Public API Design and Compatibility

When to Use This Skill

Use this skill when:

Designing public APIs for NuGet packages or libraries
Making changes to existing public APIs
Planning wire format changes for distributed systems
Implementing versioning strategies
Reviewing pull requests for breaking changes

The Three Types of Compatibility

Type	Definition	Scope
API/Source	Code compiles against newer version	Public method signatures, types
Binary	Compiled code runs against newer version	Assembly layout, method tokens
Wire	Serialized data readable by other versions	Network protocols, persistence formats

Breaking any of these creates upgrade friction for users.

Extend-Only Design

The foundation of stable APIs: never remove or modify, only extend.

Three Pillars

Previous functionality is immutable - Once released, behavior and signatures are locked
New functionality through new constructs - Add overloads, new types, opt-in features
Removal only after deprecation period - Years, not releases

Benefits

Old code continues working in new versions
New and old pathways coexist
Upgrades are non-breaking by default
Users upgrade on their schedule

Resources:

API Change Guidelines

Safe Changes (Any Release)

// ADD new overloads with default parameters
public void Process(Order order, CancellationToken ct = default);

// ADD new optional parameters to existing methods
public void Send(Message msg, Priority priority = Priority.Normal);

// ADD new types, interfaces, enums
public interface IOrderValidator { }
public enum OrderStatus { Pending, Complete, Cancelled }

// ADD new members to existing types
public class Order
{
    public DateTimeOffset? ShippedAt { get; init; }  // NEW
}

Unsafe Changes (Never or Major Version Only)

// REMOVE or RENAME public members
public void ProcessOrder(Order order);  // Was: Process()

// CHANGE parameter types or order
public void Process(int orderId);  // Was: Process(Order order)

// CHANGE return types
public Order? GetOrder(string id);  // Was: public Order GetOrder()

// CHANGE access modifiers
internal class OrderProcessor { }  // Was: public

// ADD required parameters without defaults
public void Process(Order order, ILogger logger);  // Breaks callers!

Deprecation Pattern

// Step 1: Mark as obsolete with version (any release)
[Obsolete("Obsolete since v1.5.0. Use ProcessAsync instead.")]
public void Process(Order order) { }

// Step 2: Add new recommended API (same release)
public Task ProcessAsync(Order order, CancellationToken ct = default);

// Step 3: Remove in next major version (v2.0+)
// Only after users have had time to migrate

API Approval Testing

Prevent accidental breaking changes with automated API surface testing.

Using ApiApprover + Verify

dotnet add package PublicApiGenerator
dotnet add package Verify.Xunit

[Fact]
public Task ApprovePublicApi()
{
    var api = typeof(MyLibrary.PublicClass).Assembly.GeneratePublicApi();
    return Verify(api);
}

Creates ApprovePublicApi.verified.txt:

namespace MyLibrary
{
    public class OrderProcessor
    {
        public OrderProcessor() { }
        public void Process(Order order) { }
        public Task ProcessAsync(Order order, CancellationToken ct = default) { }
    }
}

Any API change fails the test - reviewer must explicitly approve changes.

PR Review Process

PR includes changes to *.verified.txt files
Reviewers see exact API surface changes in diff
Breaking changes are immediately visible
Conscious decision required to approve

Wire Compatibility

For distributed systems, serialized data must be readable across versions.

Requirements

Direction	Requirement
Backward	Old writers → New readers (current version reads old data)
Forward	New writers → Old readers (old version reads new data)

Both are required for zero-downtime rolling upgrades.

Safely Evolving Wire Formats

Phase 1: Add read-side support (opt-in)

// New message type - readers deployed first
public sealed record HeartbeatV2(
    Address From,
    long SequenceNr,
    long CreationTimeMs);  // NEW field

// Deserializer handles both old and new
public object Deserialize(byte[] data, string manifest) => manifest switch
{
    "Heartbeat" => DeserializeHeartbeatV1(data),   // Old format
    "HeartbeatV2" => DeserializeHeartbeatV2(data), // New format
    _ => throw new NotSupportedException()
};

Phase 2: Enable write-side (opt-out, next minor version)

// Config to enable new format (off by default initially)
akka.cluster.use-heartbeat-v2 = on

Phase 3: Make default (future version)

After install base has absorbed read-side code.

Schema-Based Serialization

Prefer schema-based formats over reflection-based:

Format	Type	Wire Compatibility
Protocol Buffers	Schema-based	Excellent - explicit field numbers
MessagePack	Schema-based	Good - with contracts
System.Text.Json	Schema-based (with source gen)	Good - explicit properties
Newtonsoft.Json	Reflection-based	Poor - type names in payload
BinaryFormatter	Reflection-based	Terrible - never use

See dotnet/serialization skill for details.

Encapsulation Patterns

Internal APIs

Mark non-public APIs explicitly:

// Attribute for documentation
[InternalApi]
public class ActorSystemImpl { }

// Namespace convention
namespace MyLibrary.Internal
{
    public class InternalHelper { }  // Public for extensibility, not for users
}

Document clearly:

Types in .Internal namespaces or marked with [InternalApi] may change between any releases without notice.

Sealing Classes

// DO: Seal classes not designed for inheritance
public sealed class OrderProcessor { }

// DON'T: Leave unsealed by accident
public class OrderProcessor { }  // Users might inherit, blocking changes

Interface Segregation

// DO: Small, focused interfaces
public interface IOrderReader
{
    Order? GetById(OrderId id);
}

public interface IOrderWriter
{
    Task SaveAsync(Order order);
}

// DON'T: Monolithic interfaces (can't add methods without breaking)
public interface IOrderRepository
{
    Order? GetById(OrderId id);
    Task SaveAsync(Order order);
    // Adding new methods breaks all implementations!
}

Versioning Strategy

Semantic Versioning (Practical)

Version	Changes Allowed
Patch (1.0.x)	Bug fixes, security patches
Minor (1.x.0)	New features, deprecations, obsolete removal
Major (x.0.0)	Breaking changes, old API removal

Key Principles

No surprise breaks - Even major versions should be announced and planned
Extensions anytime - New APIs can ship in any release
Deprecate before remove - [Obsolete] for at least one minor version
Communicate timelines - Users need to plan upgrades

Chesterton's Fence

Before removing or changing something, understand why it exists.

Assume every public API is used by someone. If you want to change it:

Socialize the proposal on GitHub
Document migration path
Provide deprecation period
Ship in planned release

Pull Request Checklist

When reviewing PRs that touch public APIs:

No removed public members (use [Obsolete] instead)
No changed signatures (add overloads instead)
No new required parameters (use defaults)
API approval test updated (.verified.txt changes reviewed)
Wire format changes are opt-in (read-side first)
Breaking changes documented (release notes, migration guide)

Anti-Patterns

Breaking Changes Disguised as Fixes

// "Bug fix" that breaks users
public async Task<Order> GetOrderAsync(OrderId id)  // Was sync!
{
    // "Fixed" to be async - but breaks all callers
}

// Correct: Add new method, deprecate old
[Obsolete("Use GetOrderAsync instead")]
public Order GetOrder(OrderId id) => GetOrderAsync(id).Result;

public async Task<Order> GetOrderAsync(OrderId id) { }

Silent Behavior Changes

// Changing defaults breaks users who relied on old behavior
public void Configure(bool enableCaching = true)  // Was: false!

// Correct: New parameter with new name
public void Configure(
    bool enableCaching = false,  // Original default preserved
    bool enableNewCaching = true)  // New behavior opt-in

Polymorphic Serialization

// AVOID: Type names in wire format
{ "$type": "MyApp.Order, MyApp", "Id": 123 }

// Renaming Order class = wire break!

// PREFER: Explicit discriminators
{ "type": "order", "id": 123 }

api-design