Mark Mandel

What I expect we should do it populate address immediately with Node details (since we have that), but if we don't have the `PodIP` yet, move forwards, and only populate...

I'd like to see this fixed - seems like an unnecessary delay.

In theory, you could tied k8s cilent leader to the readiness probe for the allocation service, so only one is available behind the Service at any point and time?

> The issue with the readiness probe is that we'll have a pod in unready / unhealthy, the helm install would fail after some time as it would expect a...

> It's just a wild idea and it's probably overkill for this, but we could maybe setup a crd for the service, some kind of stand-by/active service ? That would...

oooh the worker queue's could listen to pubsub rather than the channels? 🤔 that could be neat. The only other potentially really tricky part could be managing the bits that...

> even if we have a synced cache (or list (btw, the list is sorted by priority type so it could change between 2 gsa)) between the allocators, they will...

> Let's use this thread to get the batch allocator merged. So I'll push back here a little. This code works really well under _certain circumstances_, and isn't an actual...

> So if we fix this first: https://github.com/googleforgames/agones/issues/4190, we should be ok with the batch allocator too? Should they be under the same feature gate? I think so, yes! It's...

That seems like a lot of data to be storing with a CRD (over 1000 items in a list). I mean it's _handy_ to be able to do that, but...