Data Flow

Detailed walkthrough of how market data flows through NanoARB from CME packet to order execution

Overview

This document traces a single CME market data packet through the NanoARB system, from raw UDP bytes to a trading decision and order submission. Understanding this data flow is critical for:

• Performance optimization - Identifying latency bottlenecks
• Debugging - Tracing data transformations
• Strategy development - Knowing when your code gets called
• Feature extraction - Understanding available data at each stage

End-to-End Latency Budget

From README.md (lines 221-228), the complete tick-to-trade latency:

┌────────────────────────────────────────────────────────────┐
│                  Tick-to-Trade: 780ns (median)             │
├────────────────────────────────────────────────────────────┤
│  1. LOB Update             45ns  [███░░░░░░░░░░░░░░]       │
│  2. Feature Extraction    120ns  [████████░░░░░░░░░]       │
│  3. Model Inference       580ns  [███████████████████]     │
│  4. Signal Generation      35ns  [██░░░░░░░░░░░░░░░]       │
└────────────────────────────────────────────────────────────┘

                    P95: 950ns    P99: 1.2μs

Let’s examine each stage in detail.

---

Stage 1: Market Data Ingestion

1.1 UDP Packet Reception

Where: Network interface (kernel/user space)

What happens:

1. CME multicast packet arrives on configured interface (e.g., eth0)
2. Kernel copies packet to socket buffer
3. Application reads packet via tokio::net::UdpSocket

Data format: SBE (Simple Binary Encoding) binary protocol

CME MDP 3.0 Packet Structure:
┌─────────────────────────────────────────┐
│ Packet Header (12 bytes)                │
├─────────────────────────────────────────┤
│ - Sequence Number (4 bytes)             │
│ - Sending Time (8 bytes)                │
├─────────────────────────────────────────┤
│ Message Header (varies)                 │
├─────────────────────────────────────────┤
│ - Block Length (2 bytes)                │
│ - Template ID (2 bytes)                 │
│ - Schema ID (2 bytes)                   │
│ - Version (2 bytes)                     │
├─────────────────────────────────────────┤
│ Message Body (varies by template)       │
│ - MDIncrementalRefreshBook (Template 46)│
│   - Security ID                         │
│   - RptSeq                              │
│   - Num Entries                         │
│   - Entry: [Price, Qty, Side, Action]   │
└─────────────────────────────────────────┘

1.2 SBE Decoding

Where: nano-feed/src/parser.rs:1

Function: MdpParser::parse()

impl MdpParser {
    pub fn parse(&mut self, data: &[u8]) -> FeedResult<MdpMessage> {
        // Parse header
        let (remaining, header) = parse_packet_header(data)?;
        
        // Match on template ID
        match header.template_id {
            46 => parse_book_update(remaining),
            42 => parse_trade(remaining),
            4 => parse_channel_reset(remaining),
            _ => Err(FeedError::UnsupportedTemplate),
        }
    }
}

Key optimization: Zero-copy parsing with nom combinators

// Example nom parser for price field
fn parse_price(input: &[u8]) -> IResult<&[u8], Price> {
    map(le_i64, |raw| Price::from_raw(raw))(input)
}

Output: MdpMessage::BookUpdate

pub struct BookUpdate {
    pub instrument_id: u32,
    pub timestamp: Timestamp,
    pub entries: Vec<BookEntry>,
}

pub struct BookEntry {
    pub price: Price,
    pub quantity: Quantity,
    pub side: Side,
    pub action: UpdateAction,  // Add, Delete, Change
}

Latency: ~20-30ns for typical 46-byte message

---

Stage 2: Order Book Update

2.1 Book Reconstruction

Where: nano-lob/src/orderbook.rs:1

Function: OrderBook::apply_update()

impl OrderBook {
    pub fn apply_update(&mut self, update: &BookUpdate) {
        for entry in &update.entries {
            match entry.action {
                UpdateAction::Add | UpdateAction::Change => {
                    let book_side = if entry.side == Side::Buy {
                        &mut self.bids
                    } else {
                        &mut self.asks
                    };
                    book_side.insert(entry.price, entry.quantity);
                }
                UpdateAction::Delete => {
                    let book_side = if entry.side == Side::Buy {
                        &mut self.bids
                    } else {
                        &mut self.asks
                    };
                    book_side.remove(&entry.price);
                }
            }
        }
        self.timestamp = update.timestamp;
    }
}

Data structure: BTreeMap<Price, Quantity>

Why BTreeMap?

• O(log n) insert/delete
• Ordered iteration (best bid/ask at min/max)
• Cache-friendly for small sizes

Book state after update:

Bids (descending):                 Asks (ascending):
┌─────────┬─────────┐             ┌─────────┬─────────┐
│  Price  │   Qty   │             │  Price  │   Qty   │
├─────────┼─────────┤             ├─────────┼─────────┤
│ 5000.25 │   100   │ ← Best bid  │ 5000.50 │   150   │ ← Best ask
│ 5000.00 │   250   │             │ 5000.75 │   200   │
│ 4999.75 │   180   │             │ 5001.00 │   120   │
│ 4999.50 │   300   │             │ 5001.25 │    80   │
└─────────┴─────────┘             └─────────┴─────────┘

Mid price: (5000.25 + 5000.50) / 2 = 5000.375
Spread: 5000.50 - 5000.25 = 0.25 (1 tick)

Latency: 45ns median (P95: 62ns)

2.2 Snapshot Capture

Where: nano-lob/src/snapshot.rs:1

Function: SnapshotRingBuffer::push()

pub struct LobSnapshot {
    pub timestamp: Timestamp,
    pub bids: [(Price, Quantity); 20],
    pub asks: [(Price, Quantity); 20],
}

impl SnapshotRingBuffer {
    pub fn push(&mut self, snapshot: LobSnapshot) {
        self.snapshots[self.cursor] = snapshot;
        self.cursor = (self.cursor + 1) % self.capacity;
    }
}

Purpose: Maintain 100-tick history for sequence model input

Memory layout: Stack-allocated, cache-friendly

Ring Buffer (capacity=100):
┌───┬───┬───┬───┬───┬───┬───┬───┐
│ 0 │ 1 │ 2 │...│ 98│ 99│   │   │
└───┴───┴───┴───┴───┴───┴───┴───┘
         ↑
      cursor

---

Stage 3: Feature Extraction

3.1 LOB Features

Where: nano-lob/src/features.rs:1

Function: LobFeatureExtractor::extract_all()

impl LobFeatureExtractor {
    pub fn extract_all(book: &OrderBook) -> Array1<f64> {
        let mut features = Array1::zeros(40);
        
        // Level 1-5 features (20 values)
        for i in 0..5 {
            if let Some((bid_p, bid_q)) = book.bid_at_level(i) {
                features[i * 2] = bid_p.as_f64();
                features[i * 2 + 1] = bid_q.0 as f64;
            }
            if let Some((ask_p, ask_q)) = book.ask_at_level(i) {
                features[10 + i * 2] = ask_p.as_f64();
                features[10 + i * 2 + 1] = ask_q.0 as f64;
            }
        }
        
        // Derived features (20 values)
        features[20] = Self::microprice(book);
        features[21] = Self::book_imbalance(book, 5);
        features[22] = Self::spread(book);
        features[23] = Self::mid_price_return(book);
        // ... more features
        
        features
    }
}

Feature vector (40 dimensions):

Index  Feature                      Example Value
───────────────────────────────────────────────────
0-1    Bid level 0 (price, qty)     [5000.25, 100]
2-3    Bid level 1                  [5000.00, 250]
4-5    Bid level 2                  [4999.75, 180]
6-7    Bid level 3                  [4999.50, 300]
8-9    Bid level 4                  [4999.25, 220]
10-11  Ask level 0 (price, qty)     [5000.50, 150]
12-13  Ask level 1                  [5000.75, 200]
14-15  Ask level 2                  [5001.00, 120]
16-17  Ask level 3                  [5001.25,  80]
18-19  Ask level 4                  [5001.50, 100]
20     Microprice                    5000.38
21     Book imbalance               -0.12
22     Spread (bps)                  5.0
23     Mid price return              0.0002
24     Order flow imbalance          15.0
25     VPIN                          0.35
26-39  Derived features             ...

3.2 Key Feature Calculations

Microprice

Definition: Volume-weighted mid price

pub fn microprice(book: &OrderBook) -> f64 {
    let (bid_p, bid_q) = book.best_bid().unwrap();
    let (ask_p, ask_q) = book.best_ask().unwrap();
    
    let bid_p = bid_p.as_f64();
    let ask_p = ask_p.as_f64();
    let bid_q = bid_q.0 as f64;
    let ask_q = ask_q.0 as f64;
    
    (bid_p * ask_q + ask_p * bid_q) / (bid_q + ask_q)
}

Example:

Bid: $5000.25 @ 100
Ask: $5000.50 @ 150
Microprice = (5000.25 * 150 + 5000.50 * 100) / (100 + 150)
           = (750037.5 + 500050) / 250
           = 5000.35

Order Flow Imbalance (OFI)

Definition: Net change in bid vs ask volume

pub fn order_flow_imbalance(current: &OrderBook, previous: &OrderBook) -> f64 {
    let bid_delta = current.bid_depth(5) - previous.bid_depth(5);
    let ask_delta = current.ask_depth(5) - previous.ask_depth(5);
    
    (bid_delta as f64) - (ask_delta as f64)
}

Interpretation:

• OFI > 0: More buying pressure (bullish)
• OFI < 0: More selling pressure (bearish)

VPIN (Volume-Synchronized Probability of Informed Trading)

Definition: Imbalance in signed volume

pub fn vpin(snapshots: &[LobSnapshot]) -> f64 {
    let mut buy_volume = 0.0;
    let mut sell_volume = 0.0;
    
    for i in 1..snapshots.len() {
        let prev = &snapshots[i - 1];
        let curr = &snapshots[i];
        
        if curr.mid_price() > prev.mid_price() {
            buy_volume += curr.volume();
        } else {
            sell_volume += curr.volume();
        }
    }
    
    (buy_volume - sell_volume).abs() / (buy_volume + sell_volume)
}

Range: [0, 1], higher values indicate more informed trading

Latency for all features: 120ns median (P95: 145ns)

---

Stage 4: ML Model Inference

4.1 Input Preparation

Where: nano-model/src/lib.rs

Function: SignalModel::predict()

Input shape: (batch=1, seq_len=100, features=40)

let snapshots = ring_buffer.as_tensor();  // (100, 40)
let input = snapshots.insert_axis(Axis(0));  // (1, 100, 40)

Tensor layout:

Batch dimension (1):
  |
  v
┌───────────────────────────────────────────────────┐
│ Sequence (100 timesteps)                          │
│ ┌──────────────────────────────────────────────┐  │
│ │ Features (40 values per timestep)            │  │
│ │ [bid_0_p, bid_0_q, ..., microprice, ...]     │  │
│ ├──────────────────────────────────────────────┤  │
│ │ [bid_0_p, bid_0_q, ..., microprice, ...]     │  │
│ ├──────────────────────────────────────────────┤  │
│ │                    ...                       │  │
│ └──────────────────────────────────────────────┘  │
└───────────────────────────────────────────────────┘

4.2 Mamba Model Architecture

From README.md (lines 295-323):

Input: (1, 100, 40)
   |
   v
┌─────────────────────────┐
│  Linear Projection      │  → (1, 100, 128)
│  + LayerNorm            │
└────────┬────────────────┘
         |
         v
┌─────────────────────────┐
│  Mamba Block 1          │
│  - Conv1D (kernel=4)    │  SSM with selective state
│  - SSM (S4)             │  space mechanism
│  - SiLU Gating          │
└────────┬────────────────┘
         |
         v
┌─────────────────────────┐
│  Mamba Block 2          │
└────────┬────────────────┘
         |
         v
┌─────────────────────────┐
│  Mamba Block 3          │
└────────┬────────────────┘
         |
         v
┌─────────────────────────┐
│  Mamba Block 4          │
└────────┬────────────────┘
         |
         v
┌─────────────────────────┐
│  Output Heads           │
│  - Horizon 1: 1-tick    │  → (1, 3, 3)
│  - Horizon 2: 5-tick    │     [up, flat, down]
│  - Horizon 3: 10-tick   │     probabilities
└─────────────────────────┘

Why Mamba?

• 10-50x faster than Transformers (no quadratic attention)
• Better at long sequences (100+ timesteps)
• State space models capture temporal dynamics
• Sub-microsecond inference

4.3 Output Interpretation

Output shape: (1, 3, 3) = (batch, horizons, classes)

Horizon 1 (1-tick, ~100ms):
┌──────┬───────┬───────┐
│  Up  │ Flat  │ Down  │
├──────┼───────┼───────┤
│ 0.65 │ 0.20  │ 0.15  │ ← Softmax probabilities
└──────┴───────┴───────┘

Horizon 2 (5-tick, ~500ms):
┌──────┬───────┬───────┐
│  Up  │ Flat  │ Down  │
├──────┼───────┼───────┤
│ 0.55 │ 0.25  │ 0.20  │
└──────┴───────┴───────┘

Horizon 3 (10-tick, ~1s):
┌──────┬───────┬───────┐
│  Up  │ Flat  │ Down  │
├──────┼───────┼───────┤
│ 0.45 │ 0.30  │ 0.25  │
└──────┴───────┴───────┘

Signal extraction:

let prediction = output[[0, 0, 0]] - output[[0, 0, 2]];  // P(up) - P(down)
let confidence = output[[0, 0, 0]].max(output[[0, 0, 2]]);  // Max probability

Latency: 580ns median (P95: 720ns, P99: 890ns)

---

Stage 5: Strategy Decision

5.1 Market Maker Quote Calculation

Where: nano-strategy/src/market_maker.rs:1

Function: MarketMakerStrategy::on_market_data()

impl Strategy for MarketMakerStrategy {
    fn on_market_data(&mut self, book: &dyn OrderBook) -> Vec<Order> {
        let mid = book.mid_price().unwrap();
        let tick_size = Price::from_ticks(1, 25);
        
        // 1. Base spread from config
        let base_spread = tick_size * self.config.base_spread_ticks;
        
        // 2. ML signal adjustment
        let signal_skew = if self.config.use_ml_signal {
            self.get_ml_signal() * tick_size * 2
        } else {
            Price::ZERO
        };
        
        // 3. Inventory skew
        let inventory_skew = Price::from_f64(
            self.config.skew_factor * self.position as f64 * tick_size.as_f64()
        );
        
        // 4. Calculate quotes
        let bid_price = mid - base_spread / 2 + signal_skew - inventory_skew;
        let ask_price = mid + base_spread / 2 + signal_skew - inventory_skew;
        
        vec![
            Order {
                id: OrderId::new(),
                instrument_id: self.instrument_id,
                side: Side::Buy,
                price: bid_price,
                quantity: Quantity::new(self.config.order_size),
                timestamp: Timestamp::now(),
                order_type: OrderType::Limit,
            },
            Order {
                id: OrderId::new(),
                instrument_id: self.instrument_id,
                side: Side::Sell,
                price: ask_price,
                quantity: Quantity::new(self.config.order_size),
                timestamp: Timestamp::now(),
                order_type: OrderType::Limit,
            },
        ]
    }
}

Example calculation:

Inputs:
  Mid price: $5000.375
  Base spread: 2 ticks = $0.50
  ML signal: +0.65 (bullish) → +2 ticks = $0.50
  Position: +10 contracts
  Skew factor: 0.5
  Inventory skew: 0.5 * 10 * $0.25 = $1.25

Calculation:
  Bid: 5000.375 - 0.25 + 0.50 - 1.25 = 5000.375
  Ask: 5000.375 + 0.25 + 0.50 - 1.25 = 5000.875

Orders:
  Buy  5 @ $5000.375
  Sell 5 @ $5000.875

Interpretation:

• ML signal is bullish, so both quotes shifted up by $0.50
• Long position (+10), so quotes shifted down by $1.25 to reduce inventory
• Net effect: Willing to sell at higher price, buy at fair value

Latency: 35ns (pure computation, no I/O)

5.2 Risk Checks

Where: nano-backtest/src/risk.rs:1

Function: RiskManager::check_order()

impl RiskManager {
    pub fn check_order(&self, order: &Order, current_position: i64) -> Result<()> {
        // Position limit
        let new_position = current_position + order.signed_quantity();
        if new_position.abs() > self.config.max_position {
            return Err(Error::PositionLimitExceeded);
        }
        
        // Order size limit
        if order.quantity.0 > self.config.max_order_size {
            return Err(Error::OrderSizeTooLarge);
        }
        
        // Drawdown limit
        let current_dd = (self.peak_pnl - self.current_pnl) / self.peak_pnl.abs();
        if current_dd > self.config.max_drawdown_pct {
            return Err(Error::DrawdownLimitBreached);
        }
        
        Ok(())
    }
}

Checks performed:

1. Position limit (e.g., max ±50 contracts)
2. Order size (e.g., max 10 per order)
3. Drawdown threshold (e.g., max 6% from peak)
4. Daily loss limit (e.g., max $100k per day)

If any check fails, order is rejected and strategy is notified.

---

Stage 6: Order Submission

6.1 Event Scheduling

Where: nano-backtest/src/engine.rs:215

Function: BacktestEngine::on_market_data()

for order in orders {
    // Check risk
    if let Err(e) = self.risk.check_order(&order, position) {
        tracing::warn!("Order rejected by risk: {}", e);
        continue;
    }
    
    // Schedule order with latency
    let arrival_time = self.latency.order_arrival_time(self.current_time);
    self.schedule_event(arrival_time, EventType::OrderSubmit { order });
}

Latency simulation:

impl LatencySimulator {
    pub fn order_arrival_time(&self, current_time: Timestamp) -> Timestamp {
        let latency_ns = self.order_latency_ns
            + thread_rng().gen_range(0..self.jitter_ns);
        current_time + Duration::from_nanos(latency_ns)
    }
}

Example:

Current time:    T₀ = 1,000,000,000 ns
Order latency:   100,000 ns (100 μs)
Jitter:          ±10,000 ns (±10 μs)
Arrival time:    T₀ + 105,234 ns = 1,000,105,234 ns

6.2 Exchange Matching

Where: nano-backtest/src/execution.rs:1

Function: SimulatedExchange::submit_order()

impl SimulatedExchange {
    pub fn submit_order(&mut self, order: Order, timestamp: Timestamp) -> OrderId {
        let order_id = order.id;
        
        // Add to active orders
        self.active_orders.insert(order_id, order.clone());
        
        // Calculate fee
        let fee = self.fee_model.calculate_fee(
            order.price,
            order.quantity,
            true,  // Maker (assuming limit order)
            order.side,
        );
        
        // Estimate queue position
        let queue_pos = self.estimate_queue_position(&order);
        self.queue_positions.insert(order_id, queue_pos);
        
        order_id
    }
}

Queue position estimation:

fn estimate_queue_position(&self, order: &Order) -> usize {
    // Orders behind existing volume at same price level
    let existing_qty = self.book.quantity_at_price(order.price, order.side);
    existing_qty.0 as usize
}

Fill simulation:

pub fn match_orders(&mut self, book: &OrderBook, timestamp: Timestamp) -> Vec<Fill> {
    let mut fills = vec![];
    
    for (order_id, order) in &self.active_orders {
        // Check if price crossed
        let should_fill = match order.side {
            Side::Buy => {
                book.best_ask().map_or(false, |(ask, _)| order.price >= ask)
            }
            Side::Sell => {
                book.best_bid().map_or(false, |(bid, _)| order.price <= bid)
            }
        };
        
        if should_fill {
            // Estimate fill probability based on queue position
            let queue_pos = self.queue_positions[order_id];
            let level_qty = book.quantity_at_price(order.price, order.side).0 as usize;
            let fill_prob = self.fill_model.fill_probability(queue_pos, level_qty);
            
            if thread_rng().gen::<f64>() < fill_prob {
                fills.push(Fill {
                    order_id: *order_id,
                    price: order.price,
                    quantity: order.quantity,
                    side: order.side,
                    timestamp,
                    is_maker: true,
                });
            }
        }
    }
    
    fills
}

6.3 Fill Notification

Where: nano-backtest/src/engine.rs:240

fn on_fill<S: Strategy>(&mut self, fill: Fill, strategy: &mut S) {
    // Update position
    self.positions.apply_fill_for_instrument(self.instrument_id, &fill);
    
    // Record metrics
    self.metrics.record_fill(&fill);
    
    // Notify strategy
    strategy.on_fill(&fill);
    
    // Update equity curve
    let total_pnl = self.positions.total_pnl(&self.current_prices);
    self.stats.add_equity_point(self.current_time.as_nanos(), total_pnl);
}

Position update:

impl PositionTracker {
    pub fn apply_fill(&mut self, instrument_id: u32, fill: &Fill) {
        let signed_qty = match fill.side {
            Side::Buy => fill.quantity.0 as i64,
            Side::Sell => -(fill.quantity.0 as i64),
        };
        
        let position = self.positions.entry(instrument_id).or_insert(0);
        let prev_position = *position;
        *position += signed_qty;
        
        // Update realized P&L if crossing zero
        if prev_position.signum() != position.signum() {
            let closed_qty = prev_position.abs().min(signed_qty.abs());
            let avg_entry = self.avg_entry_prices[&instrument_id];
            let pnl = (fill.price.as_f64() - avg_entry) * closed_qty as f64;
            self.realized_pnl += pnl;
        }
    }
}

---

Complete Event Timeline

Putting it all together, here’s a complete tick-to-trade timeline:

T₀:        Market data packet arrives
           ├─ UDP reception: kernel → userspace
           └─ Event: MarketData scheduled at T₀

T₀+0ns:    EventQueue pops MarketData event
           └─ BacktestEngine::on_market_data()

T₀+20ns:   SBE decoding complete
           └─ BookUpdate created

T₀+45ns:   Order book updated
           └─ Snapshot captured to ring buffer

T₀+165ns:  Feature extraction complete (45 + 120)
           └─ Array1<f64> with 40 features

T₀+745ns:  ML inference complete (165 + 580)
           └─ Prediction: [0.65, 0.20, 0.15]

T₀+780ns:  Strategy decision complete (745 + 35)
           └─ Orders: [Buy 5 @ 5000.375, Sell 5 @ 5000.875]

T₀+780ns:  Risk checks pass
           └─ Orders scheduled with latency

T₀+100μs:  Event: OrderSubmit for buy order
           └─ SimulatedExchange::submit_order()
           └─ Queue position: 50
           └─ Event: OrderAck scheduled at T₀+200μs

T₀+100μs:  Event: OrderSubmit for sell order
           └─ Similar processing

T₀+200μs:  Event: OrderAck for both orders
           └─ Strategy::on_order_ack() called

T₀+500μs:  Next market data tick arrives
           └─ Book moves, price crosses our buy order
           └─ Fill probability: 80% (queue cleared)
           └─ Random draw: 0.65 < 0.80 → FILLED
           └─ Event: OrderFill scheduled at T₀+550μs

T₀+550μs:  Event: OrderFill
           └─ Position: +5 contracts
           └─ Realized P&L: $0 (no close)
           └─ Strategy::on_fill() called

Total latency from tick to order: 780ns (within budget)

Total latency from tick to acknowledgment: ~100μs (network latency)

Total latency from tick to fill: ~500μs (depends on market)

---

Performance Profiling

Benchmarking Individual Components

Each crate includes Criterion.rs benchmarks:

# Benchmark LOB updates
cargo bench -p nano-lob orderbook

# Benchmark feature extraction
cargo bench -p nano-lob features

# Benchmark model inference
cargo bench -p nano-model inference

Sample output:

orderbook/update        time:   [44.2 ns 45.1 ns 46.3 ns]
features/microprice     time:   [8.4 ns 8.6 ns 8.9 ns]
features/extract_all    time:   [118.7 ns 120.3 ns 122.1 ns]
inference/mamba         time:   [576.2 ns 582.1 ns 589.3 ns]

Flamegraph Generation

For detailed profiling:

# Install profiling tools
cargo install flamegraph

# Run with profiling
cargo flamegraph --bench orderbook

# Output: flamegraph.svg

Typical flamegraph shows 75% of time in ML inference, 15% in feature extraction, 10% in other.

---

Optimization Techniques

1. SIMD Vectorization

Feature extraction uses SIMD for parallel computation:

use std::simd::*;

fn vectorized_book_imbalance(bids: &[f64], asks: &[f64]) -> f64 {
    let bid_vec = f64x4::from_slice(bids);
    let ask_vec = f64x4::from_slice(asks);
    let diff = bid_vec - ask_vec;
    let sum = bid_vec + ask_vec;
    (diff / sum).reduce_sum()
}

2. Lock-Free Data Structures

Crossbeam channels for inter-thread communication:

let (tx, rx) = crossbeam_channel::bounded(1024);

// Market data thread
tx.send(BookUpdate { ... })?;

// Strategy thread
let update = rx.recv()?;

3. Memory Pooling

Pre-allocate objects to avoid allocation overhead:

struct OrderPool {
    pool: Vec<Order>,
    free_list: Vec<usize>,
}

impl OrderPool {
    fn acquire(&mut self) -> &mut Order {
        let idx = self.free_list.pop().unwrap();
        &mut self.pool[idx]
    }
}

4. Compile-Time Optimizations

[profile.release]
opt-level = 3
lto = "thin"
codegen-units = 1
target-cpu = "native"  # Use CPU-specific instructions

---

Monitoring & Debugging

Tracing Events

Structured logging throughout the pipeline:

use tracing::{debug, info, warn, error, span, Level};

let span = span!(Level::DEBUG, "orderbook_update", instrument_id = %id);
let _enter = span.enter();

debug!(entries = entries.len(), "Applying book update");
book.apply_update(&update);
debug!(best_bid = ?book.best_bid(), best_ask = ?book.best_ask(), "Update complete");

Output (JSON format for log aggregation):

{
  "timestamp": "2026-03-04T10:00:00.123456Z",
  "level": "DEBUG",
  "target": "nano_lob::orderbook",
  "span": {"name": "orderbook_update", "instrument_id": 1},
  "fields": {"entries": 3},
  "message": "Applying book update"
}

Metrics Export

Prometheus metrics at http://localhost:9090/metrics:

# HELP nanoarb_lob_update_duration_ns LOB update latency
# TYPE nanoarb_lob_update_duration_ns histogram
nanoarb_lob_update_duration_ns_bucket{le="50"} 7234
nanoarb_lob_update_duration_ns_bucket{le="100"} 9512
nanoarb_lob_update_duration_ns_bucket{le="200"} 9876
nanoarb_lob_update_duration_ns_sum 445678
nanoarb_lob_update_duration_ns_count 10000

# HELP nanoarb_inference_duration_ns Model inference latency
# TYPE nanoarb_inference_duration_ns histogram
nanoarb_inference_duration_ns_bucket{le="500"} 1234
nanoarb_inference_duration_ns_bucket{le="1000"} 8765
nanoarb_inference_duration_ns_sum 5876543
nanoarb_inference_duration_ns_count 10000

---

Next Steps

• Architecture Overview - System design principles
• Crate Reference - Detailed crate documentation
• Strategy Development - Build custom strategies
• Performance Tuning - Optimize latency