Skip to content

vCPU budgets #25

Closed
Closed

Description

@sodre

Parent

#18

What to build

Caps become vCPU budgets. SpotCap and OnDemandCap (runner counts) are replaced by SpotVcpuCap (default 24) and OnDemandVcpuCap (default 8). The launcher sums the vCPUs of the capacity's live runners from their shape tags. A job that would take its capacity over budget waits on the queue for 60 s, as today. Budgets cover both architectures, and they stay soft.

Acceptance criteria

  • Launcher tests cover mixed sizes under, at and over a budget, across both architectures.
  • The deferral log line reports vCPUs, for example "deferred": "20+8 of 24 spot vCPUs".
  • The old parameters are gone. The runbook's override examples and traps table use the new names, and say which AWS quota each budget draws on.
  • A live deploy keeps the same effective limits (24 Spot and 8 On-Demand vCPUs).

Activity

  1. added a commit that references this issue on Nov 14, 2026
  2. sodre commented on Nov 14, 2026

    @sodre
    MemberAuthor

    Live: SpotVcpuCap 24, OnDemandVcpuCap 8 after deploy. With OnDemandVcpuCap=4 the launcher logged {"deferred":"4+4 of 4 on-demand vCPUs"} and both jobs ran (runs 43544010312, 43544071525 in ci-runner-canary); restored to 24/8. Seen live: a burst overshoots the budget (zae-zcm queued 8 Spot jobs at once, 36 of 24 vCPUs) because the launcher runs messages in parallel; follow-up decision pending.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    ready-for-agentFully specified, ready for an AFK agent

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions