Event sourcing backed by Azure Blob Storage, in a simple, low-maintenance .NET client library. You can use Lokad.AzureEventStore for your prototypes and small-scale projects, then move on to other solutions (such as Apache Kafka or Greg Young's Event Store) once you confirm the need for additional performance or features.

If you're curious about event sourcing, you can read this introduction from Microsoft, or the documentation for Event Store.

Lokad.AzureEventStore will fit your project comfortably if:

• Your events are serialized with Json.NET, and weigh less than 512KB each.
• Each aggregate receives less than 10 events per second.
• The materialized views for all your aggregates fit in memory.

NuGet package: Lokad.AzureEventStore.

What you get out of Lokad.AzureEventStore:

• Robust data storage: your events are written to Azure Append Blobs, which are append-only. This makes it impossible for your application to overwrite, delete or alter the existing stream of events in any way.

  The serialization format of the event stream includes checksums to detect data corruption (those are checked on every read), and is simple enough that creating a backup of a stream can be done by copying the blobs byte-for-byte.

  This library has been used as the primary persistence model at Lokad since 2014, and we have yet to encounter a case of data loss.

• No need for a server: you do not need to install a database server, keep it up-to-date with security fixes, monitor its performance, or plan a fail-over solution.

• Thread-safety within the application: multiple threads can append events to the stream, or read the current state of a projection. The library also reads events from the remote stream and updates the projection state in the background.

  All operations which involve remote access to the Append Blobs use are asynchronous (using System.Threading.Tasks).

• Consistency within each aggregate: when deciding to appending a new event to a stream, the application has access to all events present in the stream before that new event.

  For example, if two applications want to append two mutually exclusive events at the same time, one of the two applications is guaranteed to see the event generated by the other application before it attempts to write its own. This is handled transparently by the library.

• Projection state snapshots: as long as your application provides serialization methods for your projection state, the library will automatically load a previous snapshot of the state whenever available, and only read back the events that were appended since this snapshot was taken.

  You are in control of when (and how often) those snapshots are created.

• Easy migration path: the event stream is just a stream of events, serialized as JSON. It can be easily moved to another event sourcing solution.

  Projection code and event contracts are not dependent on the library and can be reused as-is.

The example project can serve as a quick introduction to the concepts behind Lokad.AzureEventStore. The project is a distributed word counter: each instance is a console application, and all instances share a dictionary of words with their corresponding counts.

If text is typed on the command line, the application splits the text into words and updates the shared count dictionary.

If the line starts with -- then the words on the line are instead removed from the dictionary.

An event represents "something that has happened", as part of an append-only stream of events representing the full history of the application. In C#, these are represented as JSON-serializable plain old objects implementing a single interface local to the project (such as IEvent in the example project).

We use two event types for this example: ValueUpdated happens whenever a word is counted, it contains the word (Key) and the new tally for that word (Value) ; ValueDeleted happens when the word is removed from the dictionary, and only references the word (Key).

    [DataContract]
    public sealed class ValueUpdated : IEvent
    {
        [DataMember]
        public string Key { get; private set; }

        [DataMember]
        public int Value { get; private set; }
    }

    [DataContract]
    public sealed class ValueDeleted : IEvent
    {
        [DataMember]
        public string Key { get; private set; }
    }

As expected of an event sourcing solution, the primary way of accessing the state application is not to read back through the event stream (which would become very time-consuming as the stream grows) but instead to create one or more materialized views of the state. Keeping these views up-to-date can be done efficiently by applying new events as they appear.

Lokad.AzureEventStore imposes the additional constraint that state should be represented by immutable data structures. This guarantees that multiple threads can read the state while a background thread applies new events, without any locks being required.

In the example project, the state is simply a dictionary that maps words to their current tally.

    public sealed class State
    {
        public State(ImmutableDictionary<string, int> bindings)
        {
            Bindings = bindings;
        }

        public ImmutableDictionary<string,int> Bindings { get; }
    }

In Lokad.AzureEventStore, a projection is the code that turns the stream of events into a state object representing the materialized view. A projection that creates a State from a stream of IEvent should implement the interface IProjection<IEvent, State>, which itself implements the interface IProjection<IEvent> of projections that work with events of type IEvent (since you may have several of these projections, with different state types, in a single application).

The following fields must be implemented:

• The initial state, to which the first event in the stream will be applied.

      public State Initial => 
          new State(ImmutableDictionary<string, int>.Empty);
  

• The type of the state. This is used for reflection purposes, to identify which projection should be invoked for a given materialized view.

      public Type State => typeof(State);
  

• The Apply function which, based on a previous state and a new event, returns a new state. The sequence number uniquely identifies the event, being incremented by 1 for each event added to the stream (the first event has a sequence number of 1).

  Note that the projection should never rely on events being provided in any specific order. The library will routinely call the projection with the same event more than once, or with events that do not even appear in the event stream, or skip certain events entirely.

  It is however guaranteed that the State passed to function Apply on a given sequence number will be the result of calling Apply on the previous sequence number (that is, the State is build by applying events in chronological order, starting with Initial).

  In a real-life project, this function would be implemented with a visitor pattern, but in the example project it is simply:

     public State Apply(uint sequence, IEvent e, State previous)
     {
         if (e is ValueDeleted d)
             return new State(previous.Bindings.Remove(d.Key));
   		  
         if (e is ValueUpdated u)
             return new State(previous.Bindings.SetItem(u.Key, u.Value));
  
         throw new ArgumentOutOfRangeException(nameof(e), "Unknown event type " + e.GetType());
     }
  

• The TryLoadAsync and TrySaveAsync, used to add checkpoint support to a projection. TrySaveAsync should return false if checkpoints are not supported, TryLoadAsync will only be called if a checkpoint was previously generated (but should throw if the checkpoint cannot be parsed).

• The FullName is used for logging, but also for checkpoint versioning if checkpoints are supported. Permitted characters are letters, numbers and the character -. The canonical name for version 18 of projection Foo is Foo-18.

  You should increase the version number whenever changes were applied to the projection that would alter the behaviour of existing events, so that the library does not mistakenly load an incorrect checkpoint but instead replays the stream from the beginning.

The storage configuration describes the location of an event stream, which is then used by the application to connect to the stream either to read or to read and write.

To create a connection string:

1. Use the new StorageConfiguration("...") constructor and pass an Azure Storage connection string. This can be the typical read-write connection string (starts with DefaultEndpointsProtocol=...) or a connection string with a Shared Access Signature that permits only listing and reading blobs (starts with BlobEndpoint=...).

   By default, Lokad.AzureEventStore will store the event stream in the root container of the storage account. To override this, you can append ;Container=... at the end of the connection string to specify another container.

2. Use the new StorageConfiguration("...") constructor and pass an absolute or relative path on your disk, such as C:\MyStream or .\MyStream. This method is not acceptable for production use ! It is provided so that applications can be run locally on a development machine without internet access (and without the Azure Storage emulator), and does not support concurrent access by multiple applications.

3. Use Testing.Initialize() to create a non-persistent in-memory stream of events (either empty, or containing a specific list of initial events).

   This function is intended for unit tests, to quickly set up the initial state of the stream.

   You can use Testing.GetEvents() later on to retrieve all events in such a stream, again for unit testing purposes.

To represent an event stream and associated materialized views, the library provides the EventStreamService<IEvent, State> class.

    var svc = EventStreamService<IEvent, State>.StartNew(        
        storage: new StorageConfiguration("..."),
        projections: new[] { new Projection() },
        projectionCache: null,
        log: null,
        cancel: cancellationToken);

When created, the service connects to the stream and starts retrieving events, building the materialized views on-the fly. If a checkpoint is found and can be loaded, then the service will only read events that were generated after the creation of the checkpoint. This phase is called the catch-up, and ends once the service reaches the end of the stream. During catch-up:

• Property service.IsReady returns false.
• Task service.Ready is running.
• Attempting to access the state or to append new events throws a StreamNotReadyException.

Once the catch-up is complete, you can access the materialized views in two ways:

• service.LocalState immediately returns the last version of the state. This will not take into account any events appended to the stream by other applications since the last time the service interacted with the stream. It will, however, be very fast.
• service.CurrentState() asynchronously returns the current version of the state, including any events written to the stream up until the moment the call was made. It performs a round-trip to the Azure Blob Storage, which might take a few milliseconds.

Call service.AppendEventsAsync(builder, cancellationToken). The builder is a function Func<State, Append<IEvent>> which takes the current state and returns the list of events to be appended.

If the builder returns at least one event, the service will then attempt to append to the event stream. If appending fails due to a conflict (because another application appended events and rendered the state obsolete), the service will fetch any new events, update the local state, and call the builder again. This repeats until the builder throws, returns no events, or the appending succeeds.

Function service.Use(params IEvent[] events) is a good helper function for creating an Append<IEvent>:

    // Another occurence of 'word' has been found
	await service.AppendEventsAsync(state =>
	{
		state.Bindings.TryGetValue(word, out int currentCount);
		return service.Use(new ValueUpdated(word, currentCount + 1));
	}, CancellationToken.None)

In practice, it is often useful to return some state from the builder back to the calling context. For instance, if the builder creates a new item in the state, it should be able to return the identifier of the new item. This is possible by using an Append<IEvent, TResult> instead, which carries data out of the builder:

    // Another occurence of 'word' has been found
	var result = await service.AppendEventsAsync(state =>
	{
		state.Bindings.TryGetValue(word, out int currentCount);
		return service.With(
		    currentCount + 1, // Return the new count
		    new ValueUpdated(word, currentCount + 1));
	}, CancellationToken.None);

	var newCount = result.Result;

The application should not be allowed to modify past events. However, as a matter of system administration, it is possible to change the event stream. This should always be a manually triggered operation, and will include the following steps:

1. Develop the migration concierge, an executable which reads from the current stream (using EventStream<T>) and writes to a new stream (using MigrationStream<T>). This will take anywhere from a few minutes to several days.
2. Start the migration concierge, wait for it to catch up to the current stream. This will usually take tens of minutes.
3. Turn off the application (or re-configure it to be read-only).
4. Wait for the migration concierge to process all events that were added to the stream before the application stopped writing. This should take a few seconds.
5. Stop the migration concierge.
6. Re-configure the application to use the new stream instead of the current stream, and start it.

You can improve the restart time by pre-generating a state cache based on the new stream (right after step 2) and uploading it so that the application can use it to start up more easily.

Example of a migration concierge which drops all events of a defunct type ObsoleteEvent:

    var currentStream = new EventStream<IMyEvent>(currentConfig);
    var newStream = new MigrationStream<IMyEvent>(newConfig);

    await newStream.MigrateFromAsync(
        currentStream, 
        // Drop events of type 'ObsoletEvent'
        (IMyEvent ev, uint _seq) => ev is ObseleteEvent ? null : ev, 
        // When reaching the end of 'currentStream', refresh it every 2 seconds
        // to see if new events have appeared
        TimeSpan.FromSeconds(2),
        default(CancellationToken));

This project is split into multiple layers. Starting at the top level (the highest level of abstraction) and moving down:

• The EventStreamService (or ESS) acts as a helpful wrapper around a permanently connected stream. On the write side, it provides the "append event to stream" primitive. On the read side, it keeps the current state up-to-date with the latest events in the stream. All its operations are thread-safe. Finally, combining both read and write access in the same object allows the "try commit, if failed update" pattern for transactions.

• The EventStreamWrapper (or ESW) does the heavy lifting for the ESS, in a single-threaded manner (the ESS just adds multi-thread support).

• The ReifiedProjection keeps the state of a projection up-to-date when it is passed events, and also supports saving and loading of the projection state to a persistent cache. The ESW will keep one or more ReifiedProjection instances to keep track of its state (or sub-components of its state).

• The EventStream provides read-write access to the events in the stream. It is optimized for sequential, beginning-to-end reads. The ESW uses the EventStream directly when pushing events, and reads the events to update the state. It is the EventStream that generates the uint sequence numbers for new events.

• The IStorageDriver provides low-level access to persistent storage optimized for concurrent appends (several drivers are available, though only one for production needs). It stores byte[] data along with uint sequence numbers, and the EventStream is responsible for converting events to and from the byte[] representation.