9.8 KiB
Universal Data Stream Architecture Audit & Optimization Plan
📊 UNIVERSAL DATA FORMAT SPECIFICATION
Our trading system is built around 5 core timeseries streams that provide a standardized data format to all models:
Core Timeseries (The Sacred 5)
- ETH/USDT Ticks (1s) - Primary trading pair real-time data
- ETH/USDT 1m - Short-term price action and patterns
- ETH/USDT 1h - Medium-term trends and momentum
- ETH/USDT 1d - Long-term market structure
- BTC/USDT Ticks (1s) - Reference asset for correlation analysis
Data Format Structure
@dataclass
class UniversalDataStream:
eth_ticks: np.ndarray # [timestamp, open, high, low, close, volume]
eth_1m: np.ndarray # [timestamp, open, high, low, close, volume]
eth_1h: np.ndarray # [timestamp, open, high, low, close, volume]
eth_1d: np.ndarray # [timestamp, open, high, low, close, volume]
btc_ticks: np.ndarray # [timestamp, open, high, low, close, volume]
timestamp: datetime
metadata: Dict[str, Any]
🏗️ CURRENT ARCHITECTURE COMPONENTS
1. Universal Data Adapter (core/universal_data_adapter.py
)
- Status: ✅ Implemented
- Purpose: Converts any data source into universal 5-timeseries format
- Key Features:
- Format validation
- Data quality assessment
- Model-specific formatting (CNN, RL, Transformer)
- Window size management
- Missing data handling
2. Unified Data Stream (core/unified_data_stream.py
)
- Status: ✅ Implemented with Subscriber Architecture
- Purpose: Central data distribution hub
- Key Features:
- Publisher-Subscriber pattern
- Consumer registration system
- Multi-consumer data distribution
- Performance tracking
- Data caching and buffering
3. Enhanced Orchestrator Integration
- Status: ✅ Implemented
- Purpose: Neural Decision Fusion using universal data
- Key Features:
- NN-driven decision making
- Model prediction fusion
- Market context preparation
- Cross-asset correlation analysis
📈 DATA FLOW MAPPING
Current Data Flow
Data Provider (Binance API)
↓
Universal Data Adapter
↓
Unified Data Stream (Publisher)
↓
┌─────────────────┬─────────────────┬─────────────────┐
│ Dashboard │ Orchestrator │ Models │
│ Subscriber │ Subscriber │ Subscriber │
└─────────────────┴─────────────────┴─────────────────┘
Registered Consumers
- Trading Dashboard - UI data updates (
ticks
,ohlcv
,ui_data
) - Enhanced Orchestrator - NN decision making (
training_data
,ohlcv
) - CNN Models - Pattern recognition (formatted CNN data)
- RL Models - Action learning (state vectors)
- COB Integration - Order book analysis (microstructure data)
🔍 ARCHITECTURE AUDIT FINDINGS
✅ STRENGTHS
- Standardized Format: All models receive consistent data structure
- Publisher-Subscriber: Efficient one-to-many data distribution
- Performance Tracking: Built-in metrics and monitoring
- Multi-Timeframe: Comprehensive temporal view
- Real-time Processing: Live data with proper buffering
⚠️ OPTIMIZATION OPPORTUNITIES
1. Memory Efficiency
- Issue: Multiple data copies across consumers
- Impact: High memory usage with many subscribers
- Solution: Implement shared memory buffers with copy-on-write
2. Processing Latency
- Issue: Sequential processing in some callbacks
- Impact: Delays in real-time decision making
- Solution: Parallel consumer notification with thread pools
3. Data Staleness
- Issue: No real-time freshness validation
- Impact: Models might use outdated data
- Solution: Timestamp-based data validity checks
4. Network Optimization
- Issue: Individual API calls for each timeframe
- Impact: Rate limiting and bandwidth waste
- Solution: Batch requests and intelligent caching
🚀 OPTIMIZATION IMPLEMENTATION PLAN
Phase 1: Memory Optimization
# Implement shared memory data structures
class SharedDataBuffer:
def __init__(self, max_size: int):
self.data = np.zeros((max_size, 6), dtype=np.float32) # OHLCV + timestamp
self.write_index = 0
self.readers = {} # Consumer ID -> last read index
def write(self, new_data: np.ndarray):
# Atomic write operation
self.data[self.write_index] = new_data
self.write_index = (self.write_index + 1) % len(self.data)
def read(self, consumer_id: str, count: int) -> np.ndarray:
# Return data since last read for this consumer
last_read = self.readers.get(consumer_id, 0)
data_slice = self._get_data_slice(last_read, count)
self.readers[consumer_id] = self.write_index
return data_slice
Phase 2: Parallel Processing
# Implement concurrent consumer notification
class ParallelDataDistributor:
def __init__(self, max_workers: int = 4):
self.executor = ThreadPoolExecutor(max_workers=max_workers)
def distribute_to_consumers(self, data_packet: Dict[str, Any]):
futures = []
for consumer in self.active_consumers:
future = self.executor.submit(self._notify_consumer, consumer, data_packet)
futures.append(future)
# Wait for all notifications to complete
for future in as_completed(futures, timeout=0.1):
try:
future.result()
except Exception as e:
logger.warning(f"Consumer notification failed: {e}")
Phase 3: Intelligent Caching
# Implement smart data caching with expiration
class SmartDataCache:
def __init__(self):
self.cache = {}
self.expiry_times = {}
self.hit_count = 0
self.miss_count = 0
def get_data(self, symbol: str, timeframe: str, force_refresh: bool = False) -> np.ndarray:
cache_key = f"{symbol}_{timeframe}"
current_time = time.time()
if not force_refresh and cache_key in self.cache:
if current_time < self.expiry_times[cache_key]:
self.hit_count += 1
return self.cache[cache_key]
# Cache miss - fetch fresh data
self.miss_count += 1
fresh_data = self._fetch_fresh_data(symbol, timeframe)
# Cache with appropriate expiration
self.cache[cache_key] = fresh_data
self.expiry_times[cache_key] = current_time + self._get_cache_duration(timeframe)
return fresh_data
📋 INTEGRATION CHECKLIST
Dashboard Integration
- Verify
web/clean_dashboard.py
uses UnifiedDataStream - Ensure proper subscriber registration
- Check data type requirements (
ui_data
,ohlcv
) - Validate real-time updates
Model Integration
- CNN models receive formatted universal data
- RL models get proper state vectors
- Neural Decision Fusion uses all 5 timeseries
- COB integration processes microstructure data
Performance Monitoring
- Stream statistics tracking
- Consumer performance metrics
- Data quality monitoring
- Memory usage optimization
🎯 IMMEDIATE ACTION ITEMS
High Priority
- Audit Dashboard Subscriber - Ensure
clean_dashboard.py
properly subscribes - Verify Model Data Flow - Check all models receive universal format
- Monitor Memory Usage - Track memory consumption across consumers
- Performance Profiling - Measure data distribution latency
Medium Priority
- Implement Shared Buffers - Reduce memory duplication
- Add Data Freshness Checks - Prevent stale data usage
- Optimize Network Calls - Batch API requests where possible
- Enhanced Error Handling - Graceful degradation on data issues
Low Priority
- Advanced Caching - Predictive data pre-loading
- Compression - Reduce data transfer overhead
- Distributed Processing - Scale across multiple processes
- Real-time Analytics - Live data quality metrics
🔧 IMPLEMENTATION STATUS
✅ Completed
- Universal Data Adapter with 5 timeseries
- Unified Data Stream with subscriber pattern
- Enhanced Orchestrator integration
- Neural Decision Fusion using universal data
🚧 In Progress
- Dashboard subscriber optimization
- Memory usage profiling
- Performance monitoring
📅 Planned
- Shared memory implementation
- Parallel consumer notification
- Advanced caching strategies
- Real-time quality monitoring
📊 SUCCESS METRICS
Performance Targets
- Data Latency: < 10ms from source to consumer
- Memory Efficiency: < 500MB total for all consumers
- Cache Hit Rate: > 80% for historical data requests
- Consumer Throughput: > 100 updates/second per consumer
Quality Targets
- Data Completeness: > 99.9% for all 5 timeseries
- Timestamp Accuracy: < 1ms deviation from source
- Format Compliance: 100% validation success
- Error Rate: < 0.1% failed distributions
🎯 CONCLUSION
The Universal Data Stream architecture is the backbone of our trading system. The 5 timeseries format ensures all models receive consistent, high-quality data. The subscriber architecture enables efficient distribution, but there are clear optimization opportunities for memory usage, processing latency, and caching.
Next Steps: Focus on implementing shared memory buffers and parallel consumer notification to improve performance while maintaining the integrity of our universal data format.
Critical: All optimization work must preserve the 5 timeseries structure as it's fundamental to our model training and decision making processes.