ADR-003: CQRS Pattern Implementation

Status	Date	Deciders
Accepted	2024-02-01	Core Team

Context

Event sourcing naturally leads to a separation between how we write data (append events) and how we read data (query projections). This separation aligns with the Command Query Responsibility Segregation (CQRS) pattern.

Key challenges we face:

1. Query Performance: Replaying events for every query is expensive
2. Query Flexibility: Different read use cases need different data shapes
3. Scalability: Read and write workloads have different characteristics
4. Consistency: Need to balance consistency with performance

We need to decide how strictly to implement CQRS and what abstractions to provide.

Decision

We will implement CQRS with clear separation between command and query sides:

Command Side (Write Model)

┌─────────────────────────────────────────────────────────────┐
│                      COMMAND SIDE                           │
│                                                             │
│  Command ──▶ Command Bus ──▶ Handler ──▶ Aggregate ──▶ Events
│                  │                                          │
│                  ▼                                          │
│            Middleware                                       │
│         (Validation, Auth)                                  │
└─────────────────────────────────────────────────────────────┘

Characteristics:

• Optimized for consistency
• Domain logic in aggregates
• Commands validated before execution
• Events stored atomically

Query Side (Read Model)

┌─────────────────────────────────────────────────────────────┐
│                      QUERY SIDE                             │
│                                                             │
│  Events ──▶ Projection Engine ──▶ Read Models ──▶ Queries   │
│                                                             │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐         │
│  │   Inline    │  │    Async    │  │    Live     │         │
│  │ (Immediate) │  │ (Background)│  │ (Real-time) │         │
│  └─────────────┘  └─────────────┘  └─────────────┘         │
└─────────────────────────────────────────────────────────────┘

Characteristics:

• Optimized for query performance
• Multiple views of same data
• Denormalized for specific use cases
• Eventually consistent (configurable)

Implementation

// Command Side
type CreateOrderCommand struct {
    CustomerID string
    Items      []OrderItem
}

func (h *OrderHandler) Handle(ctx context.Context, cmd CreateOrderCommand) (CommandResult, error) {
    order := NewOrder(uuid.New().String())
    if err := order.Create(cmd.CustomerID, cmd.Items); err != nil {
        return CommandResult{}, err
    }
    return h.store.SaveAggregate(ctx, order)
}

// Query Side
type OrderSummaryView struct {
    OrderID     string
    CustomerID  string
    TotalAmount float64
    Status      string
}

func (p *OrderSummaryProjection) Apply(ctx context.Context, event StoredEvent) error {
    switch event.Type {
    case "OrderCreated":
        // Create read model
    case "OrderShipped":
        // Update read model
    }
    return nil
}

// Query
orders, _ := orderRepo.Query(ctx, NewQuery().
    Where("CustomerID", Eq, customerId).
    Where("Status", Eq, "Pending"))

Consequences

Positive

1. Performance Optimization: Each side optimized for its purpose
2. Scalability: Can scale read and write independently
3. Flexibility: Multiple read models for different use cases
4. Simplicity: Clear boundaries between concerns
5. Testing: Can test command and query sides independently

Negative

1. Eventual Consistency: Queries may return stale data
2. Complexity: More moving parts than simple CRUD
3. Duplication: Data exists in events and projections
4. Synchronization: Must keep projections up to date

Neutral

1. Learning Curve: Teams need to understand CQRS concepts
2. Tooling: Need projection management tools

Design Decisions

1. Commands Return Results, Not Entities

Commands return a result indicating success/failure, not the updated entity:

type CommandResult struct {
    AggregateID string
    Version     int64
    Error       error
}

Rationale: Returning entities would bypass the query side and couple commands to specific read models.

2. Queries Use Dedicated Repositories

Queries go through read model repositories, not the event store:

// Good: Query read model
orders, _ := orderRepo.Query(ctx, query)

// Avoid: Loading aggregate for queries
order := NewOrder(id)
store.LoadAggregate(ctx, order) // Only for commands

3. Projection Consistency Levels

We support three consistency levels via projection types:

Type	Consistency	Use Case
Inline	Strong	Financial transactions
Async	Eventual	Reports, analytics
Live	Real-time	Dashboards, notifications

Alternatives Considered

Alternative 1: No CQRS (Event Sourcing Only)

Description: Use event sourcing without separate read models.

Rejected because:

• Query performance degrades with event count
• Forces single data model for all use cases
• Doesn't leverage event sourcing benefits fully

Alternative 2: Full CQRS with Separate Databases

Description: Completely separate databases for read and write.

Rejected because:

• Adds infrastructure complexity
• Overkill for many use cases
• Can be added later if needed

Alternative 3: Synchronous Projections Only

Description: All projections updated in same transaction.

Rejected because:

• Reduces write throughput
• Limits scalability
• Some projections don't need strong consistency

References

• CQRS by Martin Fowler
• CQRS Documents by Greg Young
• Microsoft CQRS Pattern