Persistence - Schema Evolution

Types of schema evolution

Before we explain the various techniques that can be used to safely evolve the schema of your persistent events over time, we first need to define what the actual problem is, and what the typical styles of changes are.

Since events are never deleted, we need to have a way to be able to replay (read) old events, in such way that does not force the PersistentActor to be aware of all possible versions of an event that it may have persisted in the past. Instead, we want the Actors to work on some form of "latest" version of the event and provide some means of either converting old "versions" of stored events into this "latest" event type, or constantly evolve the event definition - in a backwards compatible way - such that the new deserialization code can still read old events.

The most common schema changes you will likely are:

• adding a field to an event type,
• remove or rename field in event type,
• remove event type,
• split event into multiple smaller events.

The following sections will explain some patterns which can be used to safely evolve your schema when facing those changes.

Provided default serializers

Akka Persistence provides Google Protocol Buffers based serializers (using Akka Serialization) for it's own message types such as PersistentRepr, AtomicWrite and snapshots. Journal plugin implementations may choose to use those provided serializers, or pick a serializer which suits the underlying database better.

The below figure explains how the default serialization scheme works, and how it fits together with serializing the user provided message itself, which we will from here on refer to as the payload (highlighted in yellow):

Akka Persistence provided serializers wrap the user payload in an envelope containing all persistence-relevant information. If the Journal uses provided Protobuf serializers for the wrapper types (e.g. PersistentRepr), then the payload will be serialized using the user configured serializer, and if none is provided explicitly, Java serialization will be used for it.

The blue colored regions of the PersistentMessage indicate what is serialized using the generated protocol buffers serializers, and the yellow payload indicates the user provided event (by calling persist(payload)(...)). As you can see, the PersistentMessage acts as an envelope around the payload, adding various fields related to the origin of the event (persistenceId, sequenceNr and more).

More advanced techniques (e.g. Remove event class and ignore events) will dive into using the manifests for increasing the flexibility of the persisted vs. exposed types even more. However for now we will focus on the simpler evolution techniques, concerning simply configuring the payload serializers.

By default the payload will be serialized using Java Serialization. This is fine for testing and initial phases of your development (while you're still figuring out things and the data will not need to stay persisted forever). However, once you move to production you should really pick a different serializer for your payloads.

Configuring payload serializers

This section aims to highlight the complete basics on how to define custom serializers using Akka Serialization. Many journal plugin implementations use Akka Serialization, thus it is tremendously important to understand how to configure it to work with your event classes.

The below snippet explains in the minimal amount of lines how a custom serializer can be registered. For more in-depth explanations on how serialization picks the serializer to use etc, please refer to its documentation.

First we start by defining our domain model class, here representing a person:

public class Person {
  public final String name;
  public final String surname;
  public Person(String name, String surname) {
    this.name = name;
    this.surname = surname;
  }
}

Next we implement a serializer (or extend an existing one to be able to handle the new Person class):

/**
 * Simplest possible serializer, uses a string representation of the Person class.
 *
 * Usually a serializer like this would use a library like:
 * protobuf, kryo, avro, cap'n proto, flatbuffers, SBE or some other dedicated serializer backend
 * to perform the actual to/from bytes marshalling.
 */
public class SimplestPossiblePersonSerializer extends SerializerWithStringManifest {
  private final Charset utf8 = Charset.forName("UTF-8");

  private final String personManifest = Person.class.getName();

  // unique identifier of the serializer
  @Override public int identifier() {
    return 1234567;
  }

  // extract manifest to be stored together with serialized object
  @Override public String manifest(Object o) {
    return o.getClass().getName();
  }

  // serialize the object
  @Override public byte[] toBinary(Object obj) {
    if (obj instanceof Person) {
      Person p = (Person) obj;
      return (p.name + "|" + p.surname).getBytes(utf8);
    } else {
     throw new IllegalArgumentException(
      "Unable to serialize to bytes, clazz was: " + obj.getClass().getName());
    }
  }

  // deserialize the object, using the manifest to indicate which logic to apply
  @Override public Object fromBinary(byte[] bytes, String manifest) {
    if (personManifest.equals(manifest)) {
      String nameAndSurname = new String(bytes, utf8);
      String[] parts = nameAndSurname.split("[|]");
      return new Person(parts[0], parts[1]);
    } else {
      throw new IllegalArgumentException(
        "Unable to deserialize from bytes, manifest was: " + manifest + 
        "! Bytes length: " + bytes.length);
    }
  }

}

And finally we register the serializer and bind it to handle the docs.persistence.Person class:

# application.conf
akka {
  actor {
    serializers {
      person = "docs.persistence.SimplestPossiblePersonSerializer"
    }

    serialization-bindings {
      "docs.persistence.Person" = person
    }
  }
}

Deserialization will be performed by the same serializer which serialized the message initially because of the identifier being stored together with the message.

Please refer to the Akka Serialization documentation for more advanced use of serializers, especially the Serializer with String Manifest section since it is very useful for Persistence based applications dealing with schema evolutions, as we will see in some of the examples below.

Add fields

Situation: You need to add a field to an existing message type. For example, a SeatReservation(String letter, int row) now needs to have an associated code which indicates if it is a window or aisle seat.

Solution: Adding fields is the most common change you'll need to apply to your messages so make sure the serialization format you picked for your payloads can handle it apropriately, i.e. such changes should be binary compatible. This is easily achieved using the right serializer toolkit – we recommend something like Google Protocol Buffers or Apache Thrift however other tools may fit your needs just as well – picking a serializer backend is something you should research before picking one to run with. In the following examples we will be using protobuf, mostly because we are familiar with it, it does its job well and Akka is using it internally as well.

While being able to read messages with missing fields is half of the solution, you also need to deal with the missing values somehow. This is usually modeled as some kind of default value, or by representing the field as an Optional<T> See below for an example how reading an optional field from a serialized protocol buffers message might look like.

public enum SeatType {
  Window("W"), Aisle("A"), Other("O"), Unknown("");
  
  private final String code;
  
  private SeatType(String code) {
    this.code = code;
  }
  
  public static SeatType fromCode(String c) {
    if (Window.code.equals(c))
      return Window;
    else if (Aisle.code.equals(c))
      return Aisle;
    else if (Other.code.equals(c))
      return Other;
    else
      return Unknown;
  }
}
public class SeatReserved {
  public final String letter;
  public final int row;
  public final SeatType seatType;
  
  public SeatReserved(String letter, int row, SeatType seatType) {
    this.letter = letter;
    this.row = row;
    this.seatType = seatType;
  }
}

Next we prepare an protocol definition using the protobuf Interface Description Language, which we'll use to generate the serializer code to be used on the Akka Serialization layer (notice that the schema aproach allows us to easily rename fields, as long as the numeric identifiers of the fields do not change):

// FlightAppModels.proto
option java_package = "docs.persistence.proto";
option optimize_for = SPEED;

message SeatReserved {
  required string letter   = 1;
  required uint32 row      = 2;
  optional string seatType = 3; // the new field
}

The serializer implementation uses the protobuf generated classes to marshall the payloads. Optional fields can be handled explicitly or missing values by calling the has... methods on the protobuf object, which we do for seatType in order to use a Unknown type in case the event was stored before we had introduced the field to this event type:

/**
 * Example serializer impl which uses protocol buffers generated classes (proto.*)
 * to perform the to/from binary marshalling.
 */
public class AddedFieldsSerializerWithProtobuf extends SerializerWithStringManifest {
  @Override public int identifier() {
    return 67876;
  }

  private final String seatReservedManifest = SeatReserved.class.getName();

  @Override public String manifest(Object o){
    return o.getClass().getName();
  }

  @Override public Object fromBinary(byte[] bytes, String manifest) {
    if (seatReservedManifest.equals(manifest)) {
      // use generated protobuf serializer
      try {
        return seatReserved(FlightAppModels.SeatReserved.parseFrom(bytes));
      } catch (InvalidProtocolBufferException e) {
        throw new IllegalArgumentException(e.getMessage());
      } 
    } else {
        throw new IllegalArgumentException("Unable to handle manifest: " + manifest);
    }
  }

  @Override public byte[] toBinary(Object o) {
    if (o instanceof SeatReserved) {
      SeatReserved s = (SeatReserved) o;
      return FlightAppModels.SeatReserved.newBuilder()
        .setRow(s.row)
        .setLetter(s.letter)
        .setSeatType(s.seatType.code)
        .build().toByteArray();
      
    } else {
      throw new IllegalArgumentException("Unable to handle: " + o);
    }
  }

  // -- fromBinary helpers --

  private SeatReserved seatReserved(FlightAppModels.SeatReserved p) {
    return new SeatReserved(p.getLetter(), p.getRow(), seatType(p));
  }

  // handle missing field by assigning "Unknown" value
  private SeatType seatType(FlightAppModels.SeatReserved p) {
    if (p.hasSeatType()) 
      return SeatType.fromCode(p.getSeatType());
    else 
      return SeatType.Unknown;
  }

}

Rename fields

Situation: When first designing the system the SeatReverved event featured an code field. After some time you discover that what was originally called code actually means seatNr, thus the model should be changed to reflect this concept more accurately.

Solution 1 - using IDL based serializers: First, we will discuss the most efficient way of dealing with such kinds of schema changes – IDL based serializers.

IDL stands for Interface Description Language, and means that the schema of the messages that will be stored is based on this description. Most IDL based serializers also generate the serializer / deserializer code so that using them is not too hard. Examples of such serializers are protobuf or thrift.

Using these libraries rename operations are "free", because the field name is never actually stored in the binary representation of the message. This is one of the advantages of schema based serializers, even though that they add the overhead of having to maintain the schema. When using serializers like this, no additional code change (except renaming the field and method used during serialization) is needed to perform such evolution:

This is how such a rename would look in protobuf:

// protobuf message definition, BEFORE:
message SeatReserved {
  required string code = 1;
}

// protobuf message definition, AFTER:
message SeatReserved {
  required string seatNr = 1; // field renamed, id remains the same
}

It is important to learn about the strengths and limitations of your serializers, in order to be able to move swiftly and refactor your models fearlessly as you go on with the project.

Solution 2 - by manually handling the event versions: Another solution, in case your serialization format does not support renames as easily as the above mentioned formats, is versioning your schema. For example, you could have made your events carry an additional field called _version which was set to 1 (because it was the initial schema), and once you change the schema you bump this number to 2, and write an adapter which can perform the rename.

This approach is popular when your serialization format is something like JSON, where renames can not be performed automatically by the serializer. You can do these kinds of "promotions" either manually (as shown in the example below) or using a library like Stamina which helps to create those V1->V2->V3->...->Vn promotion chains without much boilerplate.

The following snippet showcases how one could apply renames if working with plain JSON (using a JsObject as an example JSON representation):

public class JsonRenamedFieldAdapter implements EventAdapter {
  // use your favorite json library
  private final ExampleJsonMarshaller marshaller = new ExampleJsonMarshaller();

  private final String V1 = "v1";
  private final String V2 = "v2";

  // this could be done independently for each event type
  @Override public String manifest(Object event) {
    return V2;
  }

  @Override public JsObject toJournal(Object event) {
    return marshaller.toJson(event);
  }

  @Override public EventSeq fromJournal(Object event, String manifest) {
    if (event instanceof JsObject) {
      JsObject json = (JsObject) event;
      if (V1.equals(manifest))
        json = rename(json, "code", "seatNr");
      return EventSeq.single(json);
    } else {  
      throw new IllegalArgumentException("Can only work with JSON, was: " + 
        event.getClass().getName());
    }
  }

  private JsObject rename(JsObject json, String from, String to) {
    // use your favorite json library to rename the field
    JsObject renamed = json; 
    return renamed;
  }

}

As you can see, manually handling renames induces some boilerplate onto the EventAdapter, however much of it you will find is common infrastructure code that can be either provided by an external library (for promotion management) or put together in a simple helper class.

Remove event class and ignore events

Situation: While investigating app performance you notice that insane amounts of CustomerBlinked events are being stored for every customer each time he/she blinks. Upon investigation you decide that the event does not add any value and should be deleted. You still have to be able to replay from a journal which contains those old CustomerBlinked events though.

Naive solution - drop events in EventAdapter:

The problem of removing an event type from the domain model is not as much its removal, as the implications for the recovery mechanisms that this entails. For example, a naive way of filtering out certain kinds of events from being delivered to a recovering PersistentActor is pretty simple, as one can simply filter them out in an EventAdapter:

The EventAdapter can drop old events (O) by emitting an empty EventSeq. Other events can simply be passed through (E).

This however does not address the underlying cost of having to deserialize all the events during recovery, even those which will be filtered out by the adapter. In the next section we will improve the above explained mechanism to avoid deserializing events which would be filtered out by the adapter anyway, thus allowing to save precious time during a recovery containing lots of such events (without actually having to delete them).

Improved solution - deserialize into tombstone:

In the just described technique we have saved the PersistentActor from receiving un-wanted events by filtering them out in the EventAdapter, however the event itself still was deserialized and loaded into memory. This has two notable downsides:

• first, that the deserialization was actually performed, so we spent some of out time budget on the deserialization, even though the event does not contribute anything to the persistent actors state.
• second, that we are unable to remove the event class from the system – since the serializer still needs to create the actuall instance of it, as it does not know it will not be used.

The solution to these problems is to use a serializer that is aware of that event being no longer needed, and can notice this before starting to deserialize the object.

This aproach allows us to remove the original class from our classpath, which makes for less "old" classes lying around in the project. This can for example be implemented by using an SerializerWithStringManifest (documented in depth in Serializer with String Manifest). By looking at the string manifest, the serializer can notice that the type is no longer needed, and skip the deserialization all-together:

The serializer is aware of the old event types that need to be skipped (O), and can skip deserializing them alltogether by simply returning a "tombstone" (T), which the EventAdapter converts into an empty EventSeq. Other events (E) can simply be passed through.

The serializer detects that the string manifest points to a removed event type and skips attempting to deserialize it:

public class EventDeserializationSkipped {
  public static EventDeserializationSkipped instance =
    new EventDeserializationSkipped();
  
  private EventDeserializationSkipped() {
  }
}

public class RemovedEventsAwareSerializer extends SerializerWithStringManifest {
  private final Charset utf8 = Charset.forName("UTF-8");
  private final String customerBlinkedManifest = "blinked";

  // unique identifier of the serializer
  @Override public int identifier() {
    return 8337;
  }

  // extract manifest to be stored together with serialized object
  @Override public String manifest(Object o) {
    if (o instanceof CustomerBlinked)
      return customerBlinkedManifest;
    else 
      return o.getClass().getName();
  }

  @Override public byte[] toBinary(Object o) {
    return o.toString().getBytes(utf8); // example serialization
  }

  @Override public Object fromBinary(byte[] bytes, String manifest) {
    if (customerBlinkedManifest.equals(manifest))
      return EventDeserializationSkipped.instance;
    else
     return new String(bytes, utf8);
  }
}

The EventAdapter we implemented is aware of EventDeserializationSkipped events (our "Tombstones"), and emits and empty EventSeq whenever such object is encoutered:

public class SkippedEventsAwareAdapter implements EventAdapter {
  @Override public String manifest(Object event) {
    return "";
  }
  
  @Override public Object toJournal(Object event) {
    return event;
  }

  @Override public EventSeq fromJournal(Object event, String manifest) {
    if (event == EventDeserializationSkipped.instance)
      return EventSeq.empty();
    else
      return EventSeq.single(event);
  }
}

Detach domain model from data model

Situation: You want to separate the application model (often called the "domain model") completely from the models used to persist the corresponding events (the "data model"). For example because the data representation may change independently of the domain model.

Another situation where this technique may be useful is when your serialization tool of choice requires generated classes to be used for serialization and deserialization of objects, like for example Google Protocol Buffers do, yet you do not want to leak this implementation detail into the domain model itself, which you'd like to model as plain Java classes.

Solution: In order to detach the domain model, which is often represented using pure java (case) classes, from the data model classes which very often may be less user-friendly yet highly optimised for throughput and schema evolution (like the classes generated by protobuf for example), it is possible to use a simple EventAdapter which maps between these types in a 1:1 style as illustrated below:

Domain events (A) are adapted to the data model events (D) by the EventAdapter. The data model can be a format natively understood by the journal, such that it can store it more efficiently or include additional data for the event (e.g. tags), for ease of later querying.

We will use the following domain and data models to showcase how the separation can be implemented by the adapter:

// Domain model - highly optimised for domain language and maybe "fluent" usage
public class Customer {
  public final String name;

  public Customer(String name) {
    this.name = name;
  }
}

public class Seat {
  public final String code;

  public Seat(String code) {
    this.code = code;
  }
  
  public SeatBooked bookFor(Customer customer) {
    return new SeatBooked(code, customer);
  }
}

public class SeatBooked {
  public final String code;
  public final Customer customer;
  
  public SeatBooked(String code, Customer customer) {
    this.code = code;
    this.customer = customer;
  }
}

// Data model - highly optimised for schema evolution and persistence
public class SeatBookedData {
  public final String code;
  public final String customerName;
  
  public SeatBookedData(String code, String customerName) {
    this.code = code;
    this.customerName = customerName;
  }
}

The EventAdapter takes care of converting from one model to the other one (in both directions), alowing the models to be completely detached from each other, such that they can be optimised independently as long as the mapping logic is able to convert between them:

class DetachedModelsAdapter implements EventAdapter {
  @Override public String manifest(Object event) {
    return "";
  }

  @Override public Object toJournal(Object event) {
    if (event instanceof SeatBooked) {
      SeatBooked s = (SeatBooked) event;
      return new SeatBookedData(s.code, s.customer.name);
    } else {
      throw new IllegalArgumentException("Unsupported: " + event.getClass());
    }
  }
  
  @Override public EventSeq fromJournal(Object event, String manifest) {
    if (event instanceof SeatBookedData) {
      SeatBookedData d = (SeatBookedData) event;
      return EventSeq.single(new SeatBooked(d.code, new Customer(d.customerName)));
    } else {
      throw new IllegalArgumentException("Unsupported: " + event.getClass());
    }
  }
}

The same technique could also be used directly in the Serializer if the end result of marshalling is bytes. Then the serializer can simply convert the bytes do the domain object by using the generated protobuf builders.

Store events as human-readable data model

Situation: You want to keep your persisted events in a human-readable format, for example JSON.

Solution: This is a special case of the Detach domain model from data model pattern, and thus requires some co-operation from the Journal implementation to achieve this.

An example of a Journal which may implement this pattern is MongoDB, however other databases such as PostgreSQL and Cassandra could also do it because of their built-in JSON capabilities.

In this aproach, the EventAdapter is used as the marshalling layer: it serializes the events to/from JSON. The journal plugin notices that the incoming event type is JSON (for example by performing a match on the incoming event) and stores the incoming object directly.

public class JsonDataModelAdapter implements EventAdapter {

  // use your favorite json library
  private final ExampleJsonMarshaller marshaller = 
    new ExampleJsonMarshaller();
  
  @Override public String manifest(Object event) {
    return "";
  }
  
  @Override public JsObject toJournal(Object event) {
    return marshaller.toJson(event);
  }

  @Override public EventSeq fromJournal(Object event, String manifest) {
    if (event instanceof JsObject) {
      JsObject json = (JsObject) event;
      return EventSeq.single(marshaller.fromJson(json));
    } else {
      throw new IllegalArgumentException(
        "Unable to fromJournal a non-JSON object! Was: " + event.getClass());
    }
  }
}

Alternative solution:

In fact, an AsyncWriteJournal implementation could natively decide to not use binary serialization at all, and always serialize the incoming messages as JSON - in which case the toJournal implementation of the EventAdapter would be an identity function, and the fromJournal would need to de-serialize messages from JSON.

Split large event into fine-grained events

Situation: While refactoring your domain events, you find that one of the events has become too large (coarse-grained) and needs to be split up into multiple fine-grained events.

Solution: Let us consider a situation where an event represents "user details changed". After some time we discover that this event is too coarse, and needs to be split into "user name changed" and "user address changed", because somehow users keep changing their usernames a lot and we'd like to keep this as a separate event.

The write side change is very simple, we simply persist UserNameChanged or UserAddressChanged depending on what the user actually intended to change (instead of the composite UserDetailsChanged that we had in version 1 of our model).

The EventAdapter splits the incoming event into smaller more fine grained events during recovery.

During recovery however, we now need to convert the old V1 model into the V2 representation of the change. Depending if the old event contains a name change, we either emit the UserNameChanged or we don't, and the address change is handled similarily:

interface V1 {};
interface V2 {}

// V1 event:
public class UserDetailsChanged implements V1 {
  public final String name;
  public final String address;
  public UserDetailsChanged(String name, String address) {
    this.name = name;
    this.address = address;
  }
}

// corresponding V2 events:
public class UserNameChanged implements V2 {
  public final String name;

  public UserNameChanged(String name) {
    this.name = name;
  }
}
public class UserAddressChanged implements V2 {
  public final String address;

  public UserAddressChanged(String address) {
    this.address = address;
  }
}

// event splitting adapter:
public class UserEventsAdapter implements EventAdapter {
  @Override public String manifest(Object event) {
    return "";
  }

  @Override public EventSeq fromJournal(Object event, String manifest) {
    if (event instanceof UserDetailsChanged) {
      UserDetailsChanged c = (UserDetailsChanged) event;
      if (c.name == null)
        return EventSeq.single(new UserAddressChanged(c.address));
      else if (c.address == null)
        return EventSeq.single(new UserNameChanged(c.name));
      else 
        return EventSeq.create(
            new UserNameChanged(c.name),
            new UserAddressChanged(c.address));
    } else {
      return EventSeq.single(event);
    }
  }

  @Override public Object toJournal(Object event) {
    return event;
  }
}

By returning an EventSeq from the event adapter, the recovered event can be converted to multiple events before being delivered to the persistent actor.