Advancing the realtime web
Slava Akhmechet January 27, 2015

Over the past few months the team at RethinkDB has been working on a project to make building modern, realtime apps dramatically easier. The upcoming features are the start of an exciting new database access model – instead of polling the database for changes, the developer can tell RethinkDB to continuously push updated query results to applications in realtime.

This work started as an innocuous feature to help developers integrate RethinkDB with other realtime systems. A few releases ago we shipped changefeeds – a way to subscribe to change notifications in the database. Whenever a document changes in a table, the server pushes a notification describing the change to subscribed clients. You can subscribe to changes on a table like this:

r.table('accounts').changes().run(conn)

Originally we intended this feature to help developers push data from RethinkDB to specialized data stores like ElasticSearch and message systems like RabbitMQ, but the release generated enormous excitement we didn’t expect. Digging deeper, we saw that many web developers used changefeeds as a solution to a much broader problem – how do you adapt the database to push realtime data to applications?

This turned out to be an important problem for so many developers that we expanded RethinkDB’s architecture to explicitly support realtime apps. The first batch of the new features will ship in a few days in the upcoming 1.16 release of RethinkDB, and I’m very excited to share what we’ve been working on in this post.

Why is building realtime apps so hard?

The query-response database access model works well on the web because it maps directly to HTTP’s request-response. However, modern marketplaces, streaming analytics apps, multiplayer games, and collaborative web and mobile apps require sending data directly to the client in realtime. For example, when a user changes the position of a button in a collaborative design app, the server has to notify other users that are simultaneously working on the same project. Web browsers support these use cases via WebSockets and long-lived HTTP connections, but adapting database systems to realtime needs still presents a huge engineering challenge.

A naive way to support live updates is to periodically poll the database for changes, but this solution is unworkable because it entails a tradeoff between the number of concurrent users and the polling interval. Even a small number of users polling the database will place a tremendous load on the database servers, requiring the administrator to increase the polling interval. In turn, high polling intervals very quickly result in an untenable user experience.

A scalable solution to this problem involves many cumbersome steps:

  • Hooking into replication logs of the database servers, or writing custom data invalidating logic for realtime UI components.
  • Adding messaging infrastructure (e.g. RabbitMQ) to your project.
  • Writing sophisticated routing logic to avoid broadcasting every message to every web server.
  • Reimplementing database functionality in the backend if your app requires realtime computation (e.g. realtime leaderboards).

All this requires enormous commitment of time and engineering resources. This tech presentation from Quora gives a good overview of how challenging it can be. The upcoming 1.16 release of RethinkDB is our take on helping developers build realtime apps with minimal effort, and includes the first batch of realtime push features to tackle this problem.

The database for the realtime web

A major design goal was to make the implementation non-invasive and simple to use. RethinkDB users can get started with the database by using a familiar request-response query paradigm. For example, if you’re generating a web page for a visual web design app, you can load the UI elements of a particular project like this:

> r.table('ui_elements').get_all(PROJECT_ID, index='projects').run(conn)
{ 'id': UI_ELEMENT_ID,
  'project_id': PROJECT_ID
  'type': 'button',
  'position': [100, 100],
  'size': [200, 100] }

But what if your design app is collaborative, and you want to show updates to all designers of a project in realtime? The 1.16 release of RethinkDB significantly expands the changes command to work on a much larger set of queries. The changes command lets you get the result of the query, but also asks the database to continue pushing updates to the web server as they happen in realtime, without the developer doing any additional work:

> r.table('ui_elements').get_all(PROJECT_ID, index='projects').changes().run(conn)
{ 'new_val':
  { 'id': UI_ELEMENT_ID,
    'project_id': PROJECT_ID
    'type': 'button',
    'position': [100, 100],
    'size': [200, 100] }
}

The first result of the query is just the value of the document. However, when the developer tacks on the changes command, RethinkDB will keep the cursor open, and push updates onto the cursor any time a relevant change occurs in the database. For example, if a different user moves the button in a project, the database will push a diff to every connected web server interested in the particular project, informing them of the change:

{ 'old_val':
  { 'id': UI_ELEMENT_ID,
    'project_id': PROJECT_ID
    'type': 'button',
    'position': [100, 100],
    'size': [200, 100] },
  'new_val':
  { 'id': UI_ELEMENT_ID,
    'project_id': PROJECT_ID
    'type': 'button',
    'position': [200, 200],  # the position has changed
    'size': [200, 100] }
}

Any time a web or a mobile client connects to your Python, Ruby, or Node.js application, you can create a realtime feed using the official RethinkDB drivers. The database will continuously push query result updates to your web server, which can forward the changes back to the client in realtime using WebSockets or one of the many wrapper libraries like SockJS, socket.io, or SignalR. Additionally, you’ll be able to access the functionality from most languages using one of the many community supported drivers.

The push access model eliminates the need for invalidation logic in the UI components, additional messaging infrastructure, complex routing logic on your servers, and custom code to reimplement aggregation and sorting in the application. The changes command works on a large subset of queries and is tightly integrated into RethinkDB’s architecture. For example, if you wanted to create an animated line graph of operation statistics for all tables in your production database, you could set up a feed on the internal statistics table to monitor the RethinkDB cluster itself:

> r.db('rethinkdb').table('stats').filter({ 'db': 'prod' }).changes().run(conn)

The architecture is designed to be scalable. We’re still running benchmarks, but you should be able to create thousands of concurrent changefeeds to scale your realtime apps, and the results will be pushed within milliseconds.

We’ve also built in many bells and whistles like latency awareness, that make building realtime apps much more convenient. For example, if the query results change too quickly and you don’t want to update the DOM more frequently than fifty milliseconds, you can tell changes to squash updates on a fifty millisecond window, and the database will take care of aggregating diffs and removing duplicates:

> r.table('ui_elements').get_all(PROJECT_ID, index='projects').changes(squash=0.05).run(conn)

Comparison with realtime sync services

There are many existing realtime sync services that significantly ease the pain of building realtime applications. Firebase, PubNub, and Pusher are notable examples, and there are many others. These services are excellent for getting up and running quickly. They let you sync documents across multiple browsers, offer sophisticated security models, and integrate with many existing web frameworks.

The upcoming features in RethinkDB are fundamentally different from realtime sync services in four critical ways.

Firstly, most existing realtime sync services offer very limited querying capabilities. You can query for a specific document and perhaps a range of documents, but you can’t express even simple queries that involve any computation. For example, sorting, advanced filtering, aggregation, joins, or subqueries are either limited or not available at all. This limitation turns out to be critical for real world applications, so most users end up using realtime sync services side by side with traditional database systems, and build up complex code to duplicate data between the two.

In contrast, RethinkDB is a general purpose database that allows you to easily express queries of arbitrary complexity. This eliminates the need for multiple pieces of infrastructure and additional code to duplicate data and keep it in sync across multiple services.

Secondly, the push functionality of realtime sync services is limited to single documents. You can sync documents across clients, but you can’t get a realtime incremental feed for more complex operations. In contrast, RethinkDB allows you to get a feed on queries, not just documents. For example, suppose you wanted to build a realtime leaderboard of top five gameplays in your game world. This requires sorting the gameplays by score in descending order, limiting the resultset to five top gameplays, and getting a continuous incremental feed that pushes updates to your clients any time the resultset changes. This functionality isn’t available in realtime sync services, but is trivial in RethinkDB:

r.table('gameplays').order_by(index=r.desc('score')).limit(5).changes().run(conn)

Any time the database gets updated with a new gameplay, this query will inform the developer which items dropped off the leaderboard, and which new gameplays should be included. Internally, the database doesn’t merely rerun the query any time there is a change to the gameplays table – the changefeeds are recomputed incrementally and efficiently.

Thirdly, realtime sync services are closed ecosystems that run in the cloud. While a hosted version of RethinkDB is available through our partners at Compose.io, both the protocol and the implementation are, and always will be, open-source.

Finally, most existing realtime sync services are built to allow access to their API directly from the web browser. This eliminates the need for building a backend in simple applications, and lets new users quickly deploy their apps with less hassle. As a general purpose database RethinkDB expects to be accessed from a backend server, and does not yet provide a sufficiently robust security model to be accessed directly from the web browser. We’re playing with the idea of building a secure proxy server to let web clients access RethinkDB directly from the browser, so eventually you might not need to write backend code if your application is simple enough. However, unlike realtime sync services, for now you have to access RethinkDB feeds through the backend code running in your web server.

Comparison with hooking into the replication log

Most traditional database systems offer access to their replication log, which allows clients to learn about the updates happening in the database in realtime. Many infrastructures for realtime apps are built on top of this functionality. There are three fundamental differences between RethinkDB’s changefeeds and hooking into the replication log of a database.

Firstly, like with realtime sync, hooking into the replication log gives you access to updates on individual documents. In contrast, RethinkDB’s changefeeds allow you to get feeds on query resultsets. Consider the example above, where we’re building a leaderboard of top five gameplays in a game world:

r.table('gameplays').order_by(index=r.desc('score')).limit(5).changes().run(conn)

To rebuild this functionality on top of a replication log your application would need to keep track of top five gameplays, and you’d have to write custom code to compare each new record in the gameplays table to decide if it replaces any of the gameplays in the leaderboard. More importantly, consider what happens if the game admin decides the player cheated and their gameplay score has to be reduced. Your code would have to go back to the database and recompute the query from scratch, because it has no information about which gameplay has the new record that should be on the leaderboard.

Writing this code is doable, but is fairly complex and error-prone. In a large application, the complexity can add up quickly if you have many realtime elements. In contrast, RethinkDB’s query engine eliminates this complexity by automatically taking care of the computation and sending you the correct updates as the resultset changes in realtime.

Secondly, as you move to sharded environments, working with a replication log presents additional complexity as there isn’t a single replication log to deal with. Your application would need to subscribe to multiple replication logs, and manually aggregate the events from replication logs for each shard. In contrast, RethinkDB automatically takes care of handling shards in the cluster, and changefeeds present unified views to your application.

Finally, most database systems don’t offer granular filtering functionality for replication logs, so your clients can’t get only the parts of the log they’re interested in. This presents non-trivial scalability challenges because your infrastructure has to deal with the firehose of all database events, and you need to write custom code to route only the relevant events to appropriate web servers. In contrast, RethinkDB handles scalability issues in the cluster, and each feed gives you exactly the information you need for a particular client.

RethinkDB’s changefeeds operate on a higher level of abstraction than traditional replication logs, which significantly reduces the amount of custom code and operational challenges the application developer has to consider.

Integrating with realtime web frameworks

One of the more notable projects that helps developers build realtime apps is Meteor. Meteor is an open-source platform for building realtime apps in JavaScript that promises a significantly improved developer experience. It handles a lot of the boilerplate necessary to build responsive interfaces with live updates, provides a complete platform with client-side and server-side components, and offers many advanced features like latency compensation and security out of the box. The team is making great strides in scalability and maturity of the platform, and many companies are starting to use Meteor to build the next generation of web applications.

Meteor is part of the Node.js ecosystem, and multiple other projects have popped up to bring its functionality to other languages. Volt is a framework that implements similar functionality in Ruby, and webalchemy is an alternative platform for Python. These projects are less mature, but have picked up a lot of interest in their respective ecosystems, and are likely to gain a lot of momentum once they accumulate enough functionality to let developers build high quality, scalable apps.

Meteor, Volt, and webalchemy frameworks run on top of databases, so they’re ultimately constrained by the realtime functionality and scalability of existing database systems. We’ve been collaborating with the Meteor team to ensure our design will work well with these and other similar projects. A few community members have been working on a RethinkDB integration with Meteor and Volt, and we expect robust integrations to become available in the coming months.

More work ahead

The upcoming 1.16 release contains only a subset of the functionality we’d like to include. In the next few releases we plan to expand realtime push even further:

  • We’re discussing the implementation for restartable feeds here and here. Feedback welcome!
  • We’d like to make more complex queries available via realtime push. In particular, efficient realtime push implementations for the eq_join command and map/reduce are fairly complex, and aren’t making it into 1.16.
  • Exposing the database to the internet entails serious security concerns, so we’re kicking around ideas for a secure proxy to enable direct browser access of realtime feeds.

This work is guided by three high level design principles:

  • We believe it’s important for realtime database infrastructure to be open. Both the protocol and the implementation are, and always will be, open-source.
  • The implementation should be non-invasive and very simple to use. Developers shouldn’t have to care about realtime features until they’re ready to add the functionality to their apps.
  • Realtime functionality should be efficient, scalable, and tightly integrated with the rest of the database. It shouldn’t feel like an afterthought.

Advancing the realtime web

The new functionality is a start of an exciting new database access model that eliminates many complex steps necessary for building realtime apps today. There is no need to poll the database for changes or introduce additional infrastructure like RabbitMQ. RethinkDB pushes relevant changes to the web server the instant they occur. The amount of additional code the developer has to write to implement realtime functionality in their apps is minimal, and all scalability issues are handled by the RethinkDB cluster.

We’ll be releasing the realtime extensions to RethinkDB in the next few days along with tutorials and documentation. In the meantime, you can watch the video with a live demo of the features:

We’re hoping RethinkDB 1.16 will make building realtime apps dramatically simpler and more accessible. Stay tuned for more updates, and please share your feedback with the RethinkDB team!

Found a typo? Help us improve this page.