Daniel Smith

You cannot in the abstract. I'm only seeking to solve this problem for people running a watch & a cache. (In fact it's almost a pro for option 2!)

> The linked issue also only attempts to answer the intraobject problem. Right, and this is a defect that needs to be fixed before we could merge it, see why...

Thanks @mm4tt, yeah that would make it super easy, I think.

> Is there an option 3? Could we define resource version as an unsized unsigned integer? Unfortunately this would probably also break most or all clients parsing it, should we...

> in the future a safely opaque RV can be mixed with the cluster ID of the etcd state, allowing client operations (watch, list, precondition edit) to be invalidated during...

The problem with a discontinuous switch over is that you don't really have a good way of knowing if clients have been updated or not. I prefer a 2nd field...

/lgtm /approve I think the benchmark appearing in the comments here is sufficient, thanks for making it, Alex!

We should never be adding to user-controlled traces, IMO. apiserver should add baggage to its trace headers. Rentrant calls should repeat the baggage. Then the 2nd apiserver can use this...

I don't see how webhooks could attach a trace. Letting users request apiserver initiate a trace is different than attaching a request to an already created trace, no? If we...

Maybe a preexisting trace makes sense if folks are using api calls. I guess my concern is just bounding the resource usage. A non-authz solution might look like a cap...