Because of defaulting, requests are always set if limits are. Thus, the check can never succeed. Instead, make sure that the two values are equal.
Also, remove a few other error messages and remove unnecessary Sprintf calls.
Automatic merge from submit-queue
kube-up: install new Docker pre-requisite (libltdl7) when not in image
Docker now has a dependency on libltdl7; we have to specify it manually
if we are installing docker using dpkg (vs using apt-get or similar,
which would pull it in automatically)
Fixes#28644
Automatic merge from submit-queue
Separate rate limiters for Pod evictions for different zones in NodeController
Ref. #28832
NodeController needs to be able to separately adjust rate-limits for eviction for different zones. This PR splits rate limiters.
cc @davidopp
Automatic merge from submit-queue
Implement first version of indexing in cacher
This PR adds a first version of indexing in cacher.
It has a really significant impact on performance - __in empty 2000-node cluster, apiserver cpu usage drops by ~75%.__
Not for 1.3, but we need this soon after 1.3 is done.
@lavalamp @mqliang @davidopp @gmarek @kubernetes/sig-scalability
Automatic merge from submit-queue
Don't check in generated code, part 1
This PR is a first step towards not commiting generated files, which make up a huge portion of "needs rebase" errors. It only handles deep-copy generation and conversion generation. More will come later, if the model passes muster.
This is a mega-PR. Sorry. It was necessary to do 2 generators to convince myself it worked, and the evolution of the techniques warranted multiple commits. I have tried to keep the commits self-contained and reviewable.
A quick summary of the major points in the series:
- Start by making everything call `make` rather than the various hack/* scripts. The hack scripts still exist, but give a warning to use make instead, and then they do what they did before, so it should be compatible.
- Move deepcopy generation into the Makefile, so it is done automatically
- Move conversion generation into the Makefile, so it is done automatically
- Optimize makefile for faster rebuilds
- Make CI pass
Net result: if you run "make", it will rebuild any deepcopy or conversion files it needs. It takes a few seconds to figure out there's nothing to do, but it should be a net savings. There is more to do, and we can follow this up with other generators being converted, some of which are MUCH slower than these 2.
@wojtek-t @lavalamp @smarterclayton @bgrant0607 @mikedanese @madhusudancs
Automatic merge from submit-queue
check PodAntiAffinity even when pod has no AntiAffinity constraints
Hard inter-pod anti-affinity is designed to be symmetric, in other word, the algorithm should check if AntiAffinity constraints from both the pod to be scheduled and pods already running are satisfied during scheduling.
Current implementation skipped the checking when pod has no AntiAffinity constraints, ignored the anti-affinity constraints of existing pods , which is not correct. This PR is to fix it.
This is sort of gravy. There's no good way to do this for arbitrary
directories because of the way Make works (we'd have to declare them all
PHONY).
If people hate using make directly we could wrap it in another level of
indirection, but I didn't yet. E.g.
build/make-all.sh -> make -> make-rules/all.sh -> go install
This would insulate people from 'make' atthe cost of custom build tools. I
prefer being able to say 'make' and having it just work.
This mostly takes the previously checked in files and removes them, and moves
the generation to be on-demand instead of manual. Manually verified no change
in generated output.
This allows us to start building real dependencies into Makefile.
Leave old hack/* scripts in place but advise to use 'make'. There are a few
rules that call things like 'go run' or 'build/*' that I left as-is for now.
Automatic merge from submit-queue
Deepcopy: avoid struct copies and reflection Call
- make signature of generated deepcopy methods symmetric with `in *type, out *type`, avoiding copies of big structs on the stack
- switch to `in interface{}, out interface{}` which allows us to call them with without `reflect.Call`
The first change reduces runtime of BenchmarkPodCopy-4 from `> 3500ns` to around `2300ns`.
The second change reduces runtime to around `1900ns`.
