"Providence Services"

Services (a.k.a. thrift services) have a couple of implicit handling around it. The way this is organized in providence is by defining as much as possible around the service call handling in thrift as possible.

Service Definition

Services are in essence an interface defined solely from thrift messages and types. Example simple service:

service MyService {
    i32 myMethod(1: i32 param);
}

There can also be declared extra thrift exceptions on any service method, each being an exception type message, and is declared in a throws block after the method params, like:

exception MyException {
    1: string message;
}
struct MyRequest {
    1: optional string request_field;
}

struct MyResponse {
    1: optional string response_field;
}
service MyService {
    i32 myMethod(1: i32 param) throws (1: MyException me);
    MyResponse myStub(MyRequest);
}

Generated Classes

For each service a single Service class is generated. The service class itself has no functionality, but contains a number of inner classes that do.

  • [IFace]: An interface that defines what the service is. This has a 1-to-1 mapping in methods to the declared service methods.
  • [Client]: A class that implements the IFace interface, and takes a single PClientHandler instance as constructor argument. The class is the generated code bridge between the service interface and the client handling code found in providence-core-client used to communicate with actual remote services.
  • [Processor]: A class that implements the PProcessor interface, and takes a single IFace implementation as argument. This is the interface between a server implementation, e.g. ProvidenceServlet and the service implementation actually doing the job.
class MyService {
    public interface IFace {
        int myMethod(int param) throws MyException;
        MyResponse myStub(MyRequest request);
    }

    public static class Client implements IFace {
        public Client(PServiceCallHandler handler) {
            // ...
        }

        public int myMethod(int param) throws MyException {
            // transform the method call into a generic
            // call on handler.
        }

        public MyResponse myStub(MyRequest request) {
            // use the request as message for call on handler.
        }
    }

    public static class Processor implements PProcessor {
        public Processor(IFace impl) {
            // ...
        }

        public PServiceCall handleCall(PServiceCall call) {
            // ... handle the call and call iface.myMethod() and
            // iface.myStub()
        }
    }
}

Note that the PProcessor interface extends PServiceCallHandler, so it is possible but unhelpful to wrap an implementation in a Client/Processor pair, like IFace iface = new Client(new Processor(new MyServiceImpl()));, which is largely equivalent to just Iface iface = new MyServiceImpl();.

Service Interface

The service is initially declared as a simple interface which mimics the service declaration from the thrift file, and then logic is added around that. E.g. the the ApplicationException and IOException exceptions are added in the java client implementations.

The server side handling of the service is handled in providence-core-server for thrift over HTTP, and in providence-thrift-compat for the socket server and non-blocking socket server.

Service calls

Service calls are a way to handle a request and a response so that the actual code can work with a service as if it was "just an implementation" of the generated interface. The call is defined by a 'call type (id)', grouped as request or response.

There are two types of requests:

  • call (1): Normal method call request. Requires a response (even for void return types).
  • oneway (4): A void type call request that does not require a response, and can only throw an exception if the error occurs on the client side before the actual call message is sent.

And there are two types of responses:

  • reply (2): This handled both the "normal" response of the method using the "return" type of the method, or any of the declared exception types in the "throws" section of the service call definition.
  • exception (3): Also called 'application exception' is handling of exceptions from either the client handlers, or from the providence (or thrift) server handlers themselves.

Each actual service call message must have this information:

  • The name of the method called, as declared in the service, used by the service handler to call the correct service method, and handle message serialization.
  • A call type, must be one of the types above.
  • A sequence number, which e.g. may be used in a Muxer/DeMuxer used to route responses back the the correct caller. A response's sequence number must be the same as given by the request.
  • And a request, response or exception message, which can NOT be null itself.

Application Exceptions

This is a response type where the object is a non-declared exception. E.g. the service implementation threw an undeclared runtime exception, e.g. a NullPointerException or IllegalArgumentException, or serialization failed, or there was IO or network problems. This is wrapped into the "ApplicationException", which all service calls can throw.

The application exceptions are defined as:

enum ApplicationExceptionType {
  # Any exception not matching anything below.
  UNKNOWN                 =  0;
  # the requested method (name) is not known to the service.
  UNKNOWN_METHOD          =  1;
  # the serialized message type is not compatible with the locally declared message.
  INVALID_MESSAGE_TYPE    =  2;
  # the method name in the response does not match the call.
  WRONG_METHOD_NAME       =  3;
  # the response sequence ID does not match the one from the request.
  BAD_SEQUENCE_ID         =  4;
  # The response is empty (or null).
  MISSING_RESULT          =  5;
  # Unknown or unhandled exception from the client or service handler.
  INTERNAL_ERROR          =  6;
  # Serialization error.
  PROTOCOL_ERROR          =  7;
  # TODO: Figure out what this type means.
  # AFAIK it is related to transport data problems.
  INVALID_TRANSFORM       =  8;
  # Protocol mismatch.
  INVALID_PROTOCOL        =  9;
  # The client type (???) is not supported.
  UNSUPPORTED_CLIENT_TYPE = 10;
}

exception ApplicationException {
    # Textual message about the exception
    1: string message
    # The application exception type.
    2: ApplicationExceptionType id
}

Virtual Messages

Each service method will generate a virtual params struct, and each non-oneway service method will generate a virtual response union. These messages are used to wrap the called params and response data for the service method call.

The params struct is essentially mimicking the method params as a struct. E.g.:

service Service {
    void call(1: i32 i, 2: MyStruct s) throws (1: MyFailure fail);
}

Will generate the request message definition:

struct Service.call_request {
    1: i32 i;
    2: MyStruct s;
}

For the response, a virtual struct is generated containing the response on success, using the field ID 0 (which should not be allowed to be declared, and will not be used by the non-declared auto-IDs for the exceptions), and one field for each declared exception. E.g. the example above will generate this response message definition:

union Service.call_response {
    0: i32 success;
    1: MyFailure fail;
}

Note that since this is declared as a union, one (and only one) of the fields must be set to a non-null value for the response to be valid.