brainy/docs/operations/capacity-planning.md
David Snelling 46c6af3f21 feat: implement always-adaptive caching with getCacheStats monitoring
Replaces lazy mode concept with always-adaptive caching strategy:

- Rename getLazyModeStats() → getCacheStats() with enhanced metrics
- Change lazyModeEnabled boolean → cachingStrategy enum ('preloaded' | 'on-demand')
- Update preloading threshold from 30% to 80% for better cache utilization
- Add comprehensive production monitoring and diagnostics
- Add memory detection for containers (Docker/K8s cgroups v1/v2)
- Add adaptive memory sizing from 2GB to 128GB+ systems

Breaking changes: None (backward compatible, deprecated lazy option ignored)

New APIs:
- getCacheStats(): Comprehensive cache performance statistics
- cachingStrategy field: Transparent strategy reporting
- Enhanced fairness metrics and memory pressure monitoring

Documentation:
- Add migration guide for v3.36.0
- Add operations/capacity-planning.md for enterprise deployments
- Update all examples and troubleshooting guides
- Rename monitor-lazy-mode.ts → monitor-cache-performance.ts
2025-10-10 14:09:30 -07:00

19 KiB
Raw Permalink Blame History

Capacity Planning & Operations Guide

Brainy v3.36.0+ Enterprise Operations

This guide provides production-ready capacity planning formulas, deployment strategies, and operational guidelines for scaling Brainy from development (2GB) to enterprise (128GB+) deployments.


📊 Quick Reference

Memory Allocation Formula

totalAvailable = systemMemory × utilizationFactor
modelReservation = 150MB (Q8) or 250MB (FP32)
availableForCache = totalAvailable - modelReservation
cacheSize = availableForCache × environmentRatio

Where:
- utilizationFactor = 0.80 (leave 20% for OS and other processes)
- environmentRatio = 0.25 (dev), 0.40 (container), 0.50 (production)

Adaptive Caching Strategy

estimatedVectorMemory = entityCount × 1536 bytes  // 384 dims × 4 bytes per float
hnswCacheBudget = cacheSize × 0.80  // 80% threshold for preloading decision

if estimatedVectorMemory < hnswCacheBudget:
    cachingStrategy = 'preloaded'  // All vectors loaded at init
else:
    cachingStrategy = 'on-demand'  // Vectors loaded adaptively via UnifiedCache

🎯 Deployment Scenarios

Scenario 1: Development (2GB System)

System Profile:

  • Total RAM: 2GB
  • Environment: Local development
  • Expected scale: 10K-50K entities

Memory Breakdown:

System Memory:        2048 MB
OS Reserved (20%):    -410 MB
Available:            1638 MB
Model Memory (Q8):    -150 MB
  ├─ Weights:         22 MB
  ├─ ONNX Runtime:    30 MB
  └─ Workspace:       98 MB
───────────────────────────
Available for Cache:  1488 MB
Dev Allocation (25%): 372 MB UnifiedCache
  ├─ HNSW (30%):      112 MB
  ├─ Metadata (40%):  149 MB
  ├─ Search (20%):    74 MB
  └─ Shared (10%):    37 MB

Capacity:

  • Standard Mode: Up to 70K entities (all vectors in memory)
  • Lazy Mode: Up to 500K entities (on-demand vector loading)
  • Search Latency: 5-15ms (standard), 8-20ms (lazy, cold)

Recommendations:

  • Use Q8 model for smaller footprint
  • System uses adaptive caching for datasets >70K entities
  • Monitor cache hit rate with getCacheStats()
  • ⚠️ Expect slower performance vs production systems

Configuration:

const brain = new Brainy({
  storage: { type: 'filesystem', path: './brainy-data' },
  model: { precision: 'q8' },
  cache: { /* auto-sized to 372MB */ }
})

Scenario 2: Small Production (8GB System)

System Profile:

  • Total RAM: 8GB
  • Environment: Single production server
  • Expected scale: 100K-500K entities

Memory Breakdown:

System Memory:        8192 MB
OS Reserved (20%):    -1638 MB
Available:            6554 MB
Model Memory (Q8):    -150 MB
───────────────────────────
Available for Cache:  6404 MB
Prod Allocation (50%): 3202 MB UnifiedCache
  ├─ HNSW (30%):      961 MB
  ├─ Metadata (40%):  1281 MB
  ├─ Search (20%):    640 MB
  └─ Shared (10%):    320 MB

Capacity:

  • Standard Mode: Up to 600K entities
  • Lazy Mode: Up to 5M entities
  • Search Latency: 3-8ms (standard), 5-12ms (lazy, 80% hit rate)

Recommendations:

  • Q8 model balances performance and memory
  • Adaptive on-demand caching activates automatically at ~620K entities
  • Monitor memory pressure warnings
  • Consider horizontal scaling beyond 3M entities

Configuration:

const brain = new Brainy({
  storage: { type: 'filesystem', path: '/var/lib/brainy' },
  model: { precision: 'q8' },
  // Auto-sized cache: 3202MB
})

// Monitor health
const stats = brain.hnsw.getCacheStats()
console.log(`Cache hit rate: ${stats.unifiedCache.hitRatePercent}%`)
console.log(`Caching strategy: ${stats.cachingStrategy}`)

Scenario 3: Medium Production (32GB System)

System Profile:

  • Total RAM: 32GB
  • Environment: Production server or container
  • Expected scale: 1M-10M entities

Memory Breakdown:

System Memory:        32768 MB
OS Reserved (20%):    -6554 MB
Available:            26214 MB
Model Memory (Q8):    -150 MB
───────────────────────────
Available for Cache:  26064 MB
Prod Allocation (50%): 13032 MB UnifiedCache
  ├─ HNSW (30%):      3910 MB
  ├─ Metadata (40%):  5213 MB
  ├─ Search (20%):    2606 MB
  └─ Shared (10%):    1303 MB

Capacity:

  • Standard Mode: Up to 2.5M entities
  • Lazy Mode: Up to 20M entities
  • Search Latency: 2-5ms (standard), 3-8ms (lazy, 85% hit rate)

Recommendations:

  • Consider FP32 model if accuracy is critical (adds 100MB)
  • Enable GCS/S3 storage for durability
  • Adaptive on-demand caching handles 10M+ entities efficiently
  • Monitor fairness metrics to prevent HNSW cache hogging

Configuration:

const brain = new Brainy({
  storage: {
    type: 'gcs-native',
    gcsNativeStorage: { bucketName: 'production-data' }
  },
  model: { precision: 'q8' }  // or 'fp32' for +0.5% accuracy
})

// Verify allocation
const memoryInfo = brain.hnsw.unifiedCache.getMemoryInfo()
console.log(`Cache allocated: ${Math.round(memoryInfo.memoryInfo.available / 1024 / 1024 / 1024)}GB`)
console.log(`Environment: ${memoryInfo.memoryInfo.environment}`)

Scenario 4: Large Production (128GB System)

System Profile:

  • Total RAM: 128GB
  • Environment: Dedicated production server
  • Expected scale: 10M-100M entities

Memory Breakdown:

System Memory:         131072 MB
OS Reserved (20%):     -26214 MB
Available:             104858 MB
Model Memory (FP32):   -250 MB
───────────────────────────────
Available for Cache:   104608 MB
Prod Allocation (50%): 52304 MB UnifiedCache (logarithmic scaling applies)
  ├─ HNSW (30%):       15691 MB
  ├─ Metadata (40%):   20922 MB
  ├─ Search (20%):     10461 MB
  └─ Shared (10%):     5230 MB

Logarithmic Scaling Applied: For systems >64GB, allocation uses logarithmic scaling to prevent over-allocation:

effectiveRatio = baseRatio × (1 + log10(systemGB / 64) × 0.15)
Actual cache size: ~40GB (prevents waste on 128GB systems)

Capacity:

  • Standard Mode: Up to 10M entities
  • Lazy Mode: Up to 100M+ entities
  • Search Latency: 1-3ms (standard), 2-5ms (lazy, 90%+ hit rate)

Recommendations:

  • Use FP32 model for maximum accuracy
  • Enable distributed storage (S3/GCS)
  • Monitor fairness violations (HNSW shouldn't dominate cache)
  • Consider sharding beyond 50M entities
  • Implement application-level caching for hot queries

Configuration:

const brain = new Brainy({
  storage: {
    type: 's3',
    s3Storage: {
      bucketName: 'enterprise-data',
      region: 'us-east-1'
    }
  },
  model: { precision: 'fp32' }  // Maximum accuracy
})

// Enterprise monitoring
setInterval(() => {
  const stats = brain.hnsw.getCacheStats()

  if (stats.fairness.fairnessViolation) {
    console.warn('FAIRNESS VIOLATION: HNSW using too much cache')
    console.warn(`HNSW: ${stats.fairness.hnswAccessPercent}% access, ${stats.hnswCache.sizePercent}% size`)
  }

  if (stats.unifiedCache.hitRatePercent < 75) {
    console.warn(`Low cache hit rate: ${stats.unifiedCache.hitRatePercent}%`)
    console.warn('Recommendations:', stats.recommendations)
  }
}, 60000)  // Check every minute

🐳 Container Deployments (Docker/Kubernetes)

Container Memory Detection

Brainy auto-detects container memory limits via cgroups v1/v2:

// Automatic detection
const brain = new Brainy()  // Detects cgroup limits automatically

// Verify detection
const memoryInfo = brain.hnsw.unifiedCache.getMemoryInfo()
console.log(`Container: ${memoryInfo.memoryInfo.isContainer}`)
console.log(`Source: ${memoryInfo.memoryInfo.source}`)  // 'cgroup-v2' or 'cgroup-v1'
console.log(`Limit: ${Math.round(memoryInfo.memoryInfo.available / 1024 / 1024)}MB`)

Docker Resource Limits

Small Container (2GB)

FROM node:22-alpine

WORKDIR /app
COPY package*.json ./
RUN npm ci --production

COPY . .

# Download models at build time
RUN npm run download-models

ENV NODE_OPTIONS="--max-old-space-size=1536"

CMD ["node", "dist/index.js"]
docker run \
  --memory="2g" \
  --memory-reservation="1.5g" \
  --cpus="2" \
  my-brainy-app

Expected allocation:

Container Limit:      2048 MB
Available:            1638 MB (80% usable)
Model Memory:         -150 MB
Available for Cache:  1488 MB
Container Ratio (40%): 595 MB UnifiedCache

Medium Container (8GB)

docker run \
  --memory="8g" \
  --memory-reservation="6g" \
  --cpus="4" \
  -e NODE_OPTIONS="--max-old-space-size=6144" \
  my-brainy-app

Expected allocation:

Container Limit:      8192 MB
Available:            6554 MB
Model Memory:         -150 MB
Available for Cache:  6404 MB
Container Ratio (40%): 2562 MB UnifiedCache

Kubernetes Resource Requests/Limits

Small Pod (2GB)

apiVersion: apps/v1
kind: Deployment
metadata:
  name: brainy-api
spec:
  replicas: 3
  template:
    spec:
      containers:
      - name: brainy
        image: my-brainy-app:latest
        resources:
          requests:
            memory: "1.5Gi"
            cpu: "500m"
          limits:
            memory: "2Gi"
            cpu: "1000m"
        env:
        - name: NODE_OPTIONS
          value: "--max-old-space-size=1536"

Medium Pod (8GB)

apiVersion: apps/v1
kind: Deployment
metadata:
  name: brainy-api
spec:
  replicas: 2
  template:
    spec:
      containers:
      - name: brainy
        image: my-brainy-app:latest
        resources:
          requests:
            memory: "6Gi"
            cpu: "2000m"
          limits:
            memory: "8Gi"
            cpu: "4000m"
        env:
        - name: NODE_OPTIONS
          value: "--max-old-space-size=6144"

Best Practices:

  • Set requests to 75% of limits for better scheduling
  • Download models at Docker build time (not runtime)
  • Use NODE_OPTIONS to match container memory limits
  • Monitor actual usage and adjust based on workload

📈 Scaling Strategies

Adaptive Caching Behavior

The system automatically chooses the optimal caching strategy:

  • Preloaded: Small datasets (<80% of cache) - all vectors loaded at init for zero-latency access
  • On-demand: Large datasets (>80% of cache) - vectors loaded adaptively via UnifiedCache
  • No configuration needed - system adapts automatically based on dataset size

Auto-detection logic:

const vectorMemoryNeeded = entityCount × 1536  // bytes
const hnswCacheAvailable = unifiedCache.maxSize × 0.80

if (vectorMemoryNeeded < hnswCacheAvailable) {
  // Preload strategy: all vectors loaded at init
  console.log('Caching strategy: preloaded (all vectors in memory)')
} else {
  // On-demand strategy: vectors loaded adaptively
  console.log('Caching strategy: on-demand (adaptive loading via UnifiedCache)')
}

When to Add More RAM

Consider increasing RAM when:

  • ⚠️ Cache hit rate consistently < 70%
  • ⚠️ Memory pressure warnings > 85% utilization
  • ⚠️ Search latency > 20ms on hot paths
  • ⚠️ On-demand caching active but working set is large

Decision tree:

If cache hit rate < 70%:
  └─> Is working set < 50% of total entities?
      ├─> YES: Increase cache size (add RAM)
      └─> NO: Working set too large, consider:
          ├─> Application-level caching
          ├─> Query optimization
          └─> Sharding dataset

When to Shard/Distribute

Consider sharding when:

  • ⚠️ Entity count > 50M entities on single node
  • ⚠️ Write throughput > 10K ops/sec
  • ⚠️ Need geographic distribution
  • ⚠️ Fault tolerance requirements

Sharding strategy:

// Example: Geographic sharding
const usEastBrain = new Brainy({
  storage: { type: 's3', s3Storage: { bucket: 'us-east-data' } }
})

const euWestBrain = new Brainy({
  storage: { type: 's3', s3Storage: { bucket: 'eu-west-data' } }
})

// Route queries based on user location
async function search(query, userRegion) {
  const brain = userRegion === 'US' ? usEastBrain : euWestBrain
  return await brain.search(query)
}

🔍 Monitoring & Diagnostics

Key Metrics to Track

1. Cache Performance

const stats = brain.hnsw.getCacheStats()

// Cache hit rate (target: >80%)
console.log(`Hit rate: ${stats.unifiedCache.hitRatePercent}%`)

// HNSW cache utilization
console.log(`HNSW memory: ${stats.hnswCache.estimatedMemoryMB}MB`)
console.log(`HNSW hit rate: ${stats.hnswCache.hitRatePercent}%`)

2. Memory Pressure

const memoryInfo = brain.hnsw.unifiedCache.getMemoryInfo()

console.log(`Pressure: ${memoryInfo.currentPressure.pressure}`)
// Values: 'low', 'moderate', 'high', 'critical'

if (memoryInfo.currentPressure.warnings.length > 0) {
  console.warn('Memory warnings:', memoryInfo.currentPressure.warnings)
}

3. Fairness Metrics

const stats = brain.hnsw.getCacheStats()

if (stats.fairness.fairnessViolation) {
  console.warn('Cache fairness violation detected')
  console.warn(`HNSW: ${stats.fairness.hnswAccessPercent}% access`)
  console.warn(`HNSW: ${stats.hnswCache.sizePercent}% of cache`)
}

4. Query Performance

// Track search latency
console.time('search')
const results = await brain.search('query')
console.timeEnd('search')  // Target: <10ms for hot queries

Alerting Thresholds

Set up alerts for:

  • ⚠️ Cache hit rate < 70% (sustained for 5+ minutes)
  • 🚨 Memory utilization > 90%
  • 🚨 Search latency > 50ms (p95)
  • ⚠️ Fairness violations detected

Example monitoring script:

async function monitorHealth() {
  const stats = brain.hnsw.getCacheStats()

  // Alert on low cache hit rate
  if (stats.unifiedCache.hitRatePercent < 70) {
    await sendAlert({
      severity: 'warning',
      message: `Low cache hit rate: ${stats.unifiedCache.hitRatePercent}%`,
      recommendations: stats.recommendations
    })
  }

  // Alert on memory pressure
  const memoryInfo = brain.hnsw.unifiedCache.getMemoryInfo()
  if (memoryInfo.currentPressure.pressure === 'high') {
    await sendAlert({
      severity: 'critical',
      message: 'High memory pressure detected',
      warnings: memoryInfo.currentPressure.warnings
    })
  }
}

// Run every 60 seconds
setInterval(monitorHealth, 60000)

🎯 Real-World Examples

Example 1: E-Commerce Product Catalog (500K products)

System: 16GB production server

Sizing:

Products:              500,000
Vector memory needed:  500K × 1536 bytes = 768 MB
HNSW cache available:  (16GB × 0.8 - 150MB) × 0.5 × 0.3 = 1,915 MB

Result: Standard mode (all vectors fit in HNSW cache)

Configuration:

const brain = new Brainy({
  storage: { type: 'filesystem', path: '/var/lib/brainy' },
  model: { precision: 'q8' }
})

await brain.init()

// Verify preloaded strategy (all vectors in memory)
const stats = brain.hnsw.getCacheStats()
console.log(`Caching strategy: ${stats.cachingStrategy}`)  // 'preloaded'
console.log(`Search latency: ${stats.performance.avgSearchMs}ms`)  // ~3ms

Example 2: Document Search (5M documents)

System: 32GB production server with GCS storage

Sizing:

Documents:             5,000,000
Vector memory needed:  5M × 1536 bytes = 7,680 MB
HNSW cache available:  (32GB × 0.8 - 150MB) × 0.5 × 0.3 = 3,910 MB

Result: On-demand caching (vectors loaded adaptively)

Configuration:

const brain = new Brainy({
  storage: {
    type: 'gcs-native',
    gcsNativeStorage: { bucketName: 'docs-production' }
  },
  model: { precision: 'q8' }
})

await brain.init()

// Monitor cache performance
const stats = brain.hnsw.getCacheStats()
console.log(`Caching strategy: ${stats.cachingStrategy}`)  // 'on-demand'
console.log(`Cache hit rate: ${stats.unifiedCache.hitRatePercent}%`)  // Target >80%
console.log(`Cold search latency: ${stats.performance.avgSearchMs}ms`)  // ~12ms

// Recommendations
console.log('Recommendations:', stats.recommendations)
// Example: "Cache hit rate healthy at 84.2% - no action needed"

Example 3: Knowledge Graph (20M entities)

System: 128GB dedicated server with S3 storage

Sizing:

Entities:              20,000,000
Vector memory needed:  20M × 1536 bytes = 30,720 MB
HNSW cache available:  ~15,691 MB (after logarithmic scaling)

Result: On-demand caching with high-performance adaptive loading

Configuration:

const brain = new Brainy({
  storage: {
    type: 's3',
    s3Storage: {
      bucketName: 'knowledge-graph-prod',
      region: 'us-east-1'
    }
  },
  model: { precision: 'fp32' }  // Maximum accuracy
})

await brain.init()

// Enterprise monitoring
const stats = brain.hnsw.getCacheStats()
console.log(`Entities: ${stats.autoDetection.entityCount.toLocaleString()}`)
console.log(`Caching strategy: ${stats.cachingStrategy}`)  // 'on-demand'
console.log(`Cache hit rate: ${stats.unifiedCache.hitRatePercent}%`)  // Target >85%
console.log(`HNSW cache: ${stats.hnswCache.estimatedMemoryMB}MB`)

// Fairness check
if (stats.fairness.fairnessViolation) {
  console.warn('HNSW dominating cache - consider tuning eviction policies')
}

🛠️ Troubleshooting

Issue: Low Cache Hit Rate (<70%)

Diagnosis:

const stats = brain.hnsw.getCacheStats()
console.log(`Hit rate: ${stats.unifiedCache.hitRatePercent}%`)
console.log(`Working set: ${stats.hnswCache.estimatedMemoryMB}MB`)

Solutions:

  1. Increase cache size (add RAM)
  2. Optimize query patterns (reduce random access)
  3. Implement application-level caching
  4. Consider sharding if working set > available cache

Issue: High Memory Pressure (>85%)

Diagnosis:

const memoryInfo = brain.hnsw.unifiedCache.getMemoryInfo()
console.log(`Pressure: ${memoryInfo.currentPressure.pressure}`)
console.log(`Warnings:`, memoryInfo.currentPressure.warnings)

Solutions:

  1. Reduce cache size manually (override auto-detection)
  2. Reduce entity count (archive old data - system automatically uses on-demand caching for large datasets)
  3. Increase system RAM

Issue: Fairness Violations

Diagnosis:

const stats = brain.hnsw.getCacheStats()
if (stats.fairness.fairnessViolation) {
  console.log(`HNSW access: ${stats.fairness.hnswAccessPercent}%`)
  console.log(`HNSW cache: ${stats.hnswCache.sizePercent}%`)
}

Solutions:

  1. Contact support (fairness policies may need tuning)
  2. Monitor over time (may self-correct as access patterns stabilize)
  3. File GitHub issue with diagnostics

📚 Additional Resources


Production-ready. Enterprise-scale. Zero-config. 🚀