Back Off, Bub! A Short Guide to Robust Client/Server Connection Architecture
Like this article? We recommend
Like this article? We recommend
Like this article? We recommend
Many applications these days involve some kind of client/server interaction over a network. From the naïve example of a web app making a call to a server for data (for example, in Gmail) to a more complex open-ended connection of streaming a video to your phone, all of these interactions involve making connections to services and performing some higher-level function(s) once that connection is open.
It's no surprise that these activities all share certain common patterns in their construction. Even services that simply operate within a closed environment—such as a system of servers processing log data or clearing checks—share these characteristics. For convenience, we'll call these "client/server architectures." Of course, it's not necessary that one party always be a client of the other, but for the purposes of making a connection and requesting or transferring some data, we can always model one end as a client (the party initiating the connection) and the other as a server (the party listening for and then accepting the connection).
If networks and systems were perfect, we'd just spin up a server and its client, and as soon as they started, they would connect and everything would be hunky-dory. In fact, I daresay that this is how most systems tend to be designed at the beginning—and, sadly, some even later in life.
Living in this imperfect world, we naturally must account for any number of problems with this whole scenario:
- The connection breaks due to a transient network or hardware problem.
- The server is unavailable because it's being restarted.
- The client loses network connectivity.
Any number of similar problems could occur. In the event of a lost connection, many mobile and web apps simply give up or fall into an unusable state, leaving the user to resort to closing and opening the app or web page again. This approach works to an extent, but it's not a particularly elegant or friendly solution. For automated services, it's also impractical to require human intervention for a restart. In some cases, it might even be impossible; imagine an underwater submarine or a satellite in space sending image data to a server on the surface.
On Reconnection and Naïveté
As an immediate solution, the simple answer is to have some facility in the client that detects when a connection has dropped and attempts to reconnect. While the connection is down, you can display a notice to the user, or simply buffer all network requests in a queue, to be sent out later when the connection is available again. Nice and easy.
On the surface, this solution cures all our ills, and normal service appears to have resumed. However, this design introduces a slew of complex new problems that are a lot more painful than simply a broken connection. Take this scenario, for example: The server is maxed out on resources and disconnects the client in order to make some room for the remaining set of connected clients. The rather naïve idea is that dropping a client here and there will alleviate some pressure. An even simpler scenario is that the client supplies the wrong credentials (for example, an incorrect password), and the server refuses to honor the connection.
In these cases, the client gets disconnected, notices that the connection is broken, and immediately attempts to reconnect. Today's devices and browsers can make several hundred such attempts per minute. Server systems attempting a reconnection can number in the thousands. So many connection attempts, even from only a handful of clients, will bring any server to its knees.
If you have hundreds of clients like this, your server has no hope of ever spinning up for normal service. This is the equivalent of a distributed denial of service (DDoS) attack—exactly the technique that malicious attackers use to bring down even large service clusters.
Another problem is that making so many repeated attempts exhausts client resources as well, and if it's a low-resource system (such as a smartphone), you could be wasting battery energy or causing stress to other services running on the system, all of which create an extremely unpleasant experience for your users. Even on an automated client, other services might be starved because of a disconnected client stuck in such a loop.
So what can we do instead? Should the server refuse frequent connection attempts unilaterally? Or is it the client's responsibility to do something more clever?
The Solution to Going Forward: Backing Off
The answer is both, really. A server should refuse frequent requests from the same client. This is known as rate-limiting, and it ought to use a predetermined number. We won't get into the details of that possibility here, except to say that rate-limiting ensures that all clients get a fair chance, rather than allowing a few greedy ones to hog all the resources.
For its turn, the client should do something similar—when it sees frequent error state responses from the server (the most fundamental of which is a disconnection), the client should reduce the rate at which it makes requests of the server. This is known as backing off, and it's implemented via the use of several types of algorithms, depending on capacity and requirements.
The simplest back-off algorithm just places a delay before the next connection attempt is made:
public void onDisconnect(State code) {
if (code == State.ERROR)
sleep(1, TimeUnit.SECONDS);
connect();
}
Assuming that the function onDisconnect() is called every time the connection drops, we now have a delay of one second before a subsequent connection is attempted. This is a good solution, but not great. A delay of one second is a vast improvement over a go-as-fast-as-you-can loop, but it's probably still too fast and too naïve. For example, maybe the server simply isn't available; a scheduled outage may have taken it offline for a good 30 minutes. Or, if the client has lost network connectivity, there's really no point in wasting resources making a connection attempt every second into a vacuum.
Suppose we increase this delay to once every 10 minutes? It's a decent gap, and not so long that we might miss a quick recovery. That takes care of the outage and lost-connectivity scenarios, but it simply doesn't handle transient network failures. This whole class of errors is becoming increasingly common: packets lost in a weak WiFi signal, network routes changing underfoot, load-balancing errors, or any number of such invisible gremlins. It's really unacceptable to make a user wait 10 minutes when an immediate reconnection, or an attempt after a few seconds, would have succeeded perfectly.
The trick here is to use a non-linear back-off function. So far we've seen only linear back-off functions; in other words, functions that yield a fixed delay per time interval (one second, 10 minutes, and so on). Here's an example of a quadratic back-off function:
int attempt = 0;
public void onDisconnect(State code) {
if (code == State.ERROR)
sleep(Math.pow(attempt, 2), TimeUnit.SECONDS);
attempt++;
connect();
}
public void onConnect() {
attempt = 0;
}
This code is slightly more complicated, but all it does is track the number of consecutive failed attempts to connect, and then make the delay longer every time. This code yields the following delays:
0, 1, 4, 9, 16, 25 ... // etc.
On the very first disconnection, an immediate reconnect is attempted. If that fails, we wait one second. If it fails again, we wait four seconds, and so on. This approach is great, because it neatly matches the nature of errors in networks and systems:
- A transient network failure (such as packet loss) happens in the sub-second range.
- A server or router restart may happen in the range of 1–5 seconds.
- A brief loss of connectivity ranges 9–16 seconds (for example, when passing through a tunnel while talking on a cell phone).
In essence, we "bet" that the connection failure will fall into one of these categories. What's nice for us is that this function also fits a somewhat longer outage after only around 15 or 20 attempts (at 20 attempts, the delay will be over six minutes).
On Capping and Cubicity
Now, if a connection has been offline for a fairly long time (because we're out of WiFi range, for example), and it suddenly comes back online, we have no quick way to reconnect. The delay would have grown so long by this point, a few hours into back-off delay, that the application would incorrectly assume that connectivity is not forthcoming. Of course, we don't want the linear algorithm either, for the reasons we've already stated. The solution here is adding a cap. A cap is just an arbitrarily chosen limit beyond which we won't delay. The cap varies depending on the type of services and the service-level agreement between them, but for simplicity let's choose a cap of two minutes:
int attempt = 0;
public void onDisconnect(State code) {
if (code == State.ERROR)
sleep(Math.pow(attempt, 3), TimeUnit.SECONDS);
if (attempt < 6)
attempt++;
connect();
}
public void onConnect() {
attempt = 0;
}
Notice that I've made two changes here:
- The back-off function is now cubic (attempts to the power of 3).
- We cap attempt at 5 (one less than 6).
The back-off function now yields the following delays:
0, 1, 8, 27, 64, 125
I like the cubic function best, as I think it fits the scenarios I've outlined very well. This will vary with experience and problem domain, of course.
This short foray into the complex world of client-server architectures should have given you pause. Connecting two services can simultaneously be the easiest and trickiest part of a coordinated system; and it's especially important today, as we have so many interacting pieces of software on various machines. Rather than learning from eventual (and inevitable) crises, apply these patterns early to save yourself a lot of headaches when your projects make their brave way out into the real world, with all of its wonderful rewards and lurking perils.