Karl Isenberg

Crunched some of those numbers (with the watch filtering) and got a pretty flat trend line:  These numbers are with a very simple Deployment spec replicated...

New custom autoscaling logic is a bit more complicated. It uses 1Mi for every 10 objects, on top of a 128Mi base, then rounds down to 256Mi increments, with a...

Looks like I missed a critical constraint. Apparently GKE Autopilot has a minimum memory per Pod of 512 MiB, for older clusters that don't support bursting. So I won't be...

Interesting data point here. With the watch filtering and custom autoscaling, a 672MiB limit still isn't enough to prevent `TestStressManyDeployments` from failing with 1x4000 Deployments. One difference between TestStressManyDeployments and...

Ahh. ok. design problem... ~~The `status.sync.objectCount` is being updated before each apply attempt, to match the objects in the new `status.sync.commit`. But that means the `status.sync.objectCount` is decremented before the...

It required a lot of changes, but I managed to get the sync status to persist between reconciler restarts (except errors). To get it to work I refactored the Applier...

So I modified one of the tests to deploy incrementally more Deployments, pruning them all after each step. The test took 2.1hrs, so I probably wont keep it running in...

/lgtm

I'll be doing some other changes for watch filtering soon, so I'll cut a cli-utils release afterwards.

As for kapply, I think adding a global flag is probably fine, if you add it and tests to exercise it.