The Problem With Feature Flags When Using Next.js
Do you use feature flags to get the benefits of modern, high-velocity software deployment? You do, right? Maybe you’ve noted a problem when using next.js. Not to fret; there is a solution.
Next.js provides awesome “instant” page startup using techniques like server-side rendering. But as part of that startup, the browser must retrieve the current settings of your flags. As a result, you’ll be able to enable or disable features appropriately for that user.
This may leave your app in a challenging “hurry up and wait” situation. You deliver your app’s page instantly. But then must wait for some kind of network call to determine the values of your flags. A CDN can make this experience significantly faster, but if you are looking for the absolute fastest startup experience, even that hundred millisecond round trip to the CDN may seem undesirable. You’ve invested in technology to give great page-load experiences, and then slowed that down with your necessary infrastructure.
Here, I’ll talk about a technique you can use to “have your cake and eat it too”. By using Split’s feature flagging SDK’s you can get both instant page loading, and instant (within a couple milliseconds) feature flag settings all without additional network calls.
There are many things we could discuss around feature flags in next.js. For example, you could use feature flags on the back end parts of your application to manage data persistence, in the mid-tier where server-side rendering or static-side generation is done, as well as in the browser. For right now, let’s focus on the following case.
A Refresher on Feature Flags
This, in turn, reveals a second great benefit of feature flags. You can separate deployment from releasing of the feature. Therefore, you can control that release via remote configuration, with precise control over the process. Rather than turning a feature on for all users and hoping all goes well, instead you can turn it on gradually. You can make the feature available to 1% of users, then 5% to 10% and so on. If something goes wrong with the new feature, you can turn it back off immediately. With these tools, releasing can be a no-stress process.
The capstone of feature flags, however, is the ability to measure changes in your application and correlate those back to individual features. Did the dollar value of your customer purchases go down? Was there a particular feature that caused this? Without feature flags, how would you know if an individual feature had either good or bad impact? Without that information, how can you make good business decisions? When coupled with a feature-flag aware A/B testing system, you can run experiments to discover all of the above while you release new features.
Given all these benefits, why would you not be using feature flags?
Ordinary Behavior of the Split SDK
One of the great benefits that the Split feature flag SDK offers is that it can determine whether a flag is on or off with calculations that are entirely on the local system in the SDK. This is determined entirely by local calculations within the SDK.
With local calculation, your code can decide whether to show a feature or not in the space of a couple milliseconds. This is possible since it does not have to communicate with a SaaS server to make these decisions. This has significant privacy benefits since the information used to inform this flag decision never leaves the client device. But for our purposes, it also means that once your app is up and running it is always going to execute at the fastest rate possible. As a result,this will deliver a smooth experience for your users.
For this to work, the SDK needs to get a copy of the feature flag definitions at application startup.
When the Split SDK first downloads the feature flag definitions, it stores a copy of those definitions in local storage. This will allow it to start up faster the next time the app is run in the same brower. Since an initial set of the flag definitions is available, the application can start up without waiting for the definitions to be downloaded.
Faster Than Really Fast
The picture we’ve summarized so far already is a really fast feature flag system. On all but the first run, your application already has a set of flag definitions available. Thus a network call is not needed to get started, and no network calls are needed for feature flag evaluation. This is already fast, with less than a second of overhead in the worst case (first time start up).
But for applications seeking the absolute fastest performance, can we eliminate the sub-second overhead? After all, nothing is better than better, right?
Suppose we were to deliver a copy of the feature flag definitions during the initial server-side rendering of a page. Then install them in LocalStorage before starting up the Split SDK. This seems like it would give us a complete feature flag solution with absolutely no overhead. Is it possible?
It is possible! Let’s take a look at one solution.
We can take advantage of all this and the server side rendering support in next.js:
- Keep a cached and up to date copy of all the feature flag definitions on the back end.
- When responding to a page request, include a copy of the definitions in the server-side rendered page.
- In the browser, store these definitions in LocalStorage, then start up the Split SDK, so it uses the definitions previously stored in LocalStorage instead of performing a network call.
On your server side, you need to keep an up to date version of all the feature flag definitions. You might store this information in Redis, or some other store accessible from your next.js instances.
For example, the following code fragment retrieves the flag definitions every second and stores the data in some durable storage. The point is to periodically retrieve any changes to the flag definitions, and then to integrate those into the cache.
To make this work, we create a custom application which delivers the feature flag definitions via server side rendering. It installs those into local storage and then starts up the Split SDK:
populateLocalStorage() function takes the server-side rendered flag definitions and reconstructs the cache in
How fast does this make the application startup?
This is the console output from several runs I just made:
- SDK Ready from cache. 3ms to get ready.
- SDK Ready from cache. 4ms to get ready.
- SDK Ready from cache. 4ms to get ready.
So, we can see it takes perhaps as many as four milliseconds to store the definitions, start up the SDK and be ready to go. The “cost” of running a cutting edge feature flagging system here is less than one re-rendering of a simple app’s page. Something even your most detail-oriented users will not notice.
An Added Benefit: Feature Flags 100% of the Time
This brings another beneficial result. By delivering the flag definitions in this way, your feature flagging will operate correctly without the interference of ad-blockers or other software.
Ordinarily, when the Split SDK starts up (and this is true for other SaaS based feature flagging systems, too), the network request to get feature flags can be blocked, leaving your app in an undesirable “default state”.
While the impressions and events may still not be delivered, you can be assured with this solution that your users are always seeing the correct set of features in your application.
For applications seeking the absolute highest performance, it is possible to use feature flags, while paying a tiny number of milliseconds of “overhead”.
You do not have to make a tradeoff between the highest performance apps and taking advantage of the best of modern developer-velocity practices.