docs: add distributed deployment architecture and enhancement proposals

- Add comprehensive guide for multi-instance S3-backed deployment
- Document configuration strategies for read/write separated instances
- Propose core enhancements for distributed operations
- Include implementation timeline and monitoring strategies

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
David Snelling 2025-08-04 08:42:31 -07:00
parent 90961c5ce9
commit 9b3a8a2d05
2 changed files with 958 additions and 0 deletions

View file

@ -0,0 +1,464 @@
# Proposed Brainy Enhancements for Distributed Operations
## Executive Summary
To fully support the distributed deployment scenario with multiple specialized instances sharing S3 storage, Brainy needs several enhancements focused on coordination, consistency, and operational modes.
## Core Enhancement Areas
### 1. Instance Role Management
**Current State**: Brainy operates as a standalone instance without awareness of other instances.
**Proposed Enhancement**:
```typescript
// New distributed configuration options
export interface DistributedConfig {
role: 'reader' | 'writer' | 'hybrid';
instanceId: string;
coordinationMethod: 'none' | 's3-polling' | 'websocket' | 'pubsub';
consistencyLevel: 'eventual' | 'strong' | 'bounded';
conflictResolution: 'last-write-wins' | 'vector-clock' | 'crdt';
}
// Enhanced BrainyData constructor
class BrainyData {
constructor(config: BrainyConfig & { distributed?: DistributedConfig }) {
if (config.distributed) {
this.initializeDistributedMode(config.distributed);
}
}
private initializeDistributedMode(config: DistributedConfig) {
// Set up role-specific behaviors
switch(config.role) {
case 'reader':
this.storage.setReadOnly(true);
this.enableAggressiveCaching();
this.subscribeToIndexUpdates();
break;
case 'writer':
this.storage.setWriteOnly(true);
this.enableWriteBatching();
this.publishIndexUpdates();
break;
case 'hybrid':
this.enableCoordinatedAccess();
break;
}
}
}
```
### 2. S3 Coordination Layer
**Current State**: Direct S3 operations without coordination.
**Proposed Enhancement**:
```typescript
// src/storage/s3Coordinator.ts
export class S3Coordinator {
private manifestPath = '_brainy/manifest.json';
private lockPrefix = '_brainy/locks/';
async acquireWriteLock(partition: string): Promise<LockHandle> {
const lockKey = `${this.lockPrefix}${partition}`;
const lockId = `${this.instanceId}-${Date.now()}`;
// Use S3's conditional PUT for atomic lock acquisition
try {
await this.s3.putObject({
Bucket: this.bucket,
Key: lockKey,
Body: JSON.stringify({
owner: this.instanceId,
acquired: Date.now(),
ttl: 30000 // 30 second TTL
}),
Condition: 'ObjectDoesNotExist'
});
return new LockHandle(lockId, () => this.releaseLock(lockKey));
} catch (err) {
if (err.code === 'PreconditionFailed') {
throw new LockAcquisitionError(`Partition ${partition} is locked`);
}
throw err;
}
}
async updateManifest(update: ManifestUpdate) {
// Atomic manifest updates using versioning
const manifest = await this.getManifest();
manifest.version++;
manifest.lastUpdate = Date.now();
manifest.updates.push(update);
await this.s3.putObject({
Bucket: this.bucket,
Key: this.manifestPath,
Body: JSON.stringify(manifest),
Metadata: {
'version': manifest.version.toString()
}
});
// Notify other instances
await this.broadcastUpdate(update);
}
}
```
### 3. Partition Assignment Strategy
**Current State**: All instances access all partitions.
**Proposed Enhancement**:
```typescript
// src/partitioning/distributedPartitioner.ts
export class DistributedPartitioner {
private assignments: Map<string, Set<string>> = new Map();
async assignPartitions(instances: Instance[]): Promise<PartitionAssignment> {
const writers = instances.filter(i => i.role === 'writer');
const partitions = await this.getAllPartitions();
// Use consistent hashing for stable assignments
const ring = new ConsistentHashRing(writers.map(w => w.id));
const assignment: PartitionAssignment = {};
for (const partition of partitions) {
const writer = ring.getNode(partition.id);
assignment[writer] = assignment[writer] || [];
assignment[writer].push(partition.id);
}
// Store assignments in S3 for coordination
await this.storeAssignments(assignment);
return assignment;
}
async getMyPartitions(): Promise<string[]> {
const assignments = await this.loadAssignments();
return assignments[this.instanceId] || [];
}
}
```
### 4. Write-Ahead Log for Consistency
**Current State**: Direct writes without transaction log.
**Proposed Enhancement**:
```typescript
// src/wal/writeAheadLog.ts
export class WriteAheadLog {
private logPrefix = '_brainy/wal/';
async logWrite(operation: WriteOperation): Promise<string> {
const logEntry = {
id: uuidv4(),
timestamp: Date.now(),
instanceId: this.instanceId,
operation: operation,
status: 'pending'
};
// Write to WAL first
await this.s3.putObject({
Bucket: this.bucket,
Key: `${this.logPrefix}${logEntry.id}`,
Body: JSON.stringify(logEntry)
});
// Then execute operation
try {
await this.executeOperation(operation);
await this.markComplete(logEntry.id);
} catch (err) {
await this.markFailed(logEntry.id, err);
throw err;
}
return logEntry.id;
}
async recoverFromWAL() {
// On startup, check for incomplete operations
const pendingOps = await this.getPendingOperations();
for (const op of pendingOps) {
if (this.canRecover(op)) {
await this.retryOperation(op);
}
}
}
}
```
### 5. Operational Modes
**Current State**: Single operational mode.
**Proposed Enhancement**:
```typescript
// src/modes/operationalModes.ts
export abstract class OperationalMode {
abstract canRead(): boolean;
abstract canWrite(): boolean;
abstract canDelete(): boolean;
abstract getCacheStrategy(): CacheStrategy;
}
export class ReadOnlyMode extends OperationalMode {
canRead() { return true; }
canWrite() { return false; }
canDelete() { return false; }
getCacheStrategy() {
return {
hotCacheRatio: 0.8, // More memory for cache
prefetchAggressive: true,
ttl: Infinity // Never expire cache in read-only
};
}
}
export class WriteOnlyMode extends OperationalMode {
canRead() { return false; }
canWrite() { return true; }
canDelete() { return true; }
getCacheStrategy() {
return {
hotCacheRatio: 0.2, // Minimal cache, focus on write buffer
writeBuffer: true,
batchWrites: true
};
}
}
export class HybridMode extends OperationalMode {
constructor(private coordinator: Coordinator) {}
canRead() { return true; }
canWrite() { return this.coordinator.hasWriteLock(); }
canDelete() { return this.coordinator.hasWriteLock(); }
getCacheStrategy() {
return {
hotCacheRatio: 0.5,
adaptive: true // Adjust based on workload
};
}
}
```
### 6. Health and Coordination Endpoints
**Current State**: No built-in health/coordination endpoints.
**Proposed Enhancement**:
```typescript
// src/api/coordinationAPI.ts
export class CoordinationAPI {
setupEndpoints(app: Express) {
// Health check with role information
app.get('/health', async (req, res) => {
res.json({
status: 'healthy',
role: this.config.role,
instanceId: this.instanceId,
partitions: await this.getAssignedPartitions(),
metrics: await this.getMetrics()
});
});
// Coordination endpoints
app.post('/coordinate/rebalance', async (req, res) => {
const result = await this.rebalancePartitions();
res.json(result);
});
app.get('/coordinate/assignments', async (req, res) => {
const assignments = await this.getPartitionAssignments();
res.json(assignments);
});
// Sync endpoint for configuration updates
app.post('/sync/config', async (req, res) => {
await this.updateConfiguration(req.body);
res.json({ status: 'updated' });
});
}
}
```
### 7. Event-Driven Coordination
**Current State**: No event system.
**Proposed Enhancement**:
```typescript
// src/events/distributedEvents.ts
export class DistributedEventBus {
private subscribers: Map<string, Set<EventHandler>> = new Map();
async emit(event: BrainyEvent) {
// Local handlers
const handlers = this.subscribers.get(event.type) || new Set();
for (const handler of handlers) {
await handler(event);
}
// Remote broadcast (pluggable backends)
await this.broadcast(event);
}
async broadcast(event: BrainyEvent) {
// S3-based event log (simple, no additional deps)
if (this.config.broadcastMethod === 's3') {
await this.s3.putObject({
Bucket: this.bucket,
Key: `_brainy/events/${Date.now()}-${event.type}`,
Body: JSON.stringify(event)
});
}
// WebSocket broadcast
if (this.config.broadcastMethod === 'websocket') {
this.ws.broadcast(JSON.stringify(event));
}
// Cloud Pub/Sub
if (this.config.broadcastMethod === 'pubsub') {
await this.pubsub.topic('brainy-events').publish(event);
}
}
subscribe(eventType: string, handler: EventHandler) {
if (!this.subscribers.has(eventType)) {
this.subscribers.set(eventType, new Set());
}
this.subscribers.get(eventType).add(handler);
}
}
```
### 8. Configuration Synchronization
**Current State**: Local configuration only.
**Proposed Enhancement**:
```typescript
// src/config/distributedConfig.ts
export class DistributedConfigManager {
private configCache: ConfigCache;
private configVersion: number = 0;
async loadConfig(): Promise<BrainyConfig> {
// Try S3 first for shared config
try {
const s3Config = await this.loadFromS3();
this.configVersion = s3Config.version;
return this.mergeWithLocal(s3Config);
} catch (err) {
// Fall back to local config
return this.loadLocalConfig();
}
}
async watchConfigChanges() {
// Poll S3 for config updates
setInterval(async () => {
const latestVersion = await this.getConfigVersion();
if (latestVersion > this.configVersion) {
const newConfig = await this.loadFromS3();
await this.applyConfigUpdate(newConfig);
this.configVersion = latestVersion;
}
}, 10000); // Check every 10 seconds
}
private async applyConfigUpdate(config: BrainyConfig) {
// Hot-reload configuration without restart
if (config.caching) {
this.cacheManager.updateStrategy(config.caching);
}
if (config.partitioning) {
await this.partitioner.reconfigure(config.partitioning);
}
// Emit event for other components
this.eventBus.emit({
type: 'config_updated',
payload: config
});
}
}
```
## Implementation Priority
### Phase 1: Essential Features (Week 1-2)
1. **Operational Modes** - Enable read-only/write-only modes
2. **S3 Coordination** - Basic locking and manifest management
3. **Configuration Sync** - Shared configuration from S3
### Phase 2: Coordination (Week 3-4)
4. **Partition Assignment** - Distributed partition management
5. **Event System** - Basic event broadcasting
6. **Health Endpoints** - Monitoring and coordination APIs
### Phase 3: Advanced Features (Week 5-6)
7. **Write-Ahead Log** - Consistency and recovery
8. **Advanced Coordination** - WebSocket/Pub-Sub integration
9. **Auto-rebalancing** - Dynamic partition redistribution
## Backwards Compatibility
All enhancements should be optional and backwards compatible:
```typescript
// Default behavior remains unchanged
const brainy = new BrainyData({ /* existing config */ });
// Opt-in to distributed features
const distributedBrainy = new BrainyData({
/* existing config */,
distributed: {
enabled: true,
role: 'reader',
// ... distributed options
}
});
```
## Testing Strategy
### Unit Tests
- Test each operational mode independently
- Mock S3 coordination operations
- Test configuration synchronization
### Integration Tests
- Multi-instance coordination tests
- Partition assignment and rebalancing
- Consistency under concurrent operations
### Load Tests
- Simulate millions of vectors
- Test read/write separation at scale
- Measure coordination overhead
## Performance Impact
Expected performance characteristics:
- **Read-only instances**: 10-20% faster due to optimized caching
- **Write-only instances**: 30-40% higher throughput with batching
- **Coordination overhead**: <100ms for most operations
- **Configuration sync**: <1s propagation delay
## Conclusion
These enhancements would transform Brainy into a truly distributed vector database system capable of handling large-scale deployments with specialized instances. The modular design ensures that simpler use cases remain unaffected while enabling sophisticated distributed architectures when needed.

View file

@ -0,0 +1,494 @@
# Distributed Brainy Deployment: Multi-Instance S3 Architecture
## Scenario Overview
A production deployment with 3 specialized Brainy instances sharing a single S3 bucket as the source of truth:
1. **Search Instance** (Read-Only): High-performance search across millions of vectors
2. **Bluesky Processor** (Write-Only): High-throughput ingestion from Bluesky firehose
3. **GitHub Crawler** (Write-Only): Continuous crawling and indexing of GitHub data
All instances run as Google Cloud Run containers with shared S3 storage.
## Architecture Design
### Shared Configuration Strategy
#### Option 1: Configuration Service (Recommended)
```typescript
// config-service.ts - Deployed as separate Cloud Run service
export class BrainyConfigService {
private s3Config = {
bucket: 'brainy-vectors-prod',
configPath: '_brainy/config.json',
lockPath: '_brainy/config.lock'
};
async getSharedConfig(): Promise<BrainyConfig> {
// Fetch from S3 with caching
return {
hnsw: {
M: 16,
efConstruction: 200,
seed: 42, // Critical: same seed for consistent partitioning
maxElements: 10000000
},
partitioning: {
strategy: 'semantic',
numPartitions: 128, // Must be consistent across instances
replicationFactor: 3,
hashFunction: 'xxhash' // Deterministic partitioning
},
storage: {
compressionLevel: 6,
chunkSize: 1024 * 1024, // 1MB chunks
prefixStrategy: 'date-based' // e.g., /2024/01/15/
},
caching: {
hotCacheSize: '2GB',
warmCacheSize: '8GB',
ttl: 3600
}
};
}
}
```
#### Option 2: S3-Based Config Synchronization
Store configuration in S3 with versioning and atomic updates:
```
s3://brainy-vectors-prod/
_brainy/
config.json # Shared configuration
schema.json # Vector schema definition
partitions.json # Partition mapping
instances/
search-001.json # Instance-specific overrides
bluesky-001.json
github-001.json
```
### Instance-Specific Configurations
#### 1. Search Instance (Read-Only)
```typescript
const searchConfig = {
...sharedConfig,
mode: 'read-only',
caching: {
hotCacheSize: '8GB', // Maximize cache for search
warmCacheSize: '32GB',
prefetchStrategy: 'aggressive',
bloomFilters: true // Fast negative lookups
},
hnsw: {
...sharedConfig.hnsw,
efSearch: 100, // Higher for better recall
useMmap: true // Memory-mapped files for large indices
},
s3: {
readConcurrency: 20, // High parallelism for reads
useTransferAcceleration: true,
cacheHeaders: true
},
monitoring: {
metrics: ['latency', 'recall', 'cache_hit_rate']
}
};
```
#### 2. Bluesky Processor (Write-Only)
```typescript
const blueksyConfig = {
...sharedConfig,
mode: 'write-only',
batching: {
size: 10000, // Large batches for throughput
flushInterval: 5000, // 5 seconds
parallelWrites: 4
},
deduplication: {
enabled: true,
bloomFilter: true,
windowSize: 1000000 // Check last 1M entries
},
s3: {
writeConcurrency: 10,
multipartThreshold: 50 * 1024 * 1024, // 50MB
useServerSideEncryption: true
},
indexing: {
async: true, // Don't wait for index updates
batchIndexUpdates: true
}
};
```
#### 3. GitHub Crawler (Write-Only)
```typescript
const githubConfig = {
...sharedConfig,
mode: 'write-only',
rateLimit: {
requestsPerSecond: 10, // Respect API limits
burstSize: 20
},
batching: {
size: 1000, // Smaller batches, continuous flow
flushInterval: 10000 // 10 seconds
},
embedding: {
model: 'text-embedding-3-small',
batchSize: 100,
cacheEmbeddings: true
},
s3: {
writeConcurrency: 5,
retryStrategy: 'exponential'
}
};
```
## Synchronization Mechanisms
### 1. Partition Coordinator Service
Deploy a lightweight coordinator that manages partition assignments:
```typescript
class PartitionCoordinator {
private websocket: WebSocketServer;
async assignPartition(instanceId: string, mode: 'read' | 'write') {
if (mode === 'write') {
// Ensure no partition is assigned to multiple writers
return this.getExclusivePartition(instanceId);
} else {
// Readers can access all partitions
return 'all';
}
}
async rebalance() {
// Triggered when instances join/leave
// Ensures even distribution of write load
}
}
```
### 2. Event Broadcasting via Pub/Sub
Use Google Cloud Pub/Sub for coordination:
```typescript
interface BrainyEvent {
type: 'partition_created' | 'index_updated' | 'config_changed';
timestamp: number;
payload: any;
}
// Writers publish events
await pubsub.topic('brainy-events').publish({
type: 'partition_created',
payload: { partitionId: 'p-123', vectorCount: 50000 }
});
// Readers subscribe and update local state
subscription.on('message', (message) => {
if (message.type === 'index_updated') {
await this.refreshLocalIndex(message.payload.partitionId);
}
});
```
## Performance Optimizations
### 1. Write Path Optimization
```typescript
// Parallel partition writes
class PartitionedWriter {
async write(vectors: Vector[]) {
const partitioned = this.partitionVectors(vectors);
await Promise.all(
Object.entries(partitioned).map(([partitionId, vecs]) =>
this.writeToPartition(partitionId, vecs)
)
);
}
private partitionVectors(vectors: Vector[]) {
// Use consistent hash to determine partition
return vectors.reduce((acc, vec) => {
const partition = hashToPartition(vec.id);
acc[partition] = acc[partition] || [];
acc[partition].push(vec);
return acc;
}, {});
}
}
```
### 2. Read Path Optimization
```typescript
// Distributed search with result aggregation
class DistributedSearch {
async search(query: Vector, k: number) {
// Identify relevant partitions using routing table
const partitions = await this.getRelevantPartitions(query);
// Parallel search across partitions
const results = await Promise.all(
partitions.map(p => this.searchPartition(p, query, k * 2))
);
// Merge and re-rank results
return this.mergeResults(results, k);
}
}
```
### 3. S3 Optimization Strategies
```typescript
const s3Optimizations = {
// Use S3 Transfer Acceleration for cross-region
transferAcceleration: true,
// Intelligent prefixing for parallel reads
prefixSharding: {
enabled: true,
shardCount: 16, // Distribute across 16 prefixes
strategy: 'hash' // or 'round-robin'
},
// Batch operations
batchOperations: {
getObject: 100, // Batch up to 100 GETs
putObject: 50 // Batch up to 50 PUTs
},
// Caching strategy
caching: {
cloudFront: true, // Use CDN for read-heavy workloads
s3CacheControl: 'max-age=3600'
}
};
```
## Suggested Brainy Enhancements
### 1. Native Distributed Mode
```typescript
// Proposed API
const brainy = new BrainyData({
distributed: {
mode: 'cluster',
role: 'writer' | 'reader' | 'hybrid',
coordinator: 'redis://coordinator:6379',
instanceId: process.env.INSTANCE_ID
}
});
```
### 2. S3 Lock Manager
```typescript
class S3LockManager {
async acquireLock(resource: string, ttl: number) {
// Use S3 conditional puts for distributed locking
const lockKey = `_locks/${resource}`;
const lockValue = `${this.instanceId}-${Date.now()}`;
try {
await s3.putObject({
Bucket: this.bucket,
Key: lockKey,
Body: lockValue,
Metadata: { ttl: ttl.toString() },
Condition: 'ObjectDoesNotExist'
});
return true;
} catch (err) {
if (err.code === 'PreconditionFailed') {
return false; // Lock already held
}
throw err;
}
}
}
```
### 3. Partition Discovery Service
```typescript
class PartitionDiscovery {
private cache = new Map();
async discoverPartitions(): Promise<Partition[]> {
// List S3 prefixes to discover partitions
const result = await s3.listObjectsV2({
Bucket: this.bucket,
Prefix: 'partitions/',
Delimiter: '/'
});
return result.CommonPrefixes.map(prefix => ({
id: prefix.Prefix.split('/')[1],
metadata: await this.getPartitionMetadata(prefix.Prefix)
}));
}
subscribeToChanges(callback: (event: PartitionEvent) => void) {
// Watch S3 events or use SNS/EventBridge
}
}
```
### 4. Consistency Manager
```typescript
class ConsistencyManager {
async ensureConsistency() {
// Periodic consistency checks
const tasks = [
this.verifyPartitionIntegrity(),
this.checkIndexConsistency(),
this.validateConfiguration()
];
const results = await Promise.all(tasks);
if (results.some(r => !r.valid)) {
await this.triggerRepair();
}
}
}
```
## Deployment Configuration
### Cloud Run Service Definitions
```yaml
# search-service.yaml
apiVersion: serving.knative.dev/v1
kind: Service
metadata:
name: brainy-search
spec:
template:
spec:
containers:
- image: gcr.io/project/brainy-search:latest
env:
- name: BRAINY_MODE
value: "read-only"
- name: BRAINY_ROLE
value: "search"
resources:
limits:
cpu: "4"
memory: "16Gi"
startupProbe:
httpGet:
path: /health
initialDelaySeconds: 30
# bluesky-processor.yaml
apiVersion: serving.knative.dev/v1
kind: Service
metadata:
name: brainy-bluesky
spec:
template:
spec:
containers:
- image: gcr.io/project/brainy-bluesky:latest
env:
- name: BRAINY_MODE
value: "write-only"
- name: BRAINY_ROLE
value: "bluesky-processor"
resources:
limits:
cpu: "2"
memory: "8Gi"
```
### Environment Variables
```bash
# Common to all instances
BRAINY_S3_BUCKET=brainy-vectors-prod
BRAINY_S3_REGION=us-central1
BRAINY_CONFIG_SOURCE=s3
BRAINY_CONFIG_PATH=_brainy/config.json
# Instance-specific
BRAINY_INSTANCE_ID=${K_SERVICE}-${K_REVISION}
BRAINY_ROLE=search|bluesky|github
BRAINY_MODE=read-only|write-only
```
## Monitoring and Operations
### Key Metrics to Track
1. **Search Instance**
- Query latency (p50, p95, p99)
- Cache hit ratio
- Concurrent searches
- S3 GET requests/sec
2. **Write Instances**
- Ingestion rate (vectors/sec)
- Batch size and latency
- S3 PUT requests/sec
- Partition distribution
3. **System-Wide**
- Total vector count
- Partition count and size distribution
- S3 storage usage and costs
- Cross-instance consistency lag
### Health Checks
```typescript
app.get('/health', async (req, res) => {
const health = {
instance: process.env.BRAINY_INSTANCE_ID,
role: process.env.BRAINY_ROLE,
status: 'healthy',
checks: {
s3_connectivity: await checkS3(),
config_loaded: await checkConfig(),
partition_access: await checkPartitions(),
memory_usage: process.memoryUsage()
}
};
res.json(health);
});
```
## Cost Optimization
1. **S3 Intelligent Tiering**: Automatically move cold partitions to cheaper storage classes
2. **Request Batching**: Minimize S3 API calls through batching
3. **CDN for Reads**: Use Cloud CDN for frequently accessed partitions
4. **Lifecycle Policies**: Auto-delete old snapshots and temporary data
5. **Reserved Capacity**: Use committed use discounts for Cloud Run
## Implementation Timeline
### Phase 1: Basic Setup (Week 1)
- Deploy 3 instances with shared S3 bucket
- Implement basic configuration synchronization
- Set up monitoring
### Phase 2: Optimization (Week 2-3)
- Implement partition coordinator
- Add caching layers
- Optimize S3 operations
### Phase 3: Advanced Features (Week 4+)
- Add WebSocket-based coordination
- Implement consistency checks
- Add auto-scaling based on load
## Conclusion
This architecture provides a scalable, distributed Brainy deployment that can handle millions of vectors with specialized instances for different workloads. The key is maintaining consistency through shared configuration and coordination while optimizing each instance for its specific role.