Resource Sizing

Guidelines for sizing your Laminar deployment based on workload requirements.

Overview

Proper resource sizing ensures:

Performance - Adequate resources for throughput
Stability - Avoiding OOM kills and CPU throttling
Cost efficiency - Not over-provisioning
Scalability - Room for growth

Quick Reference

Deployment Size	Events/sec	API	Controller	Workers	RocksDB Storage	Nodes
Development	< 1K	1x 256Mi	1x 512Mi	1x 2Gi	5Gi	1
Small	1K-10K	2x 512Mi	2x 1Gi	2-4x 4Gi	20Gi	3
Medium	10K-100K	3x 1Gi	2x 2Gi	4-8x 8Gi	50Gi	5
Large	100K-1M	5x 2Gi	3x 4Gi	8-16x 16Gi	200Gi	10+

Component Sizing

API Server

The API server handles HTTP requests and WebSocket connections.

api:
  replicas: 2
  resources:
    requests:
      cpu: 250m
      memory: 512Mi
    limits:
      cpu: 1000m
      memory: 1Gi

Controller

The controller manages pipeline orchestration and scheduling with embedded RocksDB.

controller:
  replicas: 2
  resources:
    requests:
      cpu: 250m
      memory: 1Gi
    limits:
      cpu: 1000m
      memory: 2Gi
  persistence:
    size: 20Gi
    storageClass: gp3
  checkpointInterval: "30s"
  maxParallelism: 16

Workers

Workers execute pipeline tasks and are the most resource-intensive component.

worker:
  replicas: 2
  resources:
    requests:
      cpu: 1000m
      memory: 4Gi
    limits:
      cpu: 2000m
      memory: 8Gi
  taskSlots: 4

Storage Sizing

RocksDB (Local State)

RocksDB storage requirements depend on pipeline state size:

Workload	Storage	IOPS	Storage Class
Development	5-10 Gi	3000	standard
Small	20-50 Gi	3000	gp3
Medium	50-100 Gi	6000	gp3
Large	200-500 Gi	10000+	io2

Object Storage (Checkpoints & Artifacts)

Estimate checkpoint and artifact storage:

Metric	Formula
Checkpoints	`pipelines × checkpoint_interval × retention`
Artifacts	`pipeline_runs × avg_artifact_size`

Example:

10 pipelines, 10s checkpoints, 7 days retention
10 × 8640 × 7 ≈ 600K checkpoints
At ~1MB each ≈ 600 GB

Node Sizing

AWS Instance Types

Size	Instance Type	vCPU	Memory	Use Case
Dev	t3.medium	2	4 GB	Development
Small	m5.xlarge	4	16 GB	Light production
Medium	m5.2xlarge	8	32 GB	Standard production
Large	m5.4xlarge	16	64 GB	Heavy workloads
Worker	c5.2xlarge	8	16 GB	CPU-intensive

GCP Machine Types

Size	Machine Type	vCPU	Memory
Dev	e2-medium	2	4 GB
Small	e2-standard-4	4	16 GB
Medium	e2-standard-8	8	32 GB
Large	e2-standard-16	16	64 GB

Azure VM Sizes

Size	VM Size	vCPU	Memory
Dev	Standard_D2s_v3	2	8 GB
Small	Standard_D4s_v3	4	16 GB
Medium	Standard_D8s_v3	8	32 GB
Large	Standard_D16s_v3	16	64 GB

Autoscaling

Horizontal Pod Autoscaler

api:
  autoscaling:
    enabled: true
    minReplicas: 2
    maxReplicas: 10
    targetCPUUtilizationPercentage: 70
    targetMemoryUtilizationPercentage: 80
 
worker:
  autoscaling:
    enabled: true
    minReplicas: 2
    maxReplicas: 20
    targetCPUUtilizationPercentage: 70

Cluster Autoscaler

Enable cluster autoscaling to add/remove nodes:

# EKS
eksctl create cluster ... --asg-access
 
# GKE
gcloud container clusters create ... --enable-autoscaling
 
# AKS
az aks create ... --enable-cluster-autoscaler

Performance Testing

Load Testing

# Generate test load
hey -n 10000 -c 100 http://laminar.example.com/api/health
 
# Monitor during test
kubectl top pods -n laminar
kubectl top nodes

Key Metrics to Monitor

Metric	Target	Action if Exceeded
API latency (p99)	< 100ms	Add replicas
CPU utilization	< 70%	Scale horizontally
Memory utilization	< 80%	Increase limits
Pipeline lag	< 1s	Add workers
RocksDB disk usage	< 80%	Expand PVC

Cost Optimization

Reserved Capacity

Use reserved instances/committed use for stable workloads:

1-year commitment: ~30% savings
3-year commitment: ~50% savings

Spot/Preemptible Instances

Use for workers (with proper checkpointing):

worker:
  nodeSelector:
    node-type: spot
  tolerations:
    - key: "spot"
      operator: "Equal"
      value: "true"
      effect: "NoSchedule"

Sizing Checklist

Identified workload characteristics (events/sec, pipeline count)
Calculated storage requirements (RocksDB + object storage)
Selected appropriate node types
Configured resource requests and limits
Enabled autoscaling for variable workloads
Set up monitoring for resource utilization
Tested under expected load

Next Steps

High Availability - HA configuration
Backup & Restore - Data protection
Troubleshooting - Common issues