plugin,etc: Rewrite to get state from Pod annotations #1163

sharnoff · 2024-12-03T06:39:10Z

Note

Hey! This is a big PR — sorry! There's a lot of smaller commits (see below), plus one GIANT commit that does the bulk of the rewrite, building on those smaller commits.

There's an overview of the changes in the commit message for the full rewrite. That might be a useful place to start to get your bearings on the changes.

Commits broken down by theme:

Background work: neonvm

neonvm-controller: Use non-controller owner ref for migration source (d6fa7d0)
In short, there's currently no reliable way to tell whether a Pod is the source or target in a live migration without checking against the migration object itself.

This is because the migration object initially has a unique "controller" reference on the target pod (with the source pod having a non-controlling owner reference), and once the migration is complete, the source pod is changed to have the unique controller reference setup.

So this change is just to keep using a non-controller reference on the source pod after the migration completes.

That means that we will have the following guarantees:
- If a Pod has a "controller" owner reference for a live migration, it's the target pod in an ongoing migration.
- If a Pod has a non-controller owner reference for a live migration, it's the source pod in a maybe-ongoing migration.
- If the Pod has a "controller" owner reference for a virtual machine, it's the current source of the VM. So if a migration source pod doesn't have an owner reference for the VM, then the migration has completed and there is no longer a use for the old source pod.
neonvm: Add helpers to get pod ownership (e77ac86)

We had some existing, similar helpers in pkg/util, but I figured it makes more sense for those to be defined in neonvm/apis/....

This change originated while working on the stateless scheduler prototype, where I found myself wanting a reliable way to determine what role a pod has in a live migration from only the pod's metadata.
neonvm-controller: Update runner pod metadata while pending (a1b55f7)

We weren't doing this previously, and it means that propagation of labels/annotations can be arbitrarily delayed while the VM starts.

Related changes in util/watch

util/watch: Store HandlerFuncs[*T] in Store[T] (ad2340f)

Also, change handleEvent[T any, P ~*T]() from a function to a method on *Store[T], now that we have the handlers translated from P → *T.

This opens up a lot of ways to make the code cleaner, and the handlers part in particular is required to implement (*Store[T]).NopUpdate() in a later commit.
util/watch: Add (*Store[T]).NopUpdate() method (1f43c44)

There's some info about this in the added comment.

tl;dr: We need it for the stateless scheduler work to be able to re-inject items into the reconcile queue while maintaining the watch as the source of truth for what each object is.
util/watch: Override GVK on incoming objects (2580422)

The K8s API server and client-go together have the behavior that objects returned from List() calls do not have TypeMeta set.

For one-off List() requests this is fine becuase you already know the type! But this interacts poorly with the generated implementations of objects' .GetObjectKind().GroupVersionKind(), as those just directly read from the TypeMeta fields (which again: are not set).

So this commit works around this behavior by getting the GVK at the start of the Watch() call and explicitly setting it on all incoming objects.
util/watch: Add (*Store[T]).Listen() method (341c0b5)

The Listen() method returns a util.BroadcastReceiver that will be updated whenever the object is modified or deleted.

This is required for now for the stateless scheduler work, so that we can be separately notified when there's changes to an object without hooking deeper into the internal state.

We can probably remove this once the scheduler plugin's agent request handler server is removed, at the end of the stateless scheduler work.

The one big commit doing the rewrite

plugin: Rewrite to get state from Pod annotations (d04ce1b)
a.k.a. "Stateless Scheduler".

This is effectively a full rewrite of the scheduler plugin. At a high level, the existing external interfaces are preserved:
- The scheduler plugin still exposes an HTTP server for the autoscaler-agent (for now); and
- The scheduler plugin is still a plugin.
However, instead of storing the state for approved resources in-memory, in the scheduler plugin, we now treat annotations on the Pod as the source of truth for requested/approved resources.

A brief overview of the internal changes to make this happen:
1. The state of resource reservations can be constructed entirely from Node and Pod objects. We do store that, and update as objects change, but it's only for convenience and not a strict requirement.
  
  One tricky piece is with scheduling. For that, we store a set of pods that have gone through the plugin methods but haven't actually had the spec.nodeName field set.
  
  For more info, the pkg/plugin/state package contains all the pure logic for manipulating resources.
2. Each watch event on Node/Pod objects is now placed into a "reconcile" queue similar to the controller framework. Reconcile operations are a tuple of (object, event type, desired reconcile time) and are retried with backoff on error/panic.
  
  For a detailed look, the 'pkg/plugin/reconcile' package defines the reconcile queue and all its related machinery.
3. The handler for autoscaler-agent requests no longer accesses the internal state and instead directly patches the VirtualMachine object to set the annotation for requested resources, and then waits for that object to be updated.
  
  Once the autoscaler-agent is converted to read and write those annotations directly, we will remove the HTTP server.
4. pkg/util/watch was changed to allow asking to be notified when
  there's changes to an object, via the new (*Store[T]).Listen() API.
  
  This was required to implement (3), and can be removed once (3) is no longer needed, if it doesn't become used in the autoscaler-agent.
5. pkg/util/watch was changed to allow triggering no-op update events, which - for our usage - will trigger requeuing the object. This solves two problems:
  1. During initial startup, we need to defer resource approval until all Pods on the Node have been processed -- otherwise, we may end up unintentionally overcommitting resources based on partial information.
    
    So during startup, we track the set of Pods with deferred approvals, and then requeue them all once startup is over by triggering no-op update events in the watch store.
  2. Whenever we handle changes for some Pod, it's worthwhile to handle certain operations on the Node -- e.g., triggering live migration if the reserved resources are too high.
    
    While we could do this as part of the Pod reconciling, we get more fair behavior (and, better balancing under load) by instead triggering re-reconciling the Pod's Node.
  Why can't this be done elsewhere? In short, consistency. Fundamentally we need to use a consistent view of the object that we're reconciling (else, it might not be no-op), and the source of truth for the current value of an object within the scheduler plugin is the watch store.

Remaining TODOs:

Flesh out the reconcile metrics (3693980)
Add metrics for k8s CRUD operations (mainly: rate of VirtualMachine update operations) (3693980)
Test node metrics (40b43c1)
Test that deferred reconciles on startup work correctly
Test that setting VM CPU/memory use fields and enabling autoscaling at the same time works correctly.
Load testing on staging (link)

Open questions:

Which commits should be squashed, or extracted into separate PRs?
We have a single global state mutex (same as before, but different usage pattern). How big a cluster can we support before that falls over? Do we need to have individual state locks per node?

github-actions · 2024-12-03T06:42:35Z

No changes to the coverage.

HTML Report

Click to open

Probably an inadvertent merge conflict between #1090 and #989 meaning we accidentally weren't using go-chef for neonvm-daemon. Noticed this while working on #1163 locally and saw that it was re-downloading all of the dependencies for neonvm-daemon every time, even though I was making changes in the scheduler and the dependencies hadn't changed.

It wasn't correct; the separator is '/', not ':'. (I think once upon a time, we used to format it with ':', but that's no longer the case). Noticed this as part of #1163.

sharnoff · 2024-12-13T17:17:22Z

Squashed 06240dd..0b64f08 (diff)
Rebased onto main: 0b64f08..6c76e9d (diff)

Omrigan

Partial review, got to the middle of handle_pod.

Omrigan · 2024-12-12T11:09:18Z

neonvm/apis/neonvm/v1/pod_helpers.go

+	return extractFromAnnotation[VirtualMachineResources](pod, VirtualMachineResourcesAnnotation)
+}
+
+func extractFromAnnotation[T any](pod *corev1.Pod, annotation string) (*T, error) {


Is this one the same as extractAnnotationJSON, let's maybe deduplicate?

Hm, if we put it in pkg/util, we get an import cycle, but I don't think it belongs in either pkg/api or neonvm/apis/neonvm/v1...

I'm inclined to leave it as-is for now, but LMK what you think

Omrigan · 2024-12-12T13:43:48Z

pkg/plugin/reconcile/event.go

+	// EventKindEphemeral represents the combination when an addition was not handled before the
+	// object was deleted, yet we still may want to process either of these.


Why we might need to handle EventKindEphemeral?

pkg/neonvm/controllers/vmmigration_controller.go

Omrigan · 2024-12-16T13:19:27Z

pkg/plugin/globalstate.go

+	requeuePod      func(uid types.UID) error
+	requeueNode     func(nodeName string) error


Why do we have different types of ids for pods and nodes? And how those are related to NamespacedName?

To answer what these IDs are, in short:

Each object in your cluster has a Name that is unique for that type of resource. Every Kubernetes object also has a UID that is unique across your whole cluster.

Some objects are not namespaced (in which case their name is just a string); some objects are (in which case their name is a NamespacedName).

As for why these two requeueFoo functions take two different IDs:

Pods can be trivially recreated with the same name. So in this PR, we always include the UID whenever dealing with a Pod object, and only ever store maps of them keyed by UID.

while Nodes can have duplicate names at different times, we also have to deal with the fact that Pods only refer to nodes by name, not UID — see the .spec.nodeName field. We already need to have some way of doing node name → node state lookups to handle pod events, so for consistency this PR only ever refers to nodes by name rather than by UID.

pkg/plugin/globalstate.go

Omrigan · 2024-12-16T13:35:47Z

pkg/plugin/handle_pod.go

+	var ok bool
+	ns, ok = s.nodes[nodeName]
+	if !ok {
+		return fmt.Errorf("pod's node %q is not present in local state", nodeName)


As discussed, maybe there is no reason to fail here, we can still accept the pod even we not yet have seen the node.

TBH that seems like it'd add a lot of complexity that I'm not sure is worth it — adding a pod potentially has to add a placeholder node but do nothing else, and deleting a pod potentially has to also delete the node it's on, if the node was already deleted and the pod is the last one.

I'm OK with complexity encapsulated into the state package because that's easy to unit-test, but I don't currently see a clean interface boundary we can draw here without unit-testing the entire global state (which is maybe fine, but a lot of effort! that's what the state package is meant to help with)

Maybe I'm missing something? WDYT?

pkg/plugin/handle_pod.go

petuhovskiy · 2024-12-20T11:54:21Z

Which commits should be squashed, or extracted into separate PRs?

The first 7 commits look quite isolated from the big "stateless scheduler" commit. They also LGTM, so you can create a PR out of these commits and merge them to main, if you want.

Omrigan · 2024-12-17T21:11:56Z

pkg/plugin/reconcile/middleware.go

+	Call(*zap.Logger, MiddlewareParams, MiddlewareHandlerFunc) (Result, error)
+}
+
+// MiddlewareHandlerFunc is an enriched version of HandlerFunc that accepts more parameters, so that


Do we really need this distinction between HandlerFunc and MiddlewareHandlerFunc?
Why not have MiddlewareParams be named smth like ObjectParams?

Mostly, I figured it was cleaner to have a simpler HandlerFunc — MiddlewareParams exists because it's easier for all the middleware to have access to common object attributes (e.g., GVK) than trying fallible extraction in each one, whereas the end handler probably doesn't need those, so it'd just be extra unused parameters.

Omrigan · 2024-12-18T15:03:14Z

pkg/plugin/reconcile/middleware.go

+
+// Middleware wraps a reconcile operation to insert its own logic
+type Middleware interface {
+	Call(*zap.Logger, MiddlewareParams, MiddlewareHandlerFunc) (Result, error)


Why handler func is passed here? It could instead be passed when a particular middleware is constructed, thus applyMiddleware is not needed.

The reason I didn't go with that approach is because storing the handler in the middleware requires either:

Putting together all the middleware with an explicit ordering of how we construct them, which risks making the code harder to refactor (because you can't "just" move things around, due to the web of dependencies); OR

Collecting the set of middleware as a set of func(handler) Middleware to be assembled at the end

I figured that providing the handler func in the call to middleware allows it to be more isolated, which -- for me, at least -- reduces the number of things I need to hold in my head when thinking about how it works.

Does that make sense? If so, WDYT?

Also worth noting that the approach you mentioned requires a single handler, rather than the distinct handler for each type that the queue currently receives as input -- discussed that here: #1163 (comment)

Omrigan · 2024-12-18T15:11:10Z

pkg/plugin/reconcile/queue.go

+
+// NewQueue builds and returns a new Queue with the provided middleware and handlers for various
+// types.
+func NewQueue(handlers map[Object]HandlerFunc, opts ...QueueOption) (*Queue, error) {


Why Queue is initialized with a map of handlers per object type, instead of just one handler, which would do the dispatch by itself?

I did it that way because it allows greater validation, in a few places:

We always know that we're seeing all possible types during construction, so if e.g. there's an issue extracting the GVK from the object (potentially due to unregistered/misregistered scheme), that'll fail on startup, rather than later on when we try to handle it

If we don't have handling for a particular object type, that'll fail at the point where it's added to the queue, rather than in the handler (which may just repeat endlessly)

Having the set of types known at construction allows builtin middleware to be aware of that — in particular, the ErrorBackoffMiddleware uses this information to pre-create separate maps for each type, instead of needing to lock a global object each time.

Omrigan · 2024-12-18T15:12:08Z

pkg/plugin/reconcile/queue.go

+
+		// Check that this isn't a duplicate
+		if _, ok := handlersByType[gvk]; ok {
+			return nil, fmt.Errorf("Duplicate handler for object type %T with GVK %q", obj, fmtGVK(gvk))


Capitalized error string. Why linter doesn't complain?

Presumably staticcheck style lints aren't enabled? The one you're referring to seems to be ST1005.

Across the rest of the codebase, of the calls to fmt.Errorf that start with a letter, about 1/3 are capitalized and 2/3 are lowercase. I'm happy to make that change separately, but it's broader than just this PR.

pkg/neonvm/controllers/vmmigration_controller.go

Omrigan · 2024-12-18T15:23:55Z

pkg/plugin/reconcile/options.go

+	apply func(*queueSettings)
+}
+
+// queueSettings is the internal, temporary structure that we use to hold the results of applying


Instead of supplying middleware-related callbacks through options, did you consider instead instantiating middlewares before initializing queue, providing those callbacks to the middleware directly. Something like:

logMiddleware := reconcile.NewLogMiddleware(handler, func(..) { // This is the callback }) panicMiddleware := reconcile.NewPanicMiddleware(logMiddleware, func(...) {}) finalHandler := panicMiddleware queue := reconcile.NewQueue(finalHandler, ...) ...

I considered doing the construction of the default middleware outside reconcile.NewQueue, but ultimately decided not to, because:

These particular pieces of middleware (logging, error backoff, panic handling) should always be included — it would be an error to not include them, because the behavior of the queue would be open to easily preventable risks (busy-looping on an error, panicking taking down the process, etc.). So, it's probably better from an API perspective to make sure that they must always be included!

The functions that they can be configured with are optional — e.g., they may only be used for incrementing metrics, so IMO it's more natural to have the API express that these are optional, even if in practice we always use them

Does that make sense? WDYT?

petuhovskiy

Finished my first reading, overall the code looks very good. Will come back for a second iteration next year :)

Made some small refactorings along the way, mostly it is about moving metrics-related code to a new package: https://github.com/neondatabase/autoscaling/pull/1182/commits

The changes are described in commit messages, feel free to accept/reject.

petuhovskiy · 2024-12-24T20:19:39Z

pkg/plugin/reconcile/middleware.go

+	maxErrorWait     = time.Minute
+)
+
+func NewErrorBackoffMiddleware(typs []schema.GroupVersionKind, callback ErrorStatsCallback) *ErrorBackoffMiddleware {


suggestion: This middleware is quite complicated. Can you please add a comment explaining what it does and why?

pkg/plugin/state/xactmap.go

pkg/plugin/state/node.go

pkg/plugin/reconcile/event.go

pkg/plugin/handle_node.go

Extracted from #1163, where `gci` incorrectly groups the "iter" import away from other stdlib pacakges, because it was only added in Go 1.23.

From now on, during the migration we will have: - Symmetrical owner refs: both VirtualMachine and VirtualMachineMigration will own both source and target pods - Asymmetrical controller refs: VirtualMachine controls source pod, VirtualMachineMigration controls target pod. This allows us to figure out the pod's role in the migration by looking at it's references. Co-authored-by: Oleg Vasilev <[email protected]>

We had some existing, similar helpers in 'pkg/util', but I figured it makes more sense for those to be defined in 'neonvm/apis/...'. This change originated while working on the stateless scheduler prototype, where I found myself wanting a reliable way to determine what role a pod has in a live migration from only the pod's metadata.

We weren't doing this previously, and it means that propagation of labels/annotations can be arbitrarily delayed while the VM starts.

Also, change handleEvent[T any, P ~*T]() from a function to a method on *Store[T], now that we have the handlers translated from P -> *T. This opens up a lot of ways to make the code cleaner, and the handlers part in particular is required to implement (*Store[T]).NopUpdate() in a later commit.

There's some info about this in the added comment. tl;dr: We need it for the stateless scheduler work to be able to re-inject items into the reconcile queue while maintaining the watch as the source of truth for what each object is.

The K8s API server and client-go together have the behavior that objects returned from List() calls do not have TypeMeta set. For one-off List() requests this is fine becuase you already know the type! But this interacts poorly with the generated implementations of objects' .GetObjectKind().GroupVersionKind(), as those just directly read from the TypeMeta fields (which again: are not set). So this commit works around this behavior by getting the GVK at the start of the Watch() call and explicitly setting it on all incoming objects.

The Listen() method returns a util.BroadcastReceiver that will be updated whenever the object is modified or deleted. This is required *for now* for the stateless scheduler work, so that we can be separately notified when there's changes to an object without hooking deeper into the internal state. We can probably remove this once the scheduler plugin's agnent request handler server is removed, at the end of the stateless scheduler work.

a.k.a. "Stateless Scheduler". This is effectively a full rewrite of the scheduler plugin. At a high level, the existing external interfaces are preserved: - The scheduler plugin still exposes an HTTP server for the autoscaler-agent (for now); and - The scheduler plugin is still a plugin. However, instead of storing the state for approved resources in-memory, in the scheduler plugin, we now treat annotations on the Pod as the source of truth for requested/approved resources. A brief overview of the internal changes to make this happen: 1. The state of resource reservations can be constructed entirely from Node and Pod objects. We *do* store that, and update as objects change, but it's only for convenience and not a strict requirement. One tricky piece is with scheduling. For that, we store a set of pods that have gone through the plugin methods but haven't actually had the spec.nodeName field set. For more info, the 'pkg/plugin/state' package contains all the pure logic for manipulating resources. 2. Each watch event on Node/Pod objects is now placed into a "reconcile" queue similar to the controller framework. Reconcile operations are a tuple of (object, event type, desired reconcile time) and are retried with backoff on error/panic. For a detailed look, the 'pkg/plugin/reconcile' package defines the reconcile queue and all its related machinery. 3. The handler for autoscaler-agent requests no longer accesses the internal state and instead directly patches the VirtualMachine object to set the annotation for requested resources, and then waits for that object to be updated. Once the autoscaler-agent is converted to read and write those annotations directly, we will remove the HTTP server. 4. 'pkg/util/watch' was changed to allow asking to be notified when there's changes to an object, via the new (*Store[T]).Listen() API. This was required to implement (3), and can be removed once (3) is no longer needed, if it doesn't become used in the autoscaler-agent. 5. 'pkg/util/watch' was changed to allow triggering no-op update events, which - for our usage - will trigger requeuing the object. This solves two problems: a. During initial startup, we need to defer resource approval until all Pods on the Node have been processed -- otherwise, we may end up unintentionally overcommitting resources based on partial information. So during startup, we track the set of Pods with deferred approvals, and then requeue them all once startup is over by triggering no-op update events in the watch store. b. Whenever we handle changes for some Pod, it's worthwhile to handle certain operations on the Node -- e.g., triggering live migration if the reserved resources are too high. While we *could* do this as part of the Pod reconciling, we get more fair behavior (and, better balancing under load) by instead triggering re-reconciling the Pod's Node. Why can't this be done elsewhere? In short, consistency. Fundamentally we need to use a consistent view of the object that we're reconciling (else, it might not be no-op), and the source of truth for the current value of an object *within the scheduler plugin* is the watch store. Co-authored-by: Arthur Petukhovsky <[email protected]>

sharnoff · 2025-01-02T21:33:42Z

Squashed 71fc62f..11b50ac (diff)
Rebased 11b50ac..d04ce1b (diff)
Extracted the neonvm commits into neonvm: Small changes for stateless scheduler #1194
Responded to most of the review comments; still a few left.
- Main thing I still want to sort out is the reconcile queue's event types — probably worth a brief discussion (maybe async) about our options there

petuhovskiy

Found a couple of small issues. I can approve everything except framework_methods.go, will take a look at it again tomorrow.

UPD: accidentally pressed approve instead of comment, but no big difference :)

petuhovskiy · 2025-01-03T15:00:21Z

autoscale-scheduler/cmd/main.go

@@ -25,6 +27,7 @@ import (
 func main() {
 	logConfig := zap.NewProductionConfig()
 	logConfig.Sampling = nil // Disable sampling, which the production config enables by default.
+	logConfig.DisableStacktrace = true


question (non-blocking): Why have you disabled stacktraces?

petuhovskiy · 2025-01-06T14:33:12Z

autoscale-scheduler/config_map.yaml

        "minUsageScore": 0.5,
        "maxUsageScore": 0,
-        "scorePeak": 0.8
+        "scorePeak": 0.8,
+        "randomize": false


question (non-blocking): Why randomize is false now? I think it was true before?

petuhovskiy · 2025-01-06T14:48:16Z

pkg/plugin/setup.go

@@ -0,0 +1,70 @@
+package plugin
+


suggestion (non-blocking): Let's rename this file to reconcile.go

petuhovskiy · 2025-01-06T19:05:41Z

pkg/plugin/state/node.go

+		newState.CPU.Watermark != n.CPU.Watermark || newState.Mem.Watermark != n.Mem.Watermark
+
+	// Propagate changes to labels:
+	for label, value := range n.Labels.Entries() {


issue (blocking): Looks like a typo, we should iterate over newState.Labels here:

suggestion:

Suggested change

for label, value := range n.Labels.Entries() {

for label, value := range newState.Labels.Entries() {

petuhovskiy · 2025-01-06T23:05:52Z