This project is fully functional, but only supports a single event stream. I've taken the lessons learnt with Rewind and applied the knowledge to the next generation of Rewind, Gorewind. The project can be found here: https://github.com/JensRantil/gorewind
Have you ever been nervous of all those DBMSs schema changes when you are deploying your applications? They are gonna take too long, or break backward compatibility? Have you ever thought "Crap, I wish I had stored that information since earlier"? Have you ever felt your writing patterns and your reading patterns differ a lot, making things harder to scale?
CQRS (Command-Query Response Segregation) is an architectural pattern that aims to solve these issues by splitting up your architectural system into two parts:
- A write side that takes care of validating input and optimizes for fast writes. The write side takes commands and outputs corresponding events if the command validates correctly.
- A read side that listens to incoming events from the write side. The read side is optimized for fast reads.
A core concept in CQRS is the event store which sits inbetween the write and the read side. The event store takes care of three things:
- persisting all events to disk.
- being a hub/broker replicating all events from the write to the read side of things.
- it allows fast querying of events so that different parts of the system can be synced back on track and new components can be brought back in play.
Rewind is an event store application that talks ZeroMQ. It is written in Python and supports multiple storage backends.
Rewind exists on PyPi and can be downloaded by issuing:
$ pip install rewind
Rewind uses basic setuptools. Installation can be used done as follows:
$ git clone https://github.com/JensRantil/rewind.git $ cd rewind $ python setup.py install
However NOTE, that this will install Rewind globally in your Python environment and is NOT recommended. Please refer to virtualenv on how to create a virtual environment.
There is a preexisting Python client package, rewind-client. If you are not writing a new client you might want to skip the following section below.
Rewind has two different wire protocols. Each is using ZeroMQ as low level transport protocol. Each wire protocol has one single ZeroMQ endpoint in Rewind:
- A request/response socket for Rewind. It is used for publishing new events and querying chronological slices of all events throughout time.
- A streaming socket. It is used by all clients that are interested in all new incoming events.
Each endpoint is configurable through command line when starting rewind. Issue rewind --help to get a list of the specific command line arguments rewind can handle.
Note that the wire protocol is still under heavy development. Pull requests and improvement proposals are welcome!
The socket for querying Rewind is the one which has the most advanced wire protocol. The socket is of type RES and takes commands. A typical converation between a client (C) and Rewind (R) looks like this:
C: Request R: Response C: Request R: Response ...
Each request is a multipart message. The first part is a string that specifies the type of request. There are multiple request types:
Used for publishing an event. The next message part is a blob of bytes that is supposed to be a serialized event of some form. Rewind does not know anything about the serialization format. It always simply stores the bytes.
Each new incoming/published event triggers that it is to be streamed out to all listening clients.
On successful reception of an event, Rewind responds with the ASCII bytes PUBLISHED. See "Error response" below for error response.
Used for querying for older events. For the QUERY request type the next two message parts must be:
- Contains an optional event id, or an empty part. Restricts the earliest (chronologically) incoming message that we are interested in to all messages received after the event with the specified event id. Note that this does not include the message with the specified event id. If this part of the message is empty, no lower restriction is made and messages will be returned starting from the first event ever seen.
- Contains an optional event id, or an empty part. Restricts the latest (chronologically) incoming message that we are interested in to all messages received before, or including, the event with the specified event id. If this part of the message is empty, no upper restriction is made and messages will be returned starting from the first event ever seen.
If you are a data structure type-of-guy you could view Rewind as a distributed insert-ordered map (event id => event) that allows querying of ranges of events based on event ids.
There are two types of responses that can be given upon a query:
- An error. See "Error response" below; or
- a resultset containing events. It's a multipart message containing frames like so; eventid #1, event #1, eventid #2, event #2, eventid #3, event #3, ... where eventid #X is the event id for event X. At most 100 messages will be returned. If Rewind did not cap number of events, the result will be appended by a last frame containing the ASCII content END. It is up to the client to make requests repeatedly if the result set is capped.
If anything goes wrong, a single message starting with the ASCII text ERROR will be sent with the response. This means an error occured. The rest of message contains a human readable (ASCII) description of the actual error that occured. This information can be highly useful for remote clients to debug any problems that might arise.
Every incoming event gets broadcast to all sockets connected to the streaming socket. The streaming socket a ZeroMQ socket of type PUB.
Every message received automatically gets assigned a unique event id (UUID, type 1) by Rewind. This event id is used for querying events (see below). Each sent message from the streaming is a multipart message that consists of two parts:
- The event ID. The client should view this as a series of bytes.
- The previous event ID. This information is useful to know whether ZeroMQ high-water mark kicked in while syncing up a client while querying for older events. If streaming has just begun, this message part can be empty and can thus be ignored.
- The event content. This is the exact same bytes that were correspondingly sent to the receiving socket.
Getting started developing rewind is quite straightforward. The library uses setuptools and standard Python project layout for tests etcetera.
This is how you check out the rewind library into a virtual environment:
cd <your development directory> virtualenv --no-site-packages rewind cd rewind git clone http://<rewind GIT URL> src
Every time you want to work on rewind you want to change directory into the source folder and activate the virtual environment scope (so that you don't touch the global Python environment):
cd src source ../bin/activate
The first time you've checked the project out, you want to initialize development mode:
python setup.py develop
Running the test suite is done by issuing:
python setup.py nosetests
. Nose is configured to automagically spit out test coverage information after the whole test suite has been executed.
As always, try to run the test suite before starting to mess with the code. That way you know nothing was broken beforehand.
Generally, I try to keep a 100% code coverage of the rewind.server package. Due to some Python 3 support hack, the coverage is around 99%. for Python 2.
The Rewind central github repository also has Travis CI integration that can be accessed at http://travis-ci.org/#!/JensRantil/rewind Every time a pull request is being made to https://github.com/JensRantil/rewind, Travis CI will make a comment about whether the pull request breaks the test suite or not.
Spelling mistakes, bad grammar, new storage backends, wire format improvements, test improvements and other feature additions are all welcome. Please issue pull requests or create an issue if you'd like to discuss it on Github.
Pick and choose:
- Rewind can look at what happened in the past and replay the events since then.
- It's time to rewind and rethink the way we are overusing DBMS's and the way we are storing our data.
This package has been developed by Jens Rantil <firstname.lastname@example.org>. You can also reach me through snailmail at:
Jens Rantil Lilla Södergatan 6A 22353 Lund SWEDEN