Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript
2.6 KB
Raw Blame
# Three-tier PriorityClass scheme (stateless agents).
#
# The scheduler preempts strictly LOWER-priority pods, and prefers evicting the
# LOWEST-priority pods that free enough capacity. Tiers:
#
#   clerum-batch            (-10)  eviction-ELIGIBLE. Recipe Jobs/CronJobs and
#                                  sandbox-recipes workloads. An interactive
#                                  host wake preferentially evicts these.
#   clerum-service          (0)    protected default tier. MCP servers,
#                                  always-on hosts, control plane. Same value
#                                  as unclassed pods, so declaring it is
#                                  documentation + explicit intent, not a
#                                  behavior change.
#   clerum-interactive-host (100)  interactive host wake pods. Preempts
#                                  lower-priority pods when the node is full.
#
# HONEST RESIDUAL: priority alone does not make the service tier untouchable.
# When a full node has no clerum-batch pods to evict, a value-100 wake pod CAN
# still preempt priority-0 pods (unclassed or clerum-service). The mitigation
# is capacity planning: interactive host requests are kept small so a wake
# fits without evicting the service tier, and batch pods are the preferred
# (and normally sufficient) eviction target. Do not describe this scheme as
# absolute protection for services.
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: clerum-batch
value: -10
globalDefault: false
preemptionPolicy: PreemptLowerPriority
description: >-
  Preemptible batch tier — recipe Jobs/CronJobs and sandbox-recipes workloads.
  May be evicted by interactive host wakes (clerum-interactive-host).
---
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: clerum-service
value: 0
globalDefault: false
preemptionPolicy: PreemptLowerPriority
description: >-
  Protected default tier — MCP servers, always-on hosts, control plane. Equal
  to the priority of unclassed pods (0); interactive wakes prefer evicting
  clerum-batch (-10), but on a node with no batch pods priority-0 pods can
  still be preempted by clerum-interactive-host (100) — see residual note.
---
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: clerum-interactive-host
value: 100
globalDefault: false
preemptionPolicy: PreemptLowerPriority
description: >-
  Interactive host wake tier — stateless Host pods scaled up on demand. Keeps
  requests small (capacity planning) so a wake prefers evicting clerum-batch
  pods rather than the priority-0 service tier.