Conjure Java Runtime (formerly http-remoting)

This repository provides an opinionated set of libraries for defining and creating RESTish/RPC servers and clients based on Feign as a client and Dropwizard/Jersey with JAX-RS service definitions as a server. Refer to the API Contract section for details on the contract between clients and servers. This library requires Java 8.

Core libraries:

• conjure-java-jaxrs-client: Clients for JAX-RS-defined service interfaces
• conjure-java-jersey-server: Configuration library for Dropwizard/Jersey servers
• conjure-java-runtime-api: API classes for service configuration, tracing, and error propagation

Usage

Maven artifacts are published to Maven Central. Example Gradle dependency configuration:

repositories {
  mavenCentral()
}

dependencies {
  compile "com.palantir.conjure.java.runtime:conjure-java-jaxrs-client:$version"
  compile "com.palantir.conjure.java.runtime:conjure-java-jersey-server:$version"
}

conjure-java-jaxrs-client

Provides the JaxRsClient factory for creating Feign-based clients for JAX-RS APIs. SSL configuration is mandatory for all clients, plain-text HTTP is not supported. Example:

SslConfiguration sslConfig = SslConfiguration.of(Paths.get("path/to/trustStore"));
UserAgent userAgent = UserAgent.of(UserAgent.Agent.of("my-user-agent", "1.0.0"));
ClientConfiguration config = ClientConfigurations.of(
        ImmutableList.of(SERVER_URI),
        SslSocketFactories.createSslSocketFactory(sslConfig),
        SslSocketFactories.createX509TrustManager(sslConfig));
HostMetricsRegistry hostMetricsRegistry = new HostMetricsRegistry();  // can call .getMetrics() and then collect them to a central metrics repository
MyService service = JaxRsClient.create(MyService.class, userAgent, hostMetricsRegistry, config);

The JaxRsClient#create factory comes in two flavours: one for creating immutable clients given a fixed ClientConfiguration, and one for creating mutable clients whose configuration (e.g., server URLs, timeouts, SSL configuration, etc.) changes when the underlying ClientConfiguration changes.

conjure-java-jersey-server

Provides Dropwizard/Jersey configuration for handling conjure types, and also exception mappers for translating common runtime exceptions as well as our own ServiceException (see the errors section) to appropriate HTTP error codes. A Dropwizard server is configured for conjure as follows:

public class MyServer extends Application<Configuration> {
    @Override
    public final void run(Configuration config, final Environment env) throws Exception {
        env.jersey().register(ConjureJerseyFeature.INSTANCE);
        env.jersey().register(new MyResource());
    }
}

tracing

Provides Zipkin-style call tracing libraries. All JaxRsClient instances are instrumented by default. Jersey server instrumentation is enabled via the ConjureJerseyFeature (see above).

Please refer to tracing-java for more details on the tracing library usage.

service-config (conjure-java-runtime-api)

Provides utilities for setting up service clients from file-based configuration. Example:

# config.yml
services:
  security:
    # default truststore for all clients
    trustStorePath: path/to/trustStore.jks
  myService:  # the key used in `factory.get("myService")` below
    uris:
      - URI
    # optionally set a myService-specific truststore
    # security:
    #   trustStorePath: path/to/trustStore.jks

ServiceConfigBlock config = readFromYaml("config.yml");
ServiceConfigurationFactory factory = ServiceConfigurationFactory.of(config);
HostMetricsRegistry hostMetricsRegistry = new HostMetricsRegistry();
MyService client = JaxRsClient.create(MyService.class, UserAgents.parse("my-agent"), hostMetricsRegistry, ClientConfigurations.of(factory.get("myService")));

keystores and ssl-config (conjure-java-runtime-api)

Provides utilities for interacting with Java trust stores and key stores and acquiring SSLSocketFactory instances using those stores, as well as a configuration class for use in server configuration files.

The SslConfiguration class specifies the configuration that should be used for a particular SSLContext. The configuration is required to include information for creating a trust store and can optionally be provided with information for creating a key store (for client authentication).

The configuration consists of the following properties:

• trustStorePath: path to a file that contains the trust store information. The format of the file is specified by the trustStoreType property.
• trustStoreType: the type of the trust store. See section below for details. The default value is JKS.
• (optional) keyStorePath: path to a file that contains the key store information. If unspecified, no key store will be associated with this configuration.
• (optional) keyStorePassword: password for the key store. Will be used to read the keystore provided by keyStorePath (if relevant for the format), and will also be used as the password for the in-memory key store created by this configuration. Required if keyStorePath is specified.
• (optional) keyStoreType: the type of the key store. See section below for details. The default value is JKS.
• (optional) keyStoreAlias: specifies the alias of the key that should be read from the key store (relevant for file formats that contain multiple keys). If unspecified, the first key returned by the store is used.

An SslConfiguration object can be constructed using the static of() factory methods of the class, or by using the SslConfiguration.Builder builder. SslConfiguration objects can be serialized and deserialized as JSON.

Once an SslConfiguration object is obtained, it can be passed as an argument to the SslSocketFactories.createSslSocketFactory method to create an SSLSocketFactory object that can be used to configure Java SSL connections.

Store Types

The following values are supported as store types:

• JKS: a trust store or key store in JKS format. When used as a trust store, the TrustedCertificateEntry entries are used as certificates. When used as a key store, the PrivateKeyEntry specified by the keyStoreAlias parameter (or the first such entry returned if the parameter is not specifeid) is used as the private key.
• PEM: for trust stores, an X.509 certificate file in PEM format, or a directory of such files. For key stores, a PEM file that contains a PKCS#1 RSA private key or PKCS#8 followed by the certificates that form the trust chain for the key in PEM format, or a directory of such files. In either case, if a directory is specified, every non-hidden file in the directory must be a file of the specified format (they will all be read).
• PKCS12: a trust store or key store in PKCS12 format. Behavior is the same as for the JKS type, but operates on stores in PKCS12 format.
• Puppet: a directory whose content conforms to the Puppet SSLdir format. For trust stores, the certificates in the certs directory are added to the trust store. For key stores, the PEM files in the private_keys directory are added as the private keys and the corresponding files in certs are used as the trust chain for the key.

errors (conjure-java-runtime-api)

Provides utilities for relaying service errors across service boundaries (see below).

API Contract

conjure-java-runtime makes the following opinionated customizations to the standard Dropwizard/Feign behavior.

Object serialization/deserialization

All parameters and return values of application/json endpoints are serialized/deserialized to/from JSON using a Jackson ObjectMapper with GuavaModule, ShimJdk7Module (same as Jackson’s Jdk7Module, but avoids Jackson 2.6 requirement) and Jdk8Module. Servers must not expose parameters or return values that cannot be handled by this object mapper.

Error propagation

Servers should use the ServiceException class to propagate application-specific errors to its callers. The ServiceException class exposes standard error codes that clients can handle in a well-defined manner; further, ServiceException implements SafeLoggable and thus allows logging infrastructure to handle "unsafe" and "safe" exception parameters appropriately. Typically, services define its error types as follows:

class Errors {
  private static final ErrorType DATASET_NOT_FOUND =
    ErrorType.create(ErrorType.Code.INVALID_ARGUMENT, "MyApplication:DatasetNotFound");

  static ServiceException datasetNotFound(DatasetId datasetId, String userName) {
    // Both the safe and unsafe params will be sent back to the client; the client needs
    // to decide themselves if any of these are safe to log. We're only marking things as
    // safe or unsafe for this server to log.
    return new ServiceException(
            DATASET_NOT_FOUND, SafeArg.of("datasetId", datasetId), UnsafeArg.of("userName", userName));
  }
}

void someMethod(String datasetId, String userName) {
  if (!exists(datasetId)) {
    throw Errors.datasetNotFound(datasetId, userName);
  }
}

The ConjureJerseyFeature installs an exception mapper for ServiceException. The exception mapper sets the response media type to application/json and returns as response body a JSON representation of a SerializableError capturing the error code, error name, and error parameters. The resulting JSON response is:

{
  "errorCode": "INVALID_ARGUMENT",
  "errorName": "MyApplication:DatasetNotFound",
  "errorInstanceId": "xxxxxxxx-xxxx-Mxxx-Nxxx-xxxxxxxxxxxx",
  "parameters": {
    "datasetId": "123abc",
    "userName": "yourUserName"
  }
}

JaxRsClient intercepts non-successful HTTP responses and throw a RemoteException wrapping the deserialized server-side SerializableError. The error codes and names of the ServiceException and SerializableError are defined by the service API, and clients should handle errors based on the error code and name:

try {
    service.someMethod();
catch (RemoteException e) {
    if (e.getError().errorName().equals("MyApplication:DatasetNotFound")) {
        handleError(e.getError().parameters().get("datasetId"));
    } else {
        throw new RuntimeException("Failed to call someMethod()", e);
    }
}

Frontends receiving such errors should use a combination of error code, error name, and parameters to display localized, user friendly error information. For example, the above error could be surfaced as "The requested dataset with id 123abc could not be found".

To support legacy server implementations, the ConjureJerseyFeature also installs exception mappers for IllegalArgumentException, NoContentException, RuntimeException and WebApplicationException. The exceptions typically yield SerializableErrors with exceptionClass=errorCode=<exception classname> and message=errorName=<exception message>. Clients should refrain from displaying the message or errorName fields to user directly. Services should prefer to throw ServiceExceptions instead of the above, since they are easier to consume for clients and support transmitting exception parameters in a safe way.

RemoteException vs ServiceException vs SerializableError vs ErrorType

• ErrorType is a record type, meant to be used as 'compile time constants' - essentially used by services to define the 'enum' of their service exceptions
• SerializableError defines the wire format for serializing ServiceExceptions in HTTP response bodies and contains the error code, error instance id, and application-defined parameters
• ServiceException is a final subclass of Exception, thrown by the server
• RemoteException is what the client sees if a remote call results in the server internally throwing a ServiceException

The workflow is:

• Server code throws an instance of ServiceException, containing some ErrorType
• The com.palantir.conjure.java.server.jersey.ServiceExceptionMapper exception mapper
  • determines the response code for this service exception
  • converts this into a SerializableError
  • serializes this into the response body as JSON
• The client sees a RemoteException, which contains the SerializableError which was sent over the wire
• The client can inspect the SerializableError and choose to act
• If the client is itself a server and does not handle the RemoteException, a SerializableError error will be sent as the response with and errorCode of INTERNAL, errorName of Default:Internal, the same errorInstanceId as the original RemoteException and no parameters.

Serialization of Optional and Nullable objects

@Nullable or Optional<?> fields in complex types are serialized using the standard Jackson mechanism:

• a present value is serialized as itself (in particular, without being wrapped in a JSON object representing the Optional object)
• an absent value is serialized as a JSON null. For example, assume a Java type of the form

public final class ComplexType {
    private final Optional<ComplexType> nested;
    private final Optional<String> string;
}

, and an instance

ComplexType value = new ComplexType(
        Optional.of(
                new ComplexType(
                        Optional.<ComplexType>absent(),
                        Optional.<String>absent(),
        Optional.of("baz"));

The JSON-serialized representation of this object is:

{"nested":{"nested":null,"string":null},"string":"baz"}

Optional return values

When a call to a service interface declaring an Optional<T> return value with media type application/json yields:

• a Optional#empty return value, then the HTTP response has error code 204 and an empty response body.
• a non-empty return value, then the HTTP response has error code 200 and the body carries the deserialized T object directly, rather than a deserialized Optional<T> object.

JaxRsClients intercept such responses, deserialize the T-typed return value and return it to the caller wrapped as an Optional<T>.

Call tracing

Clients and servers propagate call trace ids across JVM boundaries according to the Zipkin specification. In particular, clients insert X-B3-TraceId: <Trace ID> HTTP headers into all requests which get propagated by Jetty servers into subsequent client invocations.

Endpoints returning plain strings

Endpoints returning plain strings should produce media type text/plain. Return type Optional<String> is only supported for media type application/json.

Quality of service: retry, failover, throttling, backpressure

Flow control in Conjure is a collaborative effort between servers and clients.

Servers advertise an overloaded state using 429/503 responses, which clients interpret by throttling the number of in-flight requests they will send (currently according to an additive increase, multiplicative decrease based algorithm). Requests are retried a fixed number of times, scheduled with an exponential backoff algorithm.

Concurrency permits are only released when the response body is closed, so large streaming responses are correctly tracked.

conjure-java-runtime servers can use the QosException class to advertise the following conditions:

• throttle: Returns a Throttle exception indicating that the calling client should throttle its requests. The client may retry against an arbitrary node of this service.
• retryOther: Returns a RetryOther exception indicating that the calling client should retry against the given node of this service.
• unavailable: An exception indicating that (this node of) this service is currently unavailable and the client may try again at a later time, possibly against a different node of this service.

The QosExceptions have a stable mapping to HTTP status codes and response headers:

• throttle: 429 Too Many Requests, plus optional Retry-After header
• retryOther: 308 Permanent Redirect, plus Location header indicating the target host
• unavailable: 503 Unavailable

conjure-java-runtime clients handle the above error codes and take the appropriate action:

• throttle: reschedule the request with a delay: either the indicated Retry-After period, or a configured exponential backoff
• retryOther: retry the request against the indicated service node; all request parameters and headers are maintained
• unavailable: retry the request on a different host after a configurable exponential delay

Additionally, connection errors (e.g., connection refused or DNS errors) yield a retry against a different node of the service. Retries pick a target host by cycling through the list of URLs configured for a Service (see ClientConfiguration#uris). Note that the "current" URL is maintained across calls; for example, if a first call yields a retryOther/308 redirect, then any subsequent calls will be made against that URL. Similarly, if the first URL yields a DNS error and the retried call succeeds against the URL from the list, then subsequent calls are made against that URL.

The number of retries for 503 and connection errors can be configured via ClientConfiguration#maxNumRetries or ServiceConfiguration#maxNumRetries, defaulting to 4.

Metrics

The HostMetricsRegistry uses HostMetrics to track per-host response metrics. HostMetrics provides the following metrics:

• get1xx(): A timer of 1xx responses.
• get2xx(): A timer of 2xx responses.
• get3xx(): A timer of 3xx responses.
• get4xx(): A timer of 4xx responses, excluding 429s.
• get5xx(): A timer of 5xx responses, excluding 503s.
• getQos(): A timer of 429 and 503 responses.
• getOther(): A timer of all other responses.
• getIoExceptions(): A timer of all failed requests.

License

This repository is made available under the Apache 2.0 License.