dotnet-efcore-architecture

Strategic architectural patterns for EF Core data layers. Covers read/write model separation, aggregate boundary design, repository vs direct DbContext policy, N+1 query governance, row limit enforcement, and projection patterns. These patterns guide how to structure a data layer -- not how to write individual queries (see [skill:dotnet-efcore-patterns] for tactical usage).

Out of scope: Tactical EF Core usage (DbContext lifecycle, AsNoTracking, migrations, interceptors, compiled queries) is covered in [skill:dotnet-efcore-patterns]. Choosing between EF Core, Dapper, and ADO.NET is covered in [skill:dotnet-data-access-strategy]. DI container mechanics -- see [skill:dotnet-csharp-dependency-injection]. Async patterns -- see [skill:dotnet-csharp-async-patterns]. Integration testing data layers -- see [skill:dotnet-integration-testing] for database fixture and Testcontainers patterns. CI/CD pipelines -- see [skill:dotnet-gha-patterns] and [skill:dotnet-ado-patterns].

Cross-references: [skill:dotnet-efcore-patterns] for tactical DbContext usage and migrations, [skill:dotnet-data-access-strategy] for technology selection, [skill:dotnet-csharp-dependency-injection] for service registration, [skill:dotnet-csharp-async-patterns] for async query patterns.

---

Package Prerequisites

Examples in this skill use PostgreSQL (UseNpgsql). Substitute the provider package for your database:

Database	Provider Package
PostgreSQL	Npgsql.EntityFrameworkCore.PostgreSQL
SQL Server	Microsoft.EntityFrameworkCore.SqlServer
SQLite	Microsoft.EntityFrameworkCore.Sqlite

All examples also require the core Microsoft.EntityFrameworkCore package (pulled in transitively by provider packages).

---

Read/Write Model Separation

Separate read models (queries) from write models (commands) to optimize each path independently. This is not full CQRS -- it is a practical separation using EF Core features.

Approach: Separate DbContext Types

// Write context: full change tracking, navigation properties, interceptors
public sealed class WriteDbContext : DbContext
{
    public DbSet<Order> Orders => Set<Order>();
    public DbSet<Product> Products => Set<Product>();

    protected override void OnModelCreating(ModelBuilder modelBuilder)
    {
        modelBuilder.ApplyConfigurationsFromAssembly(typeof(WriteDbContext).Assembly);
    }
}

// Read context: no-tracking by default, optimized for projections
public sealed class ReadDbContext : DbContext
{
    public DbSet<Order> Orders => Set<Order>();
    public DbSet<Product> Products => Set<Product>();

    protected override void OnModelCreating(ModelBuilder modelBuilder)
    {
        modelBuilder.ApplyConfigurationsFromAssembly(typeof(ReadDbContext).Assembly);
    }

    protected override void OnConfiguring(DbContextOptionsBuilder optionsBuilder)
    {
        // Note: this is supplemental -- primary config is in DI registration
    }
}

Registration

// Write context: standard tracking, connection resiliency
builder.Services.AddDbContext<WriteDbContext>(options =>
    options.UseNpgsql(connectionString, npgsql =>
        npgsql.EnableRetryOnFailure(maxRetryCount: 3)));

// Read context: no-tracking, optionally pointed at a read replica
builder.Services.AddDbContext<ReadDbContext>(options =>
    options.UseNpgsql(readReplicaConnectionString ?? connectionString)
           .UseQueryTrackingBehavior(QueryTrackingBehavior.NoTracking));

When to Separate

Scenario	Recommendation
Simple CRUD app	Single DbContext with per-query AsNoTracking()
Read-heavy API with complex queries	Separate read/write contexts
Read replica database	Separate contexts with different connection strings
CQRS architecture	Separate contexts, possibly separate models

Start simple. Use a single DbContext and per-query AsNoTracking() until you have a concrete reason to split (different connection strings, divergent model shapes, or query complexity that justifies dedicated read models).

---

Aggregate Boundaries

An aggregate is a cluster of entities that are always loaded and saved together as a consistency boundary. EF Core maps well to aggregate-oriented design when navigation properties follow aggregate boundaries.

Defining Aggregates

// Order is the aggregate root -- it owns OrderItems
public sealed class Order
{
    public int Id { get; private set; }
    public string CustomerId { get; private set; } = default!;
    public OrderStatus Status { get; private set; }
    public DateTimeOffset CreatedAt { get; private set; }

    // Owned collection -- part of the Order aggregate
    private readonly List<OrderItem> _items = [];
    public IReadOnlyList<OrderItem> Items => _items.AsReadOnly();

    public void AddItem(int productId, int quantity, decimal unitPrice)
    {
        if (Status != OrderStatus.Draft)
            throw new InvalidOperationException("Cannot add items to a non-draft order.");

        _items.Add(new OrderItem(productId, quantity, unitPrice));
    }
}

// OrderItem belongs to the Order aggregate -- no independent access
public sealed class OrderItem
{
    public int Id { get; private set; }
    public int ProductId { get; private set; }
    public int Quantity { get; private set; }
    public decimal UnitPrice { get; private set; }

    internal OrderItem(int productId, int quantity, decimal unitPrice)
    {
        ProductId = productId;
        Quantity = quantity;
        UnitPrice = unitPrice;
    }

    private OrderItem() { } // EF Core constructor
}

EF Core Configuration for Aggregates

public sealed class OrderConfiguration : IEntityTypeConfiguration<Order>
{
    public void Configure(EntityTypeBuilder<Order> builder)
    {
        builder.HasKey(o => o.Id);

        builder.Property(o => o.CustomerId).IsRequired().HasMaxLength(50);
        builder.Property(o => o.Status).HasConversion<string>();

        // Owned collection navigation -- cascade delete, no independent DbSet
        builder.OwnsMany(o => o.Items, items =>
        {
            items.WithOwner().HasForeignKey("OrderId");
            items.Property(i => i.ProductId).IsRequired();
        });

        // Alternatively, if OrderItem needs its own table with explicit FK:
        // builder.HasMany(o => o.Items)
        //     .WithOne()
        //     .HasForeignKey("OrderId")
        //     .OnDelete(DeleteBehavior.Cascade);
        //
        // builder.Navigation(o => o.Items)
        //     .UsePropertyAccessMode(PropertyAccessMode.Field);
    }
}

Aggregate Design Rules

1. Load the entire aggregate -- do not load partial aggregates. Use Include() for the owned collections.
2. Save through the aggregate root -- call SaveChangesAsync() on the root, not on child entities independently.
3. Reference other aggregates by ID -- do not create navigation properties between aggregate roots. Use CustomerId (foreign key value), not Customer (navigation property).
4. Keep aggregates small -- large aggregates cause lock contention and slow loads. If a collection grows unbounded (e.g., audit logs), it does not belong in the aggregate.
5. One aggregate per transaction -- modifying multiple aggregates in a single transaction creates coupling. Use domain events or eventual consistency for cross-aggregate operations.

---

Repository Policy

Whether to use the repository pattern or access DbContext directly is a team decision. Both approaches are valid in .NET.

Option A: Direct DbContext Access

public sealed class CreateOrderHandler(WriteDbContext db)
{
    public async Task<int> HandleAsync(
        CreateOrderCommand command,
        CancellationToken ct)
    {
        var order = new Order(command.CustomerId);

        foreach (var item in command.Items)
        {
            order.AddItem(item.ProductId, item.Quantity, item.UnitPrice);
        }

        db.Orders.Add(order);
        await db.SaveChangesAsync(ct);

        return order.Id;
    }
}

Pros: Simple, no abstraction overhead, full LINQ power, easy to debug. Cons: Business logic can leak into query methods, harder to unit test without a database.

Option B: Repository per Aggregate Root

public interface IOrderRepository
{
    Task<Order?> GetByIdAsync(int id, CancellationToken ct);
    Task AddAsync(Order order, CancellationToken ct);
    Task SaveChangesAsync(CancellationToken ct);
}

public sealed class OrderRepository(WriteDbContext db) : IOrderRepository
{
    public async Task<Order?> GetByIdAsync(int id, CancellationToken ct)
    {
        return await db.Orders
            .Include(o => o.Items)
            .FirstOrDefaultAsync(o => o.Id == id, ct);
    }

    public async Task AddAsync(Order order, CancellationToken ct)
    {
        await db.Orders.AddAsync(order, ct);
    }

    public Task SaveChangesAsync(CancellationToken ct)
    {
        return db.SaveChangesAsync(ct);
    }
}

Pros: Testable without a database, encapsulates query logic, enforces aggregate loading rules. Cons: Extra abstraction layer, can become a leaky abstraction if LINQ is exposed, repository per aggregate can proliferate.

Decision Guide

Factor	Direct DbContext	Repository
Team size	Small, aligned	Large, varied experience
Test strategy	Integration tests with real DB	Unit tests with mocked repos
Query complexity	High (reports, projections)	Low-medium (CRUD, aggregates)
Aggregate discipline	Enforced by convention	Enforced by interface

Do not create generic repositories (IRepository<T>). They add abstraction without value -- the generic interface cannot express aggregate-specific loading rules (which Includes to use, which filters to apply). Repository interfaces should be specific to the aggregate root they serve.

---

N+1 Query Governance

N+1 queries are the most common EF Core performance problem. They occur when code iterates over a collection and executes a query per element, instead of loading all data upfront.

Detection

Enable sensitive logging in development to see SQL queries:

builder.Services.AddDbContext<AppDbContext>(options =>
    options.UseNpgsql(connectionString)
           .LogTo(Console.WriteLine, LogLevel.Information)
           .EnableSensitiveDataLogging()  // Development only
           .EnableDetailedErrors());      // Development only

Common N+1 Patterns and Fixes

Pattern 1: Lazy loading in a loop

// BAD: N+1 -- each order.Items triggers a query
var orders = await db.Orders.ToListAsync(ct);
foreach (var order in orders)
{
    var total = order.Items.Sum(i => i.Quantity * i.UnitPrice); // Lazy load!
}

// GOOD: Eager load with Include
var orders = await db.Orders
    .Include(o => o.Items)
    .ToListAsync(ct);

Pattern 2: Querying inside a loop

// BAD: N+1 -- one query per customer
foreach (var customerId in customerIds)
{
    var orders = await db.Orders
        .Where(o => o.CustomerId == customerId)
        .ToListAsync(ct);
    // ...
}

// GOOD: Single query with Contains
var orders = await db.Orders
    .Where(o => customerIds.Contains(o.CustomerId))
    .ToListAsync(ct);

Pattern 3: Missing projection

// BAD: Loads full entity graph, then maps in memory
var orders = await db.Orders
    .Include(o => o.Items)
    .Include(o => o.Customer)
    .ToListAsync(ct);
var dtos = orders.Select(o => new OrderDto(...));

// GOOD: Project in the query -- no tracking, no extra data loaded
var dtos = await db.Orders
    .Select(o => new OrderDto
    {
        Id = o.Id,
        CustomerName = o.Customer.Name,
        ItemCount = o.Items.Count,
        Total = o.Items.Sum(i => i.Quantity * i.UnitPrice)
    })
    .ToListAsync(ct);

Governance Checklist

• Disable lazy loading -- do not install Microsoft.EntityFrameworkCore.Proxies or configure UseLazyLoadingProxies(). Eager loading via Include() or projection via Select() makes data access explicit.
• Review queries in code review -- look for loops that access navigation properties or call FindAsync / FirstOrDefaultAsync per element.
• Use query tags -- db.Orders.TagWith("GetOrderSummary") makes queries identifiable in logs and profiling tools.
• Set up EF Core logging in development -- every lazy load or unexpected query is visible in the console output.

---

Row Limits and Pagination

Unbounded queries are a production risk. Always limit the number of rows returned.

Keyset Pagination (Recommended)

Keyset pagination (also called cursor-based or seek pagination) is more efficient than offset pagination for large datasets:

public async Task<PagedResult<OrderSummary>> GetOrdersAsync(
    string customerId,
    int? afterId,
    int pageSize,
    CancellationToken ct)
{
    const int maxPageSize = 100;
    pageSize = Math.Min(pageSize, maxPageSize);

    var query = db.Orders
        .AsNoTracking()
        .Where(o => o.CustomerId == customerId);

    if (afterId.HasValue)
    {
        query = query.Where(o => o.Id > afterId.Value);
    }

    var items = await query
        .OrderBy(o => o.Id)
        .Take(pageSize + 1)  // Fetch one extra to detect "has next page"
        .Select(o => new OrderSummary
        {
            Id = o.Id,
            Status = o.Status,
            CreatedAt = o.CreatedAt,
            Total = o.Items.Sum(i => i.Quantity * i.UnitPrice)
        })
        .ToListAsync(ct);

    var hasNext = items.Count > pageSize;
    if (hasNext)
    {
        items.RemoveAt(items.Count - 1);
    }

    return new PagedResult<OrderSummary>
    {
        Items = items,
        HasNextPage = hasNext,
        NextCursor = hasNext ? items[^1].Id : null
    };
}

Offset Pagination (Simple Cases)

For admin UIs or small datasets where exact page numbers matter:

var page = await db.Orders
    .AsNoTracking()
    .OrderBy(o => o.CreatedAt)
    .Skip((pageNumber - 1) * pageSize)
    .Take(pageSize)
    .ToListAsync(ct);

Warning: Offset pagination degrades at scale -- OFFSET 10000 forces the database to scan and discard 10,000 rows. Prefer keyset pagination for user-facing APIs.

Row Limit Enforcement

Set a hard upper bound on all queries to prevent accidental full-table scans:

// Interceptor approach: enforce max rows at the DbContext level
public sealed class RowLimitInterceptor : IQueryExpressionInterceptor
{
    private const int MaxRows = 1000;

    public Expression QueryCompilationStarting(
        Expression queryExpression,
        QueryExpressionEventData eventData)
    {
        // This is a simplified illustration -- actual implementation requires
        // expression tree analysis to detect existing Take() calls.
        // Consider using a code review rule or analyzer instead.
        return queryExpression;
    }
}

Practical approach: Rather than a runtime interceptor, enforce row limits through:

1. Code review convention -- every ToListAsync() must have Take(N) or be a Select() projection with Take(N).
2. API-level page size caps -- validate pageSize in the request pipeline before it reaches the query.
3. Query tags -- annotate queries with TagWith() to identify unbounded queries in monitoring.

---

Projection Patterns

Projections (Select()) are the most effective optimization for read queries. They reduce data transfer, skip change tracking, and eliminate N+1 risks.

Typed Projections

public sealed record OrderSummary
{
    public int Id { get; init; }
    public string CustomerName { get; init; } = default!;
    public int ItemCount { get; init; }
    public decimal Total { get; init; }
    public DateTimeOffset CreatedAt { get; init; }
}

var summaries = await db.Orders
    .Select(o => new OrderSummary
    {
        Id = o.Id,
        CustomerName = o.Customer.Name,
        ItemCount = o.Items.Count,
        Total = o.Items.Sum(i => i.Quantity * i.UnitPrice),
        CreatedAt = o.CreatedAt
    })
    .OrderByDescending(o => o.CreatedAt)
    .Take(50)
    .ToListAsync(ct);

Advantages Over Entity Loading

Concern	Entity + Include	Projection (Select)
Change tracking	Yes (unless AsNoTracking)	No
Data transferred	All columns	Only selected columns
N+1 risk	Yes (lazy nav props)	No (computed in SQL)
Cartesian explosion	Yes (multiple Includes)	No (single query)
Type safety	Entity types	DTO/record types

Rule: Use projections for all read-only endpoints that return DTOs. Reserve entity loading for commands that modify data.

---

Key Principles

• Separate read and write paths when you have different optimization needs -- do not force a single model to serve both
• Design aggregates around consistency boundaries -- not around database tables
• Reference other aggregates by ID -- navigation properties between aggregate roots create coupling
• Ban lazy loading -- make all data access explicit through Include() or Select()
• Enforce row limits -- every query that returns a list must have an upper bound
• Project early -- use Select() to push computation to the database and reduce data transfer
• Prefer keyset pagination over offset pagination for scalability

---

Agent Gotchas

1. Do not create navigation properties between aggregate roots -- use foreign key values (e.g., CustomerId) instead of navigation properties (e.g., Customer). Cross-aggregate navigation properties break the consistency boundary and encourage loading data that belongs to another aggregate.
2. Do not create generic repositories (IRepository<T>) -- they cannot express aggregate-specific loading rules and become leaky abstractions. Create one repository interface per aggregate root with explicit methods.
3. Do not use UseLazyLoadingProxies() -- lazy loading hides N+1 queries and makes performance unpredictable. Use Include() for eager loading or Select() for projections.
4. Do not return IQueryable<T> from repositories -- it leaks persistence concerns to callers and makes query behavior unpredictable (e.g., multiple enumeration, client-side evaluation). Return materialized results (List<T>, T?).
5. Do not write ToListAsync() without Take() on unbounded queries -- full table scans are a production incident waiting to happen. Always limit the result set.
6. Do not put audit logs or event streams inside aggregates -- unbounded collections cause slow loads and lock contention. Model them as separate entities or dedicated stores.

---

References

• EF Core performance best practices
• EF Core loading related data
• EF Core global query filters
• Domain-driven design with EF Core
• EF Core query tags
• Keyset pagination in EF Core