hurt if there are any problems. That said, I'm aware of customers running a single 132GB node that serves somewhere between 3000-5000 devs. (Please do not go this far)

I'd say at the 150 or so developer mark, we should start talking about scaling (adding a second Geo node for failover at a minimum) or if you need more robust uptime requirements, going towards a more active-active distributed cluster. I use the word "distributed" and not "highly available" as we like to find the right fit for each org. Some orgs need true HA, and it takes something close to ~30 servers to do it. Others, just can benefit from some web frontends, and separating out the DB layer, and potentially traffic shaping to separate CI load / User load.