|
|
|
|
|
|
by mercutio2
2940 days ago
|
|
|
Generally speaking, I have found Atlassian’s tools don’t scale well to megacorp usage. When there are thousands to tens of thousands of users connecting at peak times, stuff just gets extremely slow. Noticeably slower than when we all put up our own team-level Phabricator instances, although I’m obviously not adjusting for Phabricator downtime when inevitably the unofficial team sysadmin went on vacation and the server fell over. I’m sure people involved in infrastructure are working on making this better, but I don’t have any visibility into that layer. |
|
|
For thousands to tens of thousands of users, most folks need data center with 1+n nodes (w/ larger diminishing returns at 8 nodes). There's a few tips and tricks with that, like pointing CI/CD tools to a smart mirror instead of the primary instance and only mirroring specific build projects. Putting smart mirrors in the same physical location as a primary instance could also help reduce the impact of CPU-intensive requests like git fetch or clone.
Another common issue is apps/plugins. Bitbucket 5.9 added a way to get app diagnostics so infra folks can see if there's a long-running process or other weirdness. Vanilla Bitbucket needs about a 1gb heap (java) and with apps, you want around 4gb (give or take).
If you want a longer read to share, this still has some good info despite being 2~ years old: https://developer.atlassian.com/blog/2016/12/how-we-built-bi...