dotnet-io-pipelines

High-performance I/O patterns using System.IO.Pipelines. Covers PipeReader, PipeWriter, backpressure management, protocol parser implementation, and Kestrel integration. Pipelines solve the classic problems of buffer management, incomplete reads, and memory copying that plague traditional stream-based network code.

Out of scope: Async/await fundamentals and ValueTask patterns -- see [skill:dotnet-csharp-async-patterns]. Benchmarking methodology and Span<T> micro-optimization -- see [skill:dotnet-performance-patterns]. File-based I/O (FileStream, RandomAccess, MemoryMappedFile) -- see [skill:dotnet-file-io].

Cross-references: [skill:dotnet-csharp-async-patterns] for async patterns used in pipeline loops, [skill:dotnet-performance-patterns] for Span/Memory optimization techniques, [skill:dotnet-file-io] for file-based I/O patterns (FileStream, RandomAccess, MemoryMappedFile).

---

Why Pipelines Over Streams

Traditional Stream-based I/O forces developers to manage buffers manually, handle partial reads, and copy data between buffers. System.IO.Pipelines solves these problems:

Problem	Stream Approach	Pipeline Approach
Buffer management	Allocate byte[], resize manually	Automatic pooled buffer management
Partial reads	Track position, concatenate fragments	ReadResult with SequencePosition bookmarks
Backpressure	None -- writer can outpace reader	Built-in pause/resume thresholds
Memory copies	Copy between buffers at each layer	Zero-copy slicing with ReadOnlySequence<byte>
Lifetime management	Manual byte[] lifecycle	Pooled memory returned on AdvanceTo

The Pipe class connects a PipeWriter (producer) and a PipeReader (consumer) with an internal buffer pool, flow control, and completion signaling.

---

Core Concepts

Pipe, PipeReader, PipeWriter

// Create a pipe with default options (uses ArrayPool internally)
var pipe = new Pipe();

PipeWriter writer = pipe.Writer;  // Producer side
PipeReader reader = pipe.Reader;  // Consumer side

PipeWriter -- Producing Data

async Task FillPipeAsync(Stream source, PipeWriter writer,
    CancellationToken ct)
{
    const int minimumBufferSize = 512;

    while (true)
    {
        // Request a buffer from the pipe's memory pool
        Memory<byte> memory = writer.GetMemory(minimumBufferSize);

        int bytesRead = await source.ReadAsync(memory, ct);
        if (bytesRead == 0)
            break;  // End of stream

        // Tell the pipe how many bytes were written
        writer.Advance(bytesRead);

        // Flush makes data available to the reader.
        // FlushAsync may pause here if the reader is slow (backpressure).
        FlushResult result = await writer.FlushAsync(ct);
        if (result.IsCompleted)
            break;  // Reader stopped consuming
    }

    // Signal completion -- reader will see IsCompleted = true
    await writer.CompleteAsync();
}

Critical rules:

• Call GetMemory or GetSpan before writing -- never write to a previously obtained buffer after Advance
• Call Advance with the exact number of bytes written
• Call FlushAsync to make data available to the reader and to respect backpressure

PipeReader -- Consuming Data

async Task ReadPipeAsync(PipeReader reader, CancellationToken ct)
{
    while (true)
    {
        ReadResult result = await reader.ReadAsync(ct);
        ReadOnlySequence<byte> buffer = result.Buffer;

        // Try to parse messages from the buffer
        while (TryParseMessage(ref buffer, out var message))
        {
            await ProcessMessageAsync(message, ct);
        }

        // Tell the pipe how much was consumed and how much was examined.
        // consumed: data that has been fully processed (will be freed)
        // examined: data that has been looked at (won't trigger re-read
        //           until new data arrives)
        reader.AdvanceTo(buffer.Start, buffer.End);

        if (result.IsCompleted)
            break;  // Writer finished and all data consumed
    }

    await reader.CompleteAsync();
}

Critical rules:

• Always call AdvanceTo after ReadAsync -- failing to do so leaks memory
• Pass both consumed and examined positions: consumed frees memory, examined prevents busy-wait when the buffer has been scanned but does not contain a complete message
• Never access ReadResult.Buffer after calling AdvanceTo -- the memory may be recycled

---

Backpressure

Backpressure prevents fast producers from overwhelming slow consumers. The pipe pauses the writer when unread data exceeds a threshold.

PipeOptions Configuration

var pipe = new Pipe(new PipeOptions(
    pauseWriterThreshold: 64 * 1024,   // Pause writer at 64 KB buffered
    resumeWriterThreshold: 32 * 1024,  // Resume writer when buffer drops to 32 KB
    minimumSegmentSize: 4096,
    useSynchronizationContext: false));

Option	Default	Purpose
PauseWriterThreshold	65,536	FlushAsync pauses when unread bytes exceed this
ResumeWriterThreshold	32,768	FlushAsync resumes when unread bytes drop below this
MinimumSegmentSize	4,096	Minimum buffer segment allocation size
UseSynchronizationContext	false	Set false for server code to avoid context captures

How Backpressure Works

1. Writer calls FlushAsync after Advance
2. If buffered (unread) data exceeds PauseWriterThreshold, FlushAsync does not complete until the reader consumes enough data to drop below ResumeWriterThreshold
3. The writer is effectively paused -- no busy-waiting, no exceptions, just an awaitable that completes when the reader catches up

This prevents unbounded memory growth when a producer (network socket, file) is faster than the consumer (parser, business logic).

---

Protocol Parsing

Pipelines excel at parsing binary protocols because ReadOnlySequence<byte> handles fragmented data across multiple buffer segments without copying.

Length-Prefixed Protocol Parser

A common pattern: each message starts with a 4-byte big-endian length header followed by the payload.

static bool TryParseMessage(
    ref ReadOnlySequence<byte> buffer,
    out ReadOnlySequence<byte> payload)
{
    payload = default;

    // Need at least 4 bytes for the length prefix
    if (buffer.Length < 4)
        return false;

    // Read length from first 4 bytes
    int length;
    if (buffer.FirstSpan.Length >= 4)
    {
        length = BinaryPrimitives.ReadInt32BigEndian(buffer.FirstSpan);
    }
    else
    {
        // Slow path: length header spans multiple segments
        Span<byte> lengthBytes = stackalloc byte[4];
        buffer.Slice(0, 4).CopyTo(lengthBytes);
        length = BinaryPrimitives.ReadInt32BigEndian(lengthBytes);
    }

    // Validate length to prevent allocation attacks
    if (length < 0 || length > 1_048_576)  // 1 MB max
        throw new ProtocolViolationException(
            $"Invalid message length: {length}");

    // Check if the full message is available
    long totalLength = 4 + length;
    if (buffer.Length < totalLength)
        return false;

    // Extract the payload (zero-copy slice)
    payload = buffer.Slice(4, length);

    // Advance the buffer past this message
    buffer = buffer.Slice(totalLength);
    return true;
}

Delimiter-Based Protocol Parser (Line Protocol)

static bool TryReadLine(
    ref ReadOnlySequence<byte> buffer,
    out ReadOnlySequence<byte> line)
{
    // Look for the newline delimiter
    SequencePosition? position = buffer.PositionOf((byte)'\n');
    if (position is null)
    {
        line = default;
        return false;
    }

    // Slice up to (not including) the delimiter
    line = buffer.Slice(0, position.Value);

    // Advance past the delimiter
    buffer = buffer.Slice(buffer.GetPosition(1, position.Value));
    return true;
}

Working with ReadOnlySequence<byte>

ReadOnlySequence<byte> may span multiple non-contiguous memory segments. Handle both paths:

static string DecodeUtf8(ReadOnlySequence<byte> sequence)
{
    // Fast path: single contiguous segment
    if (sequence.IsSingleSegment)
    {
        return Encoding.UTF8.GetString(sequence.FirstSpan);
    }

    // Slow path: multi-segment -- rent a contiguous buffer
    int length = (int)sequence.Length;
    byte[] rented = ArrayPool<byte>.Shared.Rent(length);
    try
    {
        sequence.CopyTo(rented);
        return Encoding.UTF8.GetString(rented, 0, length);
    }
    finally
    {
        ArrayPool<byte>.Shared.Return(rented);
    }
}

---

Stream Adapter

Bridge System.IO.Pipelines with existing Stream-based APIs using PipeReader.Create and PipeWriter.Create.

// Wrap a NetworkStream for pipeline-based reading
await using var networkStream = tcpClient.GetStream();
var reader = PipeReader.Create(networkStream, new StreamPipeReaderOptions(
    bufferSize: 4096,
    minimumReadSize: 1024,
    leaveOpen: true)); // Caller manages networkStream lifetime

try
{
    await ProcessProtocolAsync(reader, cancellationToken);
}
finally
{
    await reader.CompleteAsync();
}

// Wrap a stream for pipeline-based writing
var writer = PipeWriter.Create(networkStream, new StreamPipeWriterOptions(
    minimumBufferSize: 4096,
    leaveOpen: true)); // Caller manages networkStream lifetime

try
{
    await WriteResponseAsync(writer, response, cancellationToken);
}
finally
{
    await writer.CompleteAsync();
}

---

Kestrel Integration

ASP.NET Core's Kestrel web server uses System.IO.Pipelines internally for HTTP request/response processing. Custom connection middleware can access the transport-level pipe directly.

Connection Middleware

// Custom connection middleware for protocol-level processing
builder.WebHost.ConfigureKestrel(options =>
{
    options.ListenLocalhost(9000, listenOptions =>
    {
        listenOptions.UseConnectionHandler<MyProtocolHandler>();
    });
});

public sealed class MyProtocolHandler : ConnectionHandler
{
    public override async Task OnConnectedAsync(
        ConnectionContext connection)
    {
        var reader = connection.Transport.Input;
        var writer = connection.Transport.Output;
        var ct = connection.ConnectionClosed;

        try
        {
            while (true)
            {
                ReadResult result = await reader.ReadAsync(ct);
                ReadOnlySequence<byte> buffer = result.Buffer;

                while (TryParseMessage(ref buffer, out var payload))
                {
                    var response = ProcessRequest(payload);
                    await WriteResponseAsync(writer, response);
                }

                reader.AdvanceTo(buffer.Start, buffer.End);

                if (result.IsCompleted)
                    break;
            }
        }
        finally
        {
            await reader.CompleteAsync();
            await writer.CompleteAsync();
        }
    }

    private static async Task WriteResponseAsync(
        PipeWriter writer, ReadOnlyMemory<byte> response)
    {
        // Write length prefix + payload
        var memory = writer.GetMemory(4 + response.Length);
        BinaryPrimitives.WriteInt32BigEndian(
            memory.Span, response.Length);
        response.CopyTo(memory[4..]);
        writer.Advance(4 + response.Length);
        await writer.FlushAsync();
    }
}

IDuplexPipe

Kestrel exposes connections as IDuplexPipe, combining PipeReader and PipeWriter into a single transport abstraction. This pattern also works for custom TCP servers, WebSocket handlers, and named-pipe protocols.

public interface IDuplexPipe
{
    PipeReader Input { get; }
    PipeWriter Output { get; }
}

---

Performance Tips

1. Minimize copies -- use ReadOnlySequence<byte> slicing instead of copying to byte[]. Parse directly from the sequence when possible.
2. Use GetSpan/GetMemory correctly -- request the minimum size you need. The pipe may return a larger buffer, which is fine. Do not cache the returned Span/Memory across Advance/FlushAsync calls.
3. Set useSynchronizationContext: false -- server code should never capture the synchronization context. This is the default for PipeOptions but explicit is clearer.
4. Tune pause/resume thresholds -- the defaults (64 KB / 32 KB) work for most scenarios. Increase for high-throughput bulk transfer; decrease for low-latency interactive protocols.
5. Prefer SequenceReader<byte> -- for complex parsing, SequenceReader<byte> provides TryRead, TryReadBigEndian, AdvancePast, and IsNext methods that handle multi-segment sequences transparently.

static bool TryParseHeader(
    ref ReadOnlySequence<byte> buffer,
    out int messageType,
    out int length)
{
    var reader = new SequenceReader<byte>(buffer);

    if (!reader.TryRead(out byte typeByte) ||
        !reader.TryReadBigEndian(out int len))
    {
        messageType = 0;
        length = 0;
        return false;
    }

    messageType = typeByte;
    length = len;
    buffer = buffer.Slice(reader.Position);
    return true;
}

---

Agent Gotchas

1. Do not forget to call AdvanceTo after ReadAsync -- skipping AdvanceTo leaks pooled memory and eventually causes OutOfMemoryException. Every ReadAsync must be paired with an AdvanceTo.
2. Do not access ReadResult.Buffer after calling AdvanceTo -- the underlying memory segments may be returned to the pool. Copy or parse all needed data before advancing.
3. Do not set consumed equal to examined when no complete message was found -- this creates a busy-wait loop. Set consumed to buffer.Start (nothing consumed) and examined to buffer.End (everything examined) so the pipe waits for new data.
4. Do not ignore FlushResult.IsCompleted -- it means the reader has stopped consuming. Continue writing after this and data will be silently discarded.
5. Do not use Pipe for simple stream-to-stream copying -- Stream.CopyToAsync is simpler and equally efficient. Use pipelines when you need parsing, backpressure, or zero-copy slicing.
6. Do not use BinaryPrimitives methods on spans shorter than required -- always check buffer.Length before reading fixed-width values to avoid ArgumentOutOfRangeException.

---

Knowledge Sources

• Stephen Toub, System.IO.Pipelines: High performance IO in .NET -- canonical deep dive on pipeline design, motivation, and usage patterns

References

• System.IO.Pipelines overview
• Pipe class API reference
• PipeReader API reference
• PipeWriter API reference
• SequenceReader<T>
• Kestrel connection middleware