Backbone.js, now realtime with Pusher
blog.pusher.com
blog.pusher.com
Pushing complexity out to specialized platforms is a popular trend these days. PaaS vendors are popping up in every area, but there's a threshold where this type of outsourcing makes sense.
* How complex/specialized is the problem?
* How close is it to our core product/competencies?
* What are the comparative costs in development and maintenance vs using the service?
These are the core questions I ask when I'm evaluating PaaS vs in-house solutions.
Q: How complex/specialized is the problem?
A: Unlike HTTP, WebSockets aren't widely supported, so real-time browser communication is a complex problem. However, it is a generalized issue. It is not a problem that we will have to solve alone. There are probably libraries available to assist in the implementation of our own "push" stack.
Q: How close is it to our core product/competencies?
A: A core feature of our product is real-time exchange of pricing data between clients (browsers) and the server. The faster clients receive feedback on their pricing, the better our product works. This means real-time browser communication is central to our product. It is not core, but it is very, very close.
Q: What are the comparative costs in development and maintenance vs using the service?
A: An initial review shows that there are popular libraries that abstract a significant portion of the complexity away, including fail-over for browsers that don't have support for WebSockets. Availability of commercial support isn't clear, but source code is available. We will have to build our own API, but we have the benefit of tailoring it to our application, rather than using a generic API. Back-end message queuing services will need to be added to our monitoring suite.
In a more general sense, we also evaluate how the use of PaaS would affect our product in the event of downtime. Every PaaS you use in your application spreads your exposure out over their systems as well. Can you afford for "push" to go away? What will a fail-over system cost you? We opted to make real-time communications part of our app, so the redundancy we built in to our app hosting also covers our real-time services.
For my money, I just can't see outsourcing something like client-server "push" communication.
When people ask "Why would I use Pusher instead of running my own realtime services in-house?", they often overlook the other complexities involved in the problem. Eg, scaling, maintaining, ensuring uptime, etc.
Looking at any service (IaaS, PaaS, SaaS) from the outside is an iceberg problem. You can see the public API (REST based API + WebSocket API), and you know what it does for you, however, what you don't see is what lies beneath the water, the other factors and things that have been thought about for you so you don't need to think about them.
The thing that really needs to be thought about is this: by building your own in-house realtime solution, you'll probably face similar problems to those that we have to face building and maintaining Pusher, such issues might include: 1a) How will you communicate with your realtime service? Will it be a REST API (like Pusher) or something else? 1b) How will you maintain the infrastructure required for the persistent connections and consistent distribution of messages throughout your cluster? 1c) How will you handle sudden (unexpected) growth in usage, and do you want to have to worry about that? 1d) Who will ensure that you're up-to-date on the latest websocket specification (we're at draft-07 being released in firefox 6, and draft-09 is currently being edited).
I should note at this stage, that hosting websockets is in no way similar to hosting standard HTTP, there's a lot more to the problem.
By the same logic, you could apply that to any PaaS or SaaS or IaaS product. Do you use Amazon Web Services or another cloud based host, or do you use a non-cloud based host? Thats a place where you're outsourcing the management and maintenance of your infrastructure; that is, unless of course you maintain your own data-centers and install all your own servers?
Having built a websocket server for Node.js, I know the issues involved with maintaining and keeping up to date with the latest protocol specification. Is it cheaper for you to use Socket.io instead of role your own solution, because honestly, something might go terribly wrong within Socket.io. Do you want to spend time maintaining that?
This is all about choices, for some projects, it may make sense to handle everything where you need absolute control over every aspect of your application and infrastructure. However, from experience, I can say that I probably wouldn't want to be having to maintain absolutely everything in my application. If AWS/EC2 goes down, then yes, I should worry, however, I think EC2 is less likely to go down than my own hand-installed network infrastructure.
I'm not going to try to sell you on Pusher, however, just take into full account absolutely everything before you decide to roll your own realtime solution.
– Micheil.
You make a good case for the complexity, which is one of the factors I consider. One of my unspoken concerns is that an open source library like Socket.io is going to get stale. I keep an eye on the commit log, and it looks like development is pretty active, but a look at the impact graph shows that there was a lull, and it looks like development might be changing hands. This is par for the course with open source projects, but I question whether "par for the course" is best for our project.
My other big concern is running "yet another server". We're currently running a very standard Rails stack on VPS infrastructure. We're not on a full-blown "cloud" provider, because I don't trust them. Heroku has had too much downtime for our company. AWS offers multiple AZ redundancy, but there's an inherent complexity there that we're not ready to tackle.
Unlike most startups, our revenue generating traffic isn't spread out over thousands of visits. Our revenue is made during a purchasing event that runs for a half hour. Some days, there might only be a single event. If our infrastructure is down for that event, we stand to lose $25k-$50k. I know we're not the only one with this problem (some would lose millions over the same time period), but we're a bootstrapped company. A major outage like the ones Heroku has experienced could put a serious hole in our cashflows.
Because of this, we run a very straight forward Rails hosting stack (Apache/Passenger/MySQL) on a straight forward Xen VPS host at data centers in Dallas and NYC. I fully expect that at some point in the future, one of these data centers is going to disappear off the net and we'll have to be ready. If our entire hosting provider disappeared, I could bring up every one of our customer app instances in a half-day. I know this because I've done the drills using local VPS builds from scratch, and we use third-party, off-net backup locations.
Any XaaS product we integrate in to our core product would have to demonstrate a similar focus to uptime and redundancy. The recent AWS outage that took several high-profile websites offline is a good example. Lot's of seemingly smart companies hadn't deployed across multiple AZs because of the complexity. I can't partner with a company that makes those kinds of decisions.
I know you don't want to try to sell me Pusher, but I'm genuinely curious about Pusher's commitment to uptime. I'd love to see a blog post, or knowledge base article that talks about Pusher's redundancy and disaster recovery plans. I'd love to "just trust" that XaaS companies have this sort of thing covered, but history has taught me otherwise.
Thanks again. I really appreciate the response.