Next generation MT Scheduler and Descheduler with pluggable policies

[aavarghese]

Fixes #593
🎁 Implementing next generation MT scheduler and descheduler for Knative Eventing sources with plugins for filtering pods
and scoring them to pick the best placement pod.

Proposed Changes

• Controller reads and validates the policy profile (containing predicates and priorities) for both scheduling and descheduling from ConfigMaps.
• Scheduler imports all the available plugins which invokes their init() and registers all the core and source specific (kafka) plugins.
• Scheduler runs all the predicates and priorities specified in the policy (configmap) file to eliminate all the pods that do not meet the predicate, and then prioritizes the feasible pods to find the best placement given the current state of the cluster.
• Same for descheduler (has separate predicates and priorities).
• Controls are also shared with autoscaler, and state accessor.
• Autoscaler runs autoscaling synchronously using a lock so that autoscaling does not run together with the scheduler, and also scales up number of pods using a scale factor based on number of resources in the HA domain
• State accessor computes the spread of existing and reserved vreplicas across pods, nodes and zones and stores a map of vpod key to a map of podname/nodename/zonename to vreplica count. It also stores the number of pods, nodes, zones and a list of "schedulable" pods (unschedulable pods are pods that are marked for eviction for compacting, and pods that are on unschedulable nodes (cordoned or unreachable nodes))
• Pending vreplicas that cannot be scheduled triggers the autoscaler to add new pods. This leads to rebalancing of all the already placed vreplicas together with the pending vreplicas.
• When its time to autoscale down (after refresh period), compacting the pods is attempted if space is available on lower pods. Higher pods are evicted and removed from the "schedulable" pods list. This leads to another rebalancing of all the vreplicas.
• Recovery situations are addressed by making sure pods on failed nodes/zones are restarted on new nodes (statefulset adapter settings) and state collector is aware of nodes/zones being down.

Release Note

Next generation Multi-Tenant Scheduler and Descheduler: uses a plugin interface to specify a Scheduler profile with predicates and priorities that run filtering and scoring of pods, respectively to compute the best vreplica placements. When the autoscaler adds new pods, scheduler performs a rebalancing of the already placed vreplicas along with the new vreplicas. A Descheduler profile must be installed when vreplicas need to be scaled down and placements removed. 

[lionelvillard]

/test pull-knative-sandbox-eventing-kafka-unit-tests

[codecov]

Codecov Report

Merging #768 (4f83ec9) into main (15bb0b1) will increase coverage by 0.11%.
The diff coverage is 79.05%.

@@            Coverage Diff             @@
##             main     #768      +/-   ##
==========================================
+ Coverage   74.64%   74.75%   +0.11%     
==========================================
  Files         143      154      +11     
  Lines        6633     7139     +506     
==========================================
+ Hits         4951     5337     +386     
- Misses       1435     1518      +83     
- Partials      247      284      +37     

Impacted Files	Coverage Δ
...e/lowestordinalpriority/lowest_ordinal_priority.go	66.66% <66.66%> (ø)
...alpriority/remove_with_highest_ordinal_priority.go	66.66% <66.66%> (ø)
...uler/plugins/core/evenpodspread/even_pod_spread.go	70.90% <70.90%> (ø)
...labilitynodepriority/availability_node_priority.go	72.72% <72.72%> (ø)
...labilityzonepriority/availability_zone_priority.go	74.28% <74.28%> (ø)
pkg/common/scheduler/state/interface.go	75.75% <75.75%> (ø)
pkg/common/scheduler/statefulset/scheduler.go	76.80% <75.93%> (-6.11%)	⬇️
pkg/common/scheduler/state/state.go	81.94% <81.94%> (ø)
.../kafka/nomaxresourcecount/no_max_resource_count.go	82.35% <82.35%> (ø)
...adpriority/remove_with_even_pod_spread_priority.go	83.33% <83.33%> (ø)
... and 19 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 15bb0b1...4f83ec9. Read the comment docs.

[lionelvillard]

I think this line requires some explanation. In theory the vpod scheduler shouldn't have to deal with nodes directly. A pod on an unschedulable node might still be viable, right?

[aavarghese]

Thin line I feel. A node is marked as Unschedulable in its Spec after it has been optionally drained and cordoned. K8 scheduler won't place new pods onto that Node, so idea was that we should not place "new vreplicas" onto this pod either. Because the next step for this Node might be rebooting/shutting down/upgrading etc. So this is a preventative step to add more rebalancing work. That was my thinking..

[lionelvillard]

same here. Maybe we just don't care about pod eviction. This has to be dealt somewhere else anyway

[aavarghese]

Each plugin must be given some info to prevent it from selecting a pod that is in the process of eviction. Not sure yet of how to do that without the state collector telling it so...will think about this a bit more.

[knative-metrics-robot]

The following is the coverage report on the affected files.
Say /test pull-knative-sandbox-eventing-kafka-go-coverage to re-run this coverage report

File	Old Coverage	New Coverage	Delta
pkg/apis/sources/v1beta1/kafka_scheduling.go	80.0%	81.8%	1.8
pkg/common/scheduler/plugins/core/availabilitynodepriority/availability_node_priority.go	Do not exist	81.8%
pkg/common/scheduler/plugins/core/availabilityzonepriority/availability_zone_priority.go	Do not exist	83.3%
pkg/common/scheduler/plugins/core/evenpodspread/even_pod_spread.go	Do not exist	81.8%
pkg/common/scheduler/plugins/core/lowestordinalpriority/lowest_ordinal_priority.go	Do not exist	66.7%
pkg/common/scheduler/plugins/core/podfitsresources/pod_fits_resources.go	Do not exist	100.0%
pkg/common/scheduler/plugins/core/removewithavailabilitynodepriority/remove_with_availability_node_priority.go	Do not exist	87.9%
pkg/common/scheduler/plugins/core/removewithavailabilityzonepriority/remove_with_availability_zone_priority.go	Do not exist	88.9%
pkg/common/scheduler/plugins/core/removewithevenpodspreadpriority/remove_with_even_pod_spread_priority.go	Do not exist	86.7%
pkg/common/scheduler/plugins/core/removewithhighestordinalpriority/remove_with_highest_ordinal_priority.go	Do not exist	66.7%
pkg/common/scheduler/plugins/kafka/nomaxresourcecount/no_max_resource_count.go	Do not exist	88.2%
pkg/common/scheduler/state/helpers.go	Do not exist	92.6%
pkg/common/scheduler/state/interface.go	Do not exist	86.2%
pkg/common/scheduler/state/state.go	Do not exist	88.1%
pkg/common/scheduler/statefulset/autoscaler.go	80.3%	85.3%	5.0
pkg/common/scheduler/statefulset/scheduler.go	88.1%	83.4%	-4.7

[lionelvillard]

/lgtm
/approve

[knative-prow-robot]

[APPROVALNOTIFIER] This PR is APPROVED

This pull-request has been approved by: aavarghese, lionelvillard

The full list of commands accepted by this bot can be found here.

The pull request process is described here

Details

Needs approval from an approver in each of these files:

• OWNERS [lionelvillard]

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[aavarghese]

Tracking optimizations remaining to do:

1. Plugin args should be in the same package ex: Next generation MT Scheduler and Descheduler with pluggable policies #768 (comment)
2. Node and eviction details collected in State - should be elsewhere?
3. generate example checksum Next generation MT Scheduler and Descheduler with pluggable policies #768 (comment)
4. Convert json to yaml Next generation MT Scheduler and Descheduler with pluggable policies #768 (comment)
5. Modify the scheduler adapter model from StatefulSet to Deployment?
6. Bulk scheduling of vreplicas (instead of one at a time) to improve performance
7. Rebalancing with minimal placement changes when #pods/nodes/zone changes
8. Initial number of replicas (default:1) from statefulset not a multiple of scale factor for HA Fix for [scheduler] Crash when creating state #882

/cc @lionelvillard