# Concurrency Control Flow

This diagram shows how the semaphore-based concurrency control works across multiple workers.

```mermaid
graph LR
    subgraph "Database Queue"
        Q[Pending Jobs<br/>Priority Queue]
    end
    
    subgraph "Worker-1"
        S1[Semaphore<br/>6 slots]
        J1[Job 1]
        J2[Job 2]
        J3[Job 3]
        J4[Job 4]
        J5[Job 5]
        J6[Job 6]
    end
    
    subgraph "Worker-2"
        S2[Semaphore<br/>6 slots]
        J7[Job 7]
        J8[Job 8]
        J9[Job 9]
        J10[Job 10]
        J11[Job 11]
        J12[Job 12]
    end
    
    subgraph "Worker-3"
        S3[Semaphore<br/>6 slots]
        J13[Job 13]
        J14[Job 14]
        J15[Job 15]
        J16[Job 16]
        J17[Job 17]
        J18[Job 18]
    end
    
    Q -->|Claim 6 jobs| S1
    Q -->|Claim 6 jobs| S2
    Q -->|Claim 6 jobs| S3
    
    S1 --> J1
    S1 --> J2
    S1 --> J3
    S1 --> J4
    S1 --> J5
    S1 --> J6
    
    S2 --> J7
    S2 --> J8
    S2 --> J9
    S2 --> J10
    S2 --> J11
    S2 --> J12
    
    S3 --> J13
    S3 --> J14
    S3 --> J15
    S3 --> J16
    S3 --> J17
    S3 --> J18
    
    style Q fill:#FF6B6B
    style S1 fill:#50C878
    style S2 fill:#50C878
    style S3 fill:#50C878
```

## Concurrency Control Mechanisms

### 1. Database-Level (Advisory Locks)
- **PostgreSQL Advisory Locks**: Prevent multiple workers from claiming the same job
- Atomic job claiming using `pg_try_advisory_lock()`
- Ensures exactly-once job processing

### 2. Worker-Level (Semaphore)
- **SemaphoreSlim**: Limits concurrent backtests per worker
- Default: `Environment.ProcessorCount - 2` (e.g., 6 on 8-core machine)
- Prevents CPU saturation while leaving resources for Orleans messaging

### 3. Cluster-Level (Queue Priority)
- **Priority Queue**: Jobs ordered by priority, then creation time
- VIP users get higher priority
- Fair distribution across workers

## Capacity Calculation

- **Per Worker**: 6 concurrent backtests
- **3 Workers**: 18 concurrent backtests
- **Average Duration**: ~47 minutes per backtest
- **Throughput**: ~1,080 backtests/hour
- **1000 Users × 10 backtests**: ~9 hours to process full queue