You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, the scheduler acts as the source of truth for the cluster state. We'd like to change that because of the following reasons:
Makes scheduler replication hard, as it is unclear how to reliably replicate that state.
Every scheduler restart results in a state reset, thus a period of warm-up is needed to repopulate the state by the autoscaler-agent. This mechanism is responsible for additional complication of the components logic and protocol, namely:
a. The existence of the Buffer values.
b. The notion of lastPermit.
Potentially, the upscaling communication diagram can be simplified to bypass the autoscaler-agent after the upscaling was approved by scheduler.
Feature idea(s) / DoD
A scheduler can be replicated.
A scheduler restart doesn't cause the sub-optimal scheduling decisions.
We no longer have the concepts of Buffered resources and lastPermit.
Implementation ideas
The scheduler's state and autoscaler-agent ↔ scheduler protocol can be moved entirely to annotations on the associated VirtualMachine object.
Related work
The implementation of neondatabase/neon#8111 might allow us to significantly simplify autoscaler-agent (or even merge it's remaining functionality into other components).
I think the comments were just never updated when the autoscaler-agent
was moved from a per-pod sidecar to a per-node daemonset.
Found this while working on some background work for #995.
Similar to what was done in #1055, we need to explicitly add tolerations
to the scheduler to get it to be recreated more quickly on node failure.
This is particularly necessary because we don't have #995. We could wait
for that, but it's a lot of work, and this is a small thing we can do in
the meantime.
Fixesneondatabase/cloud#17298, part of neondatabase/cloud#14114.
Problem description / Motivation
Currently, the scheduler acts as the source of truth for the cluster state. We'd like to change that because of the following reasons:
a. The existence of the
Buffer
values.b. The notion of
lastPermit
.Feature idea(s) / DoD
Buffered
resources andlastPermit
.Implementation ideas
The scheduler's state and autoscaler-agent ↔ scheduler protocol can be moved entirely to annotations on the associated VirtualMachine object.
Related work
The implementation of neondatabase/neon#8111 might allow us to significantly simplify autoscaler-agent (or even merge it's remaining functionality into other components).
Further information
There are more details in the internal RFC.
The text was updated successfully, but these errors were encountered: