FINOS 5-Spot Scheduler: Time-Based Kubernetes Scheduling for Financial Services Infrastructure

Here's a pattern every financial services infrastructure team recognises: trading systems need full compute capacity from 07:00 to 18:00 CET. Risk analytics run from 18:00 to 22:00. Overnight batch processing runs from 01:00 to 05:00. Between 05:00 and 07:00, and on weekends, most of that compute sits idle — burning electricity, depreciating hardware, and costing money for zero business value.

Cloud-native teams solve this with autoscaling: HPA and VPA react to load. But reactive autoscaling has a cold-start problem. When the trading desk opens at 07:00, the first 5-10 minutes of requests hit under-provisioned infrastructure because the autoscaler hasn't caught up yet. For financial services, those first minutes are some of the most critical of the day.

5-Spot Machine Scheduler — a Kubernetes operator being contributed to FINOS (the Fintech Open Source Foundation) — takes a different approach. It's proactive, not reactive. You define time-based schedules: "provision 8 bare-metal GPU nodes at 06:45 CET on weekdays, de-provision at 18:15." Infrastructure is ready before the workload arrives.

Why Time-Based Scheduling Matters

Financial services workloads are uniquely predictable. Unlike consumer web traffic (unpredictable spikes), financial markets operate on fixed schedules:

• European markets: 08:00-16:30 CET
• US markets: 14:30-21:00 CET
• Asian markets: 01:00-07:00 CET
• Risk calculations: End-of-day, typically 17:00-22:00
• Batch processing: Overnight, 01:00-05:00
• Development/testing: Business hours, 08:00-18:00

If your compute follows these patterns — and in regulated finance, it overwhelmingly does — then time-based scheduling isn't just an optimisation. It's the correct architecture. You know exactly when you need capacity. Schedule it.

What 5-Spot Scheduler Does

5-Spot is a Kubernetes operator that manages the lifecycle of physical machines based on time schedules:

# 5-Spot Machine Schedule CRD (conceptual)
apiVersion: scheduling.finos.org/v1alpha1
kind: MachineSchedule
metadata:
  name: trading-gpu-cluster
  namespace: trading-infra
spec:
  machinePool:
    provider: bare-metal     # or cloud, edge
    count: 8
    machineClass: gpu-a100   # Maps to machine profile
  schedule:
    timezone: Europe/London
    windows:
      - name: trading-hours
        days: [monday, tuesday, wednesday, thursday, friday]
        startTime: "06:45"   # 15 min before market open
        endTime: "18:15"     # 15 min after market close
      - name: risk-batch
        days: [monday, tuesday, wednesday, thursday, friday]
        startTime: "18:00"
        endTime: "22:00"
        machineCount: 4       # Reduced capacity for batch
  lifecycle:
    gracefulShutdownPeriod: 15m  # Drain workloads before shutdown
    preWarmMinutes: 10           # Start provisioning 10 min early
    killSwitch: true             # Emergency immediate shutdown
  priority: high                 # Resource allocation priority
  observability:
    metrics: true
    events: true
    costTracking: true

Key Capabilities

• Physical infrastructure focus: Unlike cloud autoscalers that spin up VMs, 5-Spot manages bare-metal servers, edge devices, and hybrid deployments. It integrates with Cluster API and k0smotron for machine lifecycle management
• Timezone-aware scheduling: Financial services operate across time zones. A single cluster might have London, New York, and Singapore schedules running simultaneously. 5-Spot handles timezone math, DST transitions, and holiday calendars
• Graceful lifecycle: Configurable drain periods ensure workloads migrate cleanly before machines are de-provisioned. No abrupt shutdowns. No orphaned processes. Emergency kill-switch available for immediate shutdown when needed
• Pre-warming: Machines provisioned before the schedule window opens. By the time traders arrive, infrastructure is ready, warm, and serving. Zero cold-start latency
• GitOps-native: Schedules defined as Kubernetes CRDs, stored in Git, deployed via ArgoCD. Schedule changes go through PR review — auditable, version-controlled, and reversible

The Cost Impact

The economics of time-based scheduling are straightforward:

Cost Analysis: 8x A100 GPU Nodes

Item	Detail
24/7 Operation (current):
Hours/week	168
Monthly cost	52,000 EUR (cloud) / 18,000 EUR (on-prem amortised + power)
Time-Scheduled Operation:
Trading hours (M-F, 06	45-18:15): 57.5 hours
Risk batch (M-F, 18	00-22:00): 20 hours (4 nodes)
Total active	~67.5 node-equivalent hours
Monthly cost	21,000 EUR (cloud) / 7,200 EUR (on-prem)
Savings:
Cloud	31,000 EUR/month (60% reduction)
On-prem	10,800 EUR/month (60% reduction — mostly power + cooling)
Annual cloud savings	372,000 EUR
Annual on-prem savings	129,600 EUR
Energy reduction:
24/7	8 nodes x 700W = 5.6kW continuous = 4,032 kWh/month
Scheduled	~1,620 kWh/month
Reduction	60% energy, ~2.4 tonnes CO2/year avoided

For a regulated financial institution running dozens of these clusters, the savings compound to millions annually. And the energy reduction maps directly to ESG reporting metrics.

Integration with AI Infrastructure

5-Spot is particularly valuable for AI workloads in financial services, where GPU compute is the most expensive resource:

• AI inference during market hours: Real-time inference (pricing models, risk scoring, fraud detection) needs GPU capacity during trading hours. Schedule it. De-provision after close
• Model training overnight: Large model training jobs run on the same GPU nodes during overnight windows. Same hardware, different schedule, different workload
• Development environments: Data scientists need GPU access during business hours. Schedule GPU nodes for the data science team from 08:00 to 19:00, de-provision overnight. No more forgotten dev instances running all weekend
• Batch analytics: End-of-day risk calculations, portfolio rebalancing, and regulatory reporting — schedule the compute window, run the batch, de-provision. Scale-to-zero between windows

Compliance and Auditability

For regulated financial services under DORA and NIS2, time-based scheduling provides compliance benefits:

• Auditable schedules: Every schedule change is a Git commit. Regulators can see exactly when infrastructure was provisioned, who approved the change, and why. GitOps provides the audit trail DORA requires for ICT change management
• Reduced attack surface: Machines that are de-provisioned aren't vulnerable. An overnight attack against trading infrastructure hits nothing if the nodes are shut down. Smaller operational window = smaller attack surface
• Cost governance: Budget controls enforced through scheduling. Teams can't leave GPU instances running over the weekend because the scheduler de-provisions them. Accidental cost overruns eliminated by design
• Emergency controls: Kill-switch functionality provides immediate shutdown capability required by DORA for ICT incident response. "Shut down all trading compute immediately" is one kubectl command

Combining 5-Spot with KEDA and Cluster Autoscaler

5-Spot handles predictable, time-based capacity. But within those windows, you still need reactive scaling for variable load. The architecture layers:

Layered Scheduling Architecture

Getting Involved

5-Spot is being contributed to FINOS as an open-source project. This matters for financial services adoption because:

• FINOS governance: Neutral foundation governance means no single vendor controls the roadmap. Financial institutions can contribute, influence, and trust the project
• Complementary to OpenGRIS: 5-Spot handles compute scheduling (when machines are available), while OpenGRIS handles task scheduling (what runs on those machines). Together, they provide complete grid computing orchestration
• Community validation: Born from real needs at RBC, one of the largest financial institutions. The problem it solves is universal across financial services
• Standards-based: Built on CNCF standards (Kubernetes, Cluster API). No proprietary lock-in. Integrates with existing Kubernetes infrastructure

The project is in the FINOS onboarding process. Follow the progress at the FINOS community issue.

What Open Empower Delivers

• Time-based scheduling design: Analyse your workload patterns and design optimal scheduling windows. Map market hours, batch processing, development, and AI training to time-based schedules
• 5-Spot implementation: Deploy and configure 5-Spot Scheduler on your Kubernetes infrastructure with Cluster API integration for bare-metal and hybrid cloud
• Layered scaling architecture: Combine 5-Spot (proactive) with Cluster Autoscaler (reactive nodes) and KEDA (reactive pods) for optimal cost and performance
• Cost modelling: Before/after cost analysis for your specific infrastructure. Calculate savings from time-based scheduling across all workload types
• Compliance integration: GitOps-native schedule management with DORA-compliant audit trails, change management, and emergency controls

Your infrastructure doesn't need to run 24/7 to be ready when you need it. Let's schedule it intelligently.