Object serialization

Introduction

Object serialization is the process of converting objects into a stream of bytes and, deserialization, the reverse process of creating objects from a stream of bytes. It takes place every time nodes pass objects to each other as messages, when objects are sent to or from RPC clients from the node, and when we store transactions in the database.

Corda pervasively uses a custom form of type safe binary serialization. This stands in contrast to some other systems that use weakly or untyped string-based serialization schemes like JSON or XML. The primary drivers for this were:

• A desire to have a schema describing what has been serialized alongside the actual data:

• To assist with versioning, both in terms of being able to interpret data archived long ago (e.g. trades from a decade ago, long after the code has changed) and between differing code versions.

• To make it easier to write generic code e.g. user interfaces that can navigate the serialized form of data.

• To support cross platform (non-JVM) interaction, where the format of a class file is not so easily interpreted.

• A desire to use a documented and static wire format that is platform independent, and is not subject to change with 3rd party library upgrades, etc.

• A desire to support open-ended polymorphism, where the number of subclasses of a superclass can expand over time and the subclasses do not need to be defined in the schema upfront. This is key to many Corda concepts, such as states.

• Increased security by constructing deserialized objects through supported constructors, rather than having data inserted directly into their fields without an opportunity to validate consistency or intercept attempts to manipulate supposed invariants.

• Binary formats work better with digital signatures than text based formats, as there’s much less scope for changes that modify syntax but not semantics.

Whitelisting

In classic Java serialization, any class on the JVM classpath can be deserialized. This is a source of exploits and vulnerabilities that exploit the large set of third-party libraries that are added to the classpath as part of a JVM application’s dependencies and carefully craft a malicious stream of bytes to be deserialized. In Corda, we strictly control which classes can be deserialized (and, pro-actively, serialized) by insisting that each (de)serializable class is part of a whitelist of allowed classes.

To add a class to the whitelist, you must use either of the following mechanisms:

• Add the @CordaSerializable annotation to the class. This annotation can be present on the class itself, on any super class of the class, on any interface implemented by the class or its super classes, or any interface extended by an interface implemented by the class or its super classes.
• Implement the SerializationWhitelist interface and specify a list of whitelisted classes.

There is also a built-in Corda whitelist (see the DefaultWhitelist class) that whitelists common JDK classes for convenience. This whitelist is not user-editable.

The annotation is the preferred method for whitelisting. An example is shown in tutorial-clientrpc-api. It’s reproduced here as an example of both ways you can do this for a couple of example classes.

// Not annotated, so need to whitelist manually.
data class ExampleRPCValue(val foo: String)

// Annotated, so no need to whitelist manually.
@CordaSerializable
data class ExampleRPCValue2(val bar: Int)

class ExampleRPCSerializationWhitelist : SerializationWhitelist {
    // Add classes like this.
    override val whitelist = listOf(ExampleRPCValue::class.java)
}

AMQP

Corda uses an extended form of AMQP 1.0 as its binary wire protocol. You can learn more about the Wire format Corda uses if you intend to parse Corda messages from non-JVM platforms.

Corda serialization is currently used for:

• Peer-to-peer networking.
• Persisted messages, like signed transactions and states.

For the checkpointing of flows Corda uses a private scheme that is subject to change. It is currently based on the Kryo framework, but this may not be true in future.

This separation of serialization schemes into different contexts allows us to use the most suitable framework for that context rather than attempting to force a one-size-fits-all approach. Kryo is more suited to the serialization of a program’s stack frames, as it is more flexible than our AMQP framework in what it can construct and serialize. However, that flexibility makes it exceptionally difficult to make secure. Conversely, our AMQP framework allows us to concentrate on a secure framework that can be reasoned about and thus made safer, with far fewer security holes.

Selection of serialization context should, for the most part, be opaque to CorDapp developers, the Corda framework selecting the correct context as configured.

This document describes what is currently and what will be supported in the Corda AMQP format from the perspective of CorDapp developers, to allow CorDapps to take into consideration the future state. The AMQP serialization format will continue to apply the whitelisting functionality that is already in place and described in this page.

Core Types

This section describes the classes and interfaces that the AMQP serialization format supports.

Collection Types

The following collection types are supported. Any implementation of the following will be mapped to an implementation of the interface or class on the other end. For example, if you use a Guava implementation of a collection, it will deserialize as the primitive collection type.

The declared types of properties should only use these types, and not any concrete implementation types (e.g. Guava implementations). Collections must specify their generic type, the generic type parameters will be included in the schema, and the element’s type will be checked against the generic parameters when deserialized.

java.util.Collection
java.util.List
java.util.Set
java.util.SortedSet
java.util.NavigableSet
java.util.NonEmptySet
java.util.Map
java.util.SortedMap
java.util.NavigableMap

However, as a convenience, we explicitly support the concrete implementation types below, and they can be used as the declared types of properties.

java.util.LinkedHashMap
java.util.TreeMap
java.util.EnumSet
java.util.EnumMap (but only if there is at least one entry)

JVM primitives

All the primitive types are supported.

boolean
byte
char
double
float
int
long
short

Arrays

Arrays of any type are supported, primitive or otherwise.

JDK Types

The following JDK library types are supported:

java.io.InputStream

java.lang.Boolean
java.lang.Byte
java.lang.Character
java.lang.Class
java.lang.Double
java.lang.Float
java.lang.Integer
java.lang.Long
java.lang.Short
java.lang.StackTraceElement
java.lang.String
java.lang.StringBuffer

java.math.BigDecimal

java.security.PublicKey

java.time.DayOfWeek
java.time.Duration
java.time.Instant
java.time.LocalDate
java.time.LocalDateTime
java.time.LocalTime
java.time.Month
java.time.MonthDay
java.time.OffsetDateTime
java.time.OffsetTime
java.time.Period
java.time.YearMonth
java.time.Year
java.time.ZonedDateTime
java.time.ZonedId
java.time.ZoneOffset

java.util.BitSet
java.util.Currency
java.util.UUID

Third-Party Types

The following 3rd-party types are supported:

kotlin.Unit
kotlin.Pair

org.apache.activemq.artemis.api.core.SimpleString

Corda Types

Any classes and interfaces in the Corda codebase annotated with @CordaSerializable are supported.

All Corda exceptions that are expected to be serialized inherit from CordaThrowable via either CordaException (for checked exceptions) or CordaRuntimeException (for unchecked exceptions). Any Throwable that is serialized but does not conform to CordaThrowable will be converted to a CordaRuntimeException, with the original exception type and other properties retained within it.

Custom Types

You own types must adhere to the following rules to be supported:

Classes

General Rules

• The class must be compiled with parameter names included in the .class file. This is the default in Kotlin but must be turned on in Java using the -parameters command line option to javac

  In circumstances where classes cannot be recompiled, such as when using a third-party library, a proxy serializer can be used to avoid this problem. Details on creating such an object can be found on the Pluggable Serializers for CorDapps page.

• The class must be annotated with @CordaSerializable

• The declared types of constructor arguments, getters, and setters must be supported, and where generics are used, the generic parameter must be a supported type, an open wildcard (*), or a bounded wildcard which is currently widened to an open wildcard

• Any superclass must adhere to the same rules, but can be abstract

• Object graph cycles are not supported, so an object cannot refer to itself, directly or indirectly

Constructor Instantiation

The primary way Corda’s AMQP serialization framework instantiates objects is via a specified constructor. This is used to first determine which properties of an object are to be serialized, then, on deserialization, it is used to instantiate the object with the serialized values.

It is recommended that serializable objects in Corda adhere to the following rules, as they allow immutable state objects to be deserialized:

• A Java Bean getter for each of the properties in the constructor, with a name of the form getX. For example, for a constructor parameter foo, there must be a getter called getFoo(). If foo is a boolean, the getter may optionally be called isFoo() (this is why the class must be compiled with parameter names turned on)
• A constructor which takes all of the properties that you wish to record in the serialized form. This is required in order for the serialization framework to reconstruct an instance of your class
• If more than one constructor is provided, the serialization framework needs to know which one to use. The @ConstructorForDeserialization annotation can be used to indicate which one. For a Kotlin class, without the @ConstructorForDeserialization annotation, the primary constructor will be selected

In Kotlin, this maps cleanly to a data class where there getters are synthesized automatically. For example, suppose we have the following data class:

data class Example (val a: Int, val b: String)

Properties a and b will be included in the serialized form.

However, properties not mentioned in the constructor will not be serialized. For example, in the following code, property c will not be considered part of the serialized form:

data class Example (val a: Int, val b: String) {
    var c: Int = 20
}

var e = Example (10, "hello")
e.c = 100;

val e2 = e.serialize().deserialize() // e2.c will be 20, not 100!!!

Setter Instantiation

As an alternative to constructor-based initialisation, Corda can also determine the important elements of an object by inspecting the getter and setter methods present on the class. If a class has only a default constructor and properties then the serializable properties will be determined by the presence of both a getter and setter for that property that are both publicly visible (i.e. the class adheres to the classic idiom of mutable JavaBeans).

On deserialization, a default instance will first be created, and then the setters will be invoked on that object to populate it with the correct values.

For example:

class Example(var a: Int, var b: Int, var c: Int)

class Example {
    private int a;
    private int b;
    private int c;

    public int getA() { return a; }
    public int getB() { return b; }
    public int getC() { return c; }

    public void setA(int a) { this.a = a; }
    public void setB(int b) { this.b = b; }
    public void setC(int c) { this.c = c; }
}

Inaccessible Private Properties

Whilst the Corda AMQP serialization framework supports private object properties without publicly accessible getter methods, this development idiom is strongly discouraged.

For example.

class C(val a: Int, private val b: Int)

class C {
    public Integer a;
    private Integer b;

    public C(Integer a, Integer b) {
        this.a = a;
        this.b = b;
    }
}

When designing Corda states, it should be remembered that they are not, despite appearances, traditional OOP style objects. They are signed over, transformed, serialized, and relationally mapped. As such, all elements should be publicly accessible by design.

Providing a public getter, as per the following example, is acceptable:

class C(val a: Int, b: Int) {
    var b: Int = b
       private set
}

class C {
    public Integer a;
    private Integer b;

    C(Integer a, Integer b) {
        this.a = a;
        this.b = b;
    }

    public Integer getB() {
        return b;
    }
}

Mismatched Class Properties / Constructor Parameters

Consider an example where you wish to ensure that a property of class whose type is some form of container is always sorted using some specific criteria yet you wish to maintain the immutability of the class.

This could be codified as follows:

@CordaSerializable
class ConfirmRequest(statesToConsume: List<StateRef>, val transactionId: SecureHash) {
    companion object {
        private val stateRefComparator = compareBy<StateRef>({ it.txhash }, { it.index })
    }

    private val states = statesToConsume.sortedWith(stateRefComparator)
}

The intention in the example is to always ensure that the states are stored in a specific order regardless of the ordering of the list used to initialise instances of the class. This is achieved by using the first constructor parameter as the basis for a private member. However, because that member is not mentioned in the constructor (whose parameters determine what is serializable as discussed above) it would not be serialized. In addition, as there is no provided mechanism to retrieve a value for statesToConsume we would fail to build a serializer for this Class.

In this case a secondary constructor annotated with @ConstructorForDeserialization would not be a valid solution as the two signatures would be the same. Best practice is thus to provide a getter for the constructor parameter which explicitly associates it with the actual member variable.

@CordaSerializable
class ConfirmRequest(statesToConsume: List<StateRef>, val transactionId: SecureHash) {
    companion object {
        private val stateRefComparator = compareBy<StateRef>({ it.txhash }, { it.index })
    }

    private val states = statesToConsume.sortedWith(stateRefComparator)

    //Explicit "getter" for a property identified from the constructor parameters
    fun getStatesToConsume() = states
}

Mutable Containers

Because Java fundamentally provides no mechanism by which the mutability of a class can be determined this presents a problem for the serialization framework. When reconstituting objects with container properties (lists, maps, etc) we must chose whether to create mutable or immutable objects. Given the restrictions, we have decided it is better to preserve the immutability of immutable objects rather than force mutability on presumed immutable objects.

data class C(val l : MutableList<String>)

val bytes = C(mutableListOf ("a", "b", "c")).serialize()
val newC = bytes.deserialize()

newC.l.add("d")

The call to newC.l.add will throw an UnsupportedOperationException.

There are several workarounds that can be used to preserve mutability on reconstituted objects. Firstly, if the class isn’t a Kotlin data class and thus isn’t restricted by having to have a primary constructor.

class C {
    val l : MutableList<String>

    @Suppress("Unused")
    constructor (l : MutableList<String>) {
        this.l = l.toMutableList()
    }
}

val bytes = C(mutableListOf ("a", "b", "c")).serialize()
val newC = bytes.deserialize()

// This time this call will succeed
newC.l.add("d")

Secondly, if the class is a Kotlin data class, a secondary constructor can be used.

data class C (val l : MutableList<String>){
    @ConstructorForDeserialization
    @Suppress("Unused")
    constructor (l : Collection<String>) : this (l.toMutableList())
}

val bytes = C(mutableListOf ("a", "b", "c")).serialize()
val newC = bytes.deserialize()

// This will also work
newC.l.add("d")

Thirdly, to preserve immutability of objects (a recommend design principle - Copy on Write semantics) then mutating the contents of the class can be done by creating a new copy of the data class with the altered list passed (in this example) passed in as the Constructor parameter.

data class C(val l : List<String>)

val bytes = C(listOf ("a", "b", "c")).serialize()
val newC = bytes.deserialize()

val newC2 = newC.copy (l = (newC.l + "d"))

Enums

All enums are supported, provided they are annotated with @CordaSerializable. Corda supports interoperability of enumerated type versions. This allows such types to be changed over time without breaking backward (or forward) compatibility. The rules and mechanisms for doing this are discussed in Enum Evolution.

Exceptions

The following rules apply to supported Throwable implementations.

• If you wish for your exception to be serializable and transported type safely it should inherit from either CordaException or CordaRuntimeException
• If not, the Throwable will deserialize to a CordaRuntimeException with the details of the original Throwable contained within it, including the class name of the original Throwable

Kotlin Objects

Kotlin’s non-anonymous object s (i.e. constructs like object foo : Contract {...}) are singletons and treated differently. They are recorded into the stream with no properties, and deserialize back to the singleton instance. Currently, the same is not true of Java singletons, which will deserialize to new instances of the class. This is hard to fix because there’s no perfectly standard idiom for Java singletons.

Kotlin’s anonymous object s (i.e. constructs like object : Contract {...}) are not currently supported and will not serialize correctly. They need to be re-written as an explicit class declaration.

Class synthesis

Corda serialization supports dynamically synthesising classes from the supplied schema when deserializing, without the supporting classes being present on the classpath. This can be useful where generic code might expect to be able to use reflection over the deserialized data, for scripting languages that run on the JVM, and also for ensuring classes not on the classpath can be deserialized without loading potentially malicious code.

If the original class implements some interfaces then the carpenter will make sure that all of the interface methods are backed by fields. If that’s not the case then an exception will be thrown during deserialization. This check can be turned off with SerializationContext.withLenientCarpenter. This can be useful if only the field getters are needed, say in an object viewer.

Calculated values

In some cases, for example the exitKeys field in FungibleState, a property in an interface may normally be implemented as a calculated value, with a “getter” method for reading it but neither a corresponding constructor parameter nor a “setter” method for writing it. In this case, it will not automatically be included among the properties to be serialized, since the receiving class would ordinarily be able to re-calculate it on demand. However, a synthesized class will not have the method implementation which knows how to calculate the value, and a cast to the interface will fail because the property is not serialized and so the “getter” method present in the interface will not be synthesized.

The solution is to annotate the method with the SerializableCalculatedProperty annotation, which will cause the value exposed by the method to be read and transmitted during serialization, but discarded during normal deserialization. The synthesized class will then include a backing field together with a “getter” for the serialized calculated value, and will remain compatible with the interface.

If the annotation is added to the method in the interface, then all implementing classes must calculate the value and none may have a corresponding backing field; alternatively, it can be added to the overriding method on each implementing class where the value is calculated and there is no backing field. If the field is a Kotlin val, then the annotation should be targeted at its getter method, e.g. @get:SerializableCalculatedProperty.

Future enhancements

Possible future enhancements include:

• Java singleton support. We will add support for identifying classes which are singletons and identifying the static method responsible for returning the singleton instance
• Instance internalizing support. We will add support for identifying classes that should be resolved against an instances map to avoid creating many duplicate instances that are equal (similar to String.intern())

Type Evolution

Type evolution is the mechanism by which classes can be altered over time yet still remain serializable and deserializable across all versions of the class. This ensures an object serialized with an older idea of what the class “looked like” can be deserialized and a version of the current state of the class instantiated.

More detail can be found in Default Class Evolution.