Eric Florenzano’s Blog

The Technology Behind Convore

Feb 16, 2011

We launched Convore last week, and the first question developers tend to ask when they find Convore is "what technology powers this site?" It is asked so often, in fact, that we have started to copy and paste the same short response again and again. That response was good enough to satisfy people who simply wanted to know if we were Rails or Django, or whether we were using node.js for the real-time stuff, but this article will expand upon that-- not only giving more details for the curious, but also giving us a link to point people at when they ask the question in the future. I always wish other people were totally open about their architectures, so that I can learn from their good choices and their bad, so I'd like to be as open as possible about ours. Let's dive in!

The basics

All of our application code is powered by Python. Our front-end html page generation is done by Django, which we use in a surprisingly traditional way given the real-time nature of Convore as a product. Everything is assembled at once: all messages, the sidebar, and the header are all rendered on the server instead of being pulled in after-the-fact with JavaScript. All of the important data is canonically stored in PostgreSQL, including messages, topics, groups, unread counts, and user profiles. Search functionality is provided by Solr, which is interfaced into our application by way of the handy Haystack Django application.

The message lifecycle

When a new message comes into the system, first it's parsed by a series of regular expressions designed to pull out interesting bits of information from the message. Right now all we're looking for is username references and links (and further, whether those links point at images which should be rendered in-line.) At the end of this parsing stage, we have a structured message parse list, which is converted into JSON.

So, for example if someone posted the message:

@ericflo @simonw Here's how we connect/disconnect from Redis in production:

The resulting JSON parse list would look like this:

        "type": "username",
        "user_id": 1,
        "username": "ericflo",
        "markup": "<a href=\"/users/ericflo/\">@ericflo</a>"
        "type": "username",
        "user_id": 56,
        "username": "simonw",
        "markup": " <a href=\"/users/simonw/\">@simonw</a>"
        "type": "text",
        "markup": " Here&#39;s how we connect/disconnect from Redis in production: "
        "type": "url",
        "url": "",
        "markup": "<a href=\"\" target=\"_blank\"></a>"

After this is constructed, we log all our available information about this message, and then save to the database-- both the raw message as it was received, and the JSON-encoded parsed node list.

Now a task is sent to Celery (by way of Redis) notifying it that this new message has been received. This Celery task now increments the unread count for everyone who has access to the topic that the message was posted in, and then it publishes to a Redis pub/sub for the group that the message was posted to. Finally, the task scans through the message, looking for any users that were mentioned in the message, and writes entries to the database for every mention.

On the other end of that pub/sub are the many open http requests that our users have initiated, which are waiting for any new messages or information. Those all simultaneously return the new message information, at which point they reconnect again, waiting for the next message to arrive.

The real-time endpoint

Our live updates endpoint is actually a very simple and lightweight pure-WSGI Python application, hosted using Eventlet. It spawns off a coroutine for each request, and in that coroutine, it looks up all the groups that a user is a member of, and then opens a connection to Redis subscribing to all of those channels. Each of these Eventlet-hosted Python applications has the ability to host hundreds-to-thousands of open connections, and we run several instances on each of our front-end machines. It has a few more responsibilities, like marking a topic as read before it returns a response, but the most important thing is to be a bridge between the user and Redis pub/sub.

Future improvements

There are so many places where our architecture can be improved. This is our first version, and now that real users are using the system, already some of our initial assumptions are being challenged. For instance, we thought that pub/sub to a channel per group would be enough, but what that means is that everyone in a group sees the exact same events as everyone else in that group.

This means we don't have the ability to customize each user's experience based on their preferences--no way to put a user on ignore, filter certain messages, etc. It also means that we aren't able to sync up a user's experience across tabs or browsers, since we don't really want to broadcast to everyone in the group that one user has visited a topic, thereby removing any unread messages in that topic. So going forward we're going to have to break up that per-group pub/sub into per-user pub/sub.

Another area that could be improved is our unread counts. Right now they're stored as rows in our PostgreSQL database, which makes it extremely easy to batch update them and do aggregate queries on them, but the number of these rows is increasing rapidly, and without some kind of sharding scheme, it will at some point become more difficult to work with such a large amount of rows. My feeling is that this will eventually need to be moved into a non-relational data store, and we'll need to write a service layer in front of it to deal with pre-aggregating and distributing updates, but nothing is set in stone just yet.

Finally, Python may not be the best language for this real-time endpoint. Eventlet is a fantastic Python library and it allowed us to build something extremely fast that has scaled to several thousand concurrent connections without breaking a sweat on launch day, but it has its limits. There is a large body of work out there on handling a large number of open connections, using Java's NIO framework, Erlang's mochiweb, or node.js.

That's all folks

We're pretty proud of what we've built in a very short time, and we're glad it has held up as well as it has on our launch day and afterwards. We're excited about the problems we're now being faced with, both scaling the technology, and scaling the product. I hope this article has quenched any curiosity out there about how Convore works. If there are any questions, feel free to join Convore and ask away!

(Or discuss it on Hacker News)