SSWG Proposal

gRPC Swift

Proposal: SSWG-NNNN
Authors: Daniel Alm, George Barnett, Tim Burks, Michael Rebello
Sponsor(s): TBD
Review Manager: TBD
Status: Implemented
Implementation: grpc-swift
Forum Threads: TODO

Package Description

A gRPC client and server library with code generation.


Package Name	`GRPC`
Module Name	`GRPC`
Proposed Maturity Level	Sandbox
License	Apache 2
Dependencies	SwiftNIO 2.2, SwiftNIO HTTP2 1.5, SwiftNIO SSL 2.4, SwiftNIO Transport Services 1.1, SwiftProtobuf 1.5, SwiftLog 1.0, Commander 0.8

Introduction

gRPC is a widely used protocol with implementations in a number of languages. Most of these implementations wrap a core C-library which can lead to memory safety issues and is difficult to debug. This leads to further rough edges on iOS where clients have to deal with network connectivity changes (e.g. LTE to WiFi). Having a gRPC server and client implementation in Swift built on top of NIO will help to eliminate or reduce each of these issues.

Motivation

TODO

Describe the reasoning for this package to be proposed to the Swift Server Working Group to be recommended for use across the Swift Server ecosystem.

If there are missing capabilities or flexibility with existing packages, outline them with clear examples.

Proposed solution

We will use SwiftNIO to provide the network layer, SwiftNIO SSL for TLS, SwiftNIO HTTP2 for HTTP/2, and SwiftProtobuf will be used for message serialization. We will also use SwiftNIO Transport Services to provide first-class support for Apple Platforms.

gRPC Background

Since gRPC generates client and server bindings from an interface definition language (Protocol Buffers) we should layout an example service so that the following proposal is more concrete. gRPC has four call types:

unary (client sends one request, server sends one response),
client streaming (client sends zero or more requests, server sends one response)
server streaming (client sends one request, server sends zero or more responses)
bidirectional streaming (client sends zero or more requests, server sends zero or more responses).

As an example we will consider an Echo service with one function for each of the four call types. Each function takes Echo_EchoRequest(s) as input and returns Echo_EchoResponse(s). The functions are:

get (unary),
collect (client streaming),
expand (server streaming), and
update (bidirectional streaming).

The semantics of each call are explained in the server section below and the code is available in the gRPC Swift repository.

Server

Users implement a service's business logic in a handler conforming to a generated "provider" protocol which transitively conforms to CallHandlerProvider. The implementation of the generated protocol is passed to the sever on initialization.

For the Echo service the generated protocol is:

protocol Echo_EchoProvider: CallHandlerProvider {
  func get(
    request: Echo_EchoRequest,
    context: StatusOnlyCallContext
  ) -> EventLoopFuture<Echo_EchoResponse>

  func collect(
    context: UnaryResponseCallContext<Echo_EchoResponse>
  ) -> EventLoopFuture<(StreamEvent<Echo_EchoRequest>) -> Void>

  func expand(
    request: Echo_EchoRequest,
    context: StreamingResponseCallContext<Echo_EchoResponse>
  ) -> EventLoopFuture<GRPCStatus>

  func update(
    context: StreamingResponseCallContext<Echo_EchoResponse>
  ) -> EventLoopFuture<(StreamEvent<Echo_EchoRequest>) -> Void>
}

get accepts a single request and a context which has access to the event loop and request head; the function returns a future response.
collect provides a context which has access to the event loop, request head and a response promise; the function returns a future stream event handler which should fulfil the response promise.
expand accepts a single request and streaming context which provides a sendResponse method in addition to the event loop and request head provided to get, the call is terminated by returning the future status of the call.
update provides the same context as expand and returns a future stream event handler which should fulfil the status promise provided by the context.

A server can be started by instantiating a Server with configuration which includes a list of providers:

let group = MultiThreadedEventLoopGroup(numberOfThreads: 1)
let configuration = Server.Configuration(
  target: .hostAndPort("localhost", 8080),
  eventLoopGroup: group,
  serviceProviders: [EchoProvider()]
)

let server: EventLoopFuture<Server> = Server.start(configuration: configuration)

Client

To make a unary call to the Echo service:

// Initialize an Echo client
let echo: EchoClient = ...

// Call the Get function
let get = echo.get(Echo_EchoRequest.with { $0.text = "foo bar baz" })

// The server may send back initial metadata.
get.initialMetadata.whenSuccess { (headers: HTTPHeaders) in
  print("Get initial metadata: \(headers)")
}

// get is unary, so it has a response future.
get.response.whenSuccess { (response: Echo_EchoResponse in
  print("Get response: \(response)")
}

// Each call also has a status future containing the gRPC status code
// and (optional) message.
get.status.whenSuccess { (status: GRPCStatus) in
  print("Get status: \(status)")
}

// The server may also send back trailing metadata:
get.trailingMetadata.whenSuccess { (trailers: HTTPHeaders) in
  print("Get trailing metadata: \(trailers)")
}

To make a bidirectional streaming call to the Echo service:

// Call the Update function, providing a response handler.
let update = echo.update { (response: Echo_EchoResponse) in
  print("Update response: \(response)")
}

// The client is streaming so has methods to send messages to the server:
update.sendMessage(Echo_EchoRequest.with { $0.text = "foo bar baz" }, promise: nil)

// It is also responsible for closing the request stream:
update.sendEnd(promise: nil)

// Note: there are versions of the above methods which return EventLoopFuture<Void>
// instead of accepting a promise. There is also a method to send a batch of messages.

initialMetadata, trailingMetadata, and status also exist on the call object as in the unary call.

The client streaming and server streaming calls are just a combination of the unary and bidirectional streaming calls above:

The client streaming call provides methods for sending messages and has a response future
The server streaming call accepts a single request on initialization has a response handler.

Code Generation

We will provide plugins for the Protobuf compiler protoc to generate code for the client and server. SwiftProtobuf provides a plugin library which wewill make use of to do most of the heavy lifting.

Out of Scope

The following are out of scope for this proposal:

Supporting different serialization formats (e.g. FlatBuffers); only the Protocol Buffers format (via SwiftProtobuf) is supported.
Support for QUIC.

Detailed design

Server

Servers are started with some configuration:

let server: EventLoopFuture<Server> = Server.start(configuration: configuration)

The configuration (Server.Configuration) includes what the server should bind to (host and port, unix domain socket), the event loop group it use, TLS configuration (optional), error delegate (optional) and a list of CallHandlerProviders which it may use to serve requests.

The server also supports gRPC-Web via a handler which configures the pipeline based on the HTTP version of the request.

When the server recieves a request, the GRPCChannelHandler checks the path of the request (i.e the RPC being called) and determines whether a CallHandlerProvider exists which may service the request. If one exists then an appropriate GRPCCallHandler is returned which delegates logic to a method implemented by the user. The NIO pipeline for the server's child channels follows:

NIOSSLHandler (if TLS is being used)
If HTTP/1 (i.e. gRPC-Web):
- HTTPServerPipelineHandler
- WebCORSHandler
Otherwise (i.e. standard gRPC):
- NIOHTTP2Handler
- HTTP2StreamMultiplexer
- HTTP2ToHTTP1ServerCodec
HTTP1ToRawGRPCServerCodec translates HTTP/1 types to gRPC metadata and length-prefixed messages, it also handles request/response state, compression (not yet implemented) and message buffering since messages may span multiple frames. It is "Raw" since the emitted messages are just bytes and not yet typed.
GRPCChannelHandler configures the pipeline on receiving the request head by looking at the request URI and finding an appropriate service provider. This handler is removed from the pipeline when the handler has configured the rest of the pipeline.
GRPCCallHandler handles the delivery of requests and responses to and from the user implemented call handlers.

Where the GRPCCallHandler is one of:

UnaryCallHandler,
ClientStreamingCallHandler,
ServerStreamingCallHandler,
BidirectionalStreamingCallHandler

An example implementation follows:

class EchoProvider: Echo_EchoProvider {
  // get: Return a prefixed version of the request message
  func get(request: Echo_EchoRequest, context: StatusOnlyCallContext) -> EventLoopFuture<Echo_EchoResponse> {
    // Return the response future.
    return context.eventLoop.makeSucceededFuture(Echo_EchoResponse.with {
      $0.text = "Swift echo get: \(request.text)"
    })
  }

  // collect: Join the requests on " ", send the joined requests in a single response
  func collect(context: UnaryResponseCallContext<Echo_EchoResponse>) -> EventLoopFuture<(StreamEvent<Echo_EchoRequest>) -> Void> {
    var parts: [String] = []

    return context.eventLoop.makeSucceededFuture({ event in
      switch event {
      case .message(let message):
        // Buffer the message.
        parts.append(message.text)

      case .end:
        // We have a complete response now. Join the parts and succeed the promise.
        context.responsePromise.succeed(Echo_EchoResponse.with {
          $0.text = "Swift echo collect: " + parts.joined(separator: " ")
        })
      }
    })
  }

  // expand: Split the request message on " ", send each part in a separate response
  func expand(request: Echo_EchoRequest, context: StreamingResponseCallContext<Echo_EchoResponse>) -> EventLoopFuture<GRPCStatus> {
    var endOfSendOperationQueue = context.eventLoop.makeSucceededFuture(())

    // Split the request and add each part to a queue of messages to respond with.
    for (i, part) in request.text.components(separatedBy: " ").enumerated() {
      let response = Echo_EchoResponse.with {
        $0.text = "Swift echo expand (\(i)): \(part)"
      }
      endOfSendOperationQueue = endOfSendOperationQueue.flatMap {
        context.sendResponse(response)
      }
    }

    // Once the queue is done, return a status OK future.
    return endOfSendOperationQueue.map {
      GRPCStatus.ok
    }
  }

  // update: Return a prefixed version of each message in the request stream
  func update(context: StreamingResponseCallContext<Echo_EchoResponse>) -> EventLoopFuture<(StreamEvent<Echo_EchoRequest>) -> Void> {
    var endOfSendOperationQueue = context.eventLoop.makeSucceededFuture(())
    var count = 0

    return context.eventLoop.makeSucceededFuture({ event in
      switch event {
      case .message(let message):
        let response = Echo_EchoResponse.with {
          $0.text = "Swift echo update (\(count)): \(message.text)"
        }

        // Queue the response message.
        endOfSendOperationQueue = endOfSendOperationQueue.flatMap {
          context.sendResponse(response)
        }
        count += 1

      case .end:
        // End of request stream: fulfill the status promise.
        endOfSendOperationQueue.map {
          GRPCStatus.ok
        }.cascade(to: context.statusPromise)
      }
    })
  }
}

Client

Making Calls

Codifying what has been said previously about the four call types, each of them build on top of ClientCall:

public protocol ClientCall {
  associatedtype RequestMessage: Message
  associatedtype ResponseMessage: Message

  /// Initial response metadata.
  var initialMetadata: EventLoopFuture<HTTPHeaders> { get }

  /// Status of this call which may be populated by the server or client.
  var status: EventLoopFuture<GRPCStatus> { get }

  /// Trailing response metadata.
  var trailingMetadata: EventLoopFuture<HTTPHeaders> { get }

  /// Cancel the current call.
  func cancel()
}

The calls which have a single response from the server (unary and client streaming) implement UnaryResponseClientCall which extends ClientCall to include a future response:

public protocol UnaryResponseClientCall: ClientCall {
  /// The response message returned from the service if the call is successful.
  /// This may be failed if the call encounters an error.
  var response: EventLoopFuture<ResponseMessage> { get }
}

For calls which have any number of responses from the server (server streaming and bidirectional streaming), constructing the call requires a response handler: (ResponseMessage) -> Void.

Calls sending a single request to the server (unary and server streaming) accept a single request on initialisation. Calls which send any number of requests to the server (client streaming and bidirectional streaming) return a call which conforms to StreamingRequestClientCall which extends ClientCall to provide methods for sending messages to the server:

public protocol StreamingRequestClientCall: ClientCall {
  /// Sends a message to the service.
  func sendMessage(_ message: RequestMessage) -> EventLoopFuture<Void>
  func sendMessage(_ message: RequestMessage, promise: EventLoopPromise<Void>?)

  /// Sends a sequence of messages to the service.
  func sendMessages<S: Sequence>(_ messages: S) -> EventLoopFuture<Void> where S.Element == RequestMessage
  func sendMessages<S: Sequence>(_ messages: S, promise: EventLoopPromise<Void>?) where S.Element == RequestMessage

  /// Terminates a stream of messages sent to the service.
  func sendEnd() -> EventLoopFuture<Void>
  func sendEnd(promise: EventLoopPromise<Void>?)
}

Each of the call types can be made from factory methods on the GRPCClient protocol. The call signatures are:

public func makeUnaryCall<Request: Message, Response: Message>(
  path: String,
  request: Request,
  callOptions: CallOptions? = nil,
  responseType: Response.Type = Response.self
) -> UnaryCall<Request, Response>

public func makeServerStreamingCall<Request: Message, Response: Message>(
  path: String,
  request: Request,
  callOptions: CallOptions? = nil,
  responseType: Response.Type = Response.self,
  handler: @escaping (Response) -> Void
) -> ServerStreamingCall<Request, Response>

public func makeClientStreamingCall<Request: Message, Response: Message>(
  path: String,
  callOptions: CallOptions? = nil,
  requestType: Request.Type = Request.self,
  responseType: Response.Type = Response.self
) -> ClientStreamingCall<Request, Response>

public func makeBidirectionalStreamingCall<Request: Message, Response: Message>(
  path: String,
  callOptions: CallOptions? = nil,
  requestType: Request.Type = Request.self,
  responseType: Response.Type = Response.self,
  handler: @escaping (Response) -> Void
) -> BidirectionalStreamingCall<Request, Response>

This keeps the code generation straightforward: the generated client stubs call these functions with some static information, such as the path (e.g. "/Echo/Get") and the appropriate request and response types.

In cases where no client has been generated, an AnyServiceClient can be used. It provides the above methods but has no stubs for a service:

let anyServiceClient: AnyServiceClient = ...

// Equivalent to: echoClient.get(Echo_EchoRequest.with { $0.text = "foo bar baz" })
let get = anyServiceClient.makeUnaryCall(
  path: "/Echo/Get",
  request: Echo_EchoRequest.with { $0.text = "foo bar baz" },
  responseType: Echo_EchoResponse.self
)

Making Clients

So far we have glossed over how to construct a client. A client requires a connection to a gRPC server, in other gRPC implementations this it typically called a Channel. To avoid confusion with NIO it is named ClientConnection. A ClientConnection is initialized with some configuration:

let configuration = ClientConnection.Configuration(
  target: .hostAndPort("localhost", "8080"),
  eventLoopGroup: group,
  // Delegates for observing errors and connectivity state changes:
  errorDelegate: nil,
  connectivityStateDelegate: nil,
  // TLS configuration, a subset of NIO's TLSConfiguration:
  tls: nil,
  // Connection backoff configuration:
  connectionBackoff: ConnectionBackoff()
)

let connection = ClientConnection(configuration: configuration)

Clients take a ClientConnection and an optional CallOptions struct on initialization:

// Call options used for each call unless specified at call time.
// Has support for custom metadata (headers) and call timeouts amongst a few
// other things.
let defaultCallOptions = CallOptions(timeout: try .seconds(90))

// Create a client, this would usually be generated from a proto.
let echo = EchoClient(
  connection: connection,
  defaultCallOptions: defaultCallOptions  // optional
)

ClientConnection lifecycle

The ClientConnection is initialized with Configuration as described above. During initialization the NIO Channel for the connection is created and stored in an EventLoopFuture. The connection is created using the exponential backoff algorithm described by gRPC. The state of the connection is monitored (using the states defined by gRPC: idle, connecting, ready, transient failure, and shutdown) and will automatically reconnect (with backoff) if the channel is closed but the close was not initiated by the user.

Users may optionally provide a connectivity state delegate to observe these changes:

public protocol ConnectivityStateDelegate {
  func connectivityStateDidChange(from oldState: ConnectivityState, to newState: ConnectivityState)
}

NIO Pipeline

The client’s channel pipeline follows:

NIOSSLHandler (if TLS is being used)
NIOHTTP2Handler
HTTP2StreamMultiplexer

Each call is made on an HTTP/2 stream channel whose pipeline is:

HTTP2ToHTTP1ClientCodec
HTTP1ToRawGRPCClientCodec: translates HTTP/1 types into gRPC metadata and length-prefixed messages, it also handles request/response state, compression (not yet implemented) and message buffering since messages may span multiple frames. It is “Raw” since the emitted messages are just bytes and not yet typed.
GRPCClientCodec: handles encoding/decoding of messages.
ClientRequestChannelHandler: handles sending messages from the client, has unary and streaming versions.
ClientResponseChannelHandler: handles receiving messages from the server, has unary and streaming versions. Holds the promises for the varying futures exposed in the ClientCall protocols. It also holds the logic for timing out calls and handling errors.

A few notes:

The HTTP2 to HTTP1 translation simplifies implementation of the HTTP1ToRawGRPCClientCodec
Other handlers exist in the pipeline for error handling and verification (i.e.
TLS handshake was successful and a valid protocol was negotiated) but were omitted for brevity.
The differences between the four call types are just in their construction and request and response handlers.

Apple Platform Support

The library also provides a means to run using NIO Transport Services instead where it’s supported on Apple platforms. The user only has to provide a correctly typed EventLoopGroup in their Configuration and we’ll pick the appropriate bootstrap. To aid this we provide some utility functions:

public enum NetworkPreference {
  // NIOTS when available, NIO otherwise
  case best
  // Pick manually
  case userDefined(NetworkImplementation).
}

public enum NetworkImplementation {
  // i.e. NIOTS (this has the appropriate @available/#if canImport(Network))
  case networkFramework
  // i.e. NIO
  case posix
}

public enum PlatformSupport {
  // Returns an EventLoopGroup of the appropriate type based on user preference.
  public static func makeEventLoopGroup(
    loopCount: Int,
    networkPreference: NetworkPreference = .best
  ) -> EventLoopGroup {
    // ...
  }
}

One thing which should be called out is that the TLS support is provided by SwiftNIO SSL, even when Network.framework is being used. Ideally we would provide TLS via Network.framework if it’s being used, however abstracting over the different configuration for the two interfaces is not trivial.

Possible Areas of Improvement

Much of the configuration in the gRPC core library is via "channel arguments". Some of these options are surfaced via Configuration and CallOptions, however, providing a mechanism where new options can be added without breaking API (i.e. adding an additional field to a struct) would be beneficial.
Removing the HTTP1 code from the client pipeline may yield a small performance improvement. This is purely implementation, however, and could be done at any time.
The server cannot be configured to choose the level of support for gRPC Web (that is: gRPC Web is always supported), making this configurable would be better for users who only want to support standard gRPC.
Neither the client or server can have the bootstraps easily configured.
When using NIO Transport Services on Apple Platforms, TLS is always provided via NIO's NIOSSLHandler and not via Network.framework.

Maturity Justification

TODO

Explain why this solution should be accepted at the proposed maturity level.

Alternatives considered

TODO

Describe alternative approaches to addressing the same problem, and why you chose this approach instead.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

SSWG Proposal

gRPC Swift

Package Description

Introduction

Motivation

Proposed solution

gRPC Background

Server

Client

Code Generation

Out of Scope

Detailed design

Server

Client

Apple Platform Support

Possible Areas of Improvement

Maturity Justification

Alternatives considered

Clone this wiki locally