Lean Enterprise: How High Performance Organizations Innovate at Scale (Lean (O'Reilly))

Author: Jez Humble, Joanne Molesky, Barry O'Reilly
4.4
This Month Stack Overflow 1

Comments

by anonymous   2019-01-13

I think that in cases involving performance optimization, you really want to create an A/B test to determine which one you should do. There really is no cookie cutter answer for image preloading that applies to everybody as a best practice.

One of the biggest tenets of a favorite book of mine, Lean Enterprise, is use to A/B tests to prove or disprove a HIPPO (highly paid person's opinion). Certain opinions carry a lot of weight, both in your organization and on the internet. Because of their importance and reputation, their opinions veer towards the realm of fact, even though it may not be.

The practice of measuring performance empirically is also touted by another book I love which deals with performance tuning - Code Complete 2nd Ed. In that book, McConnell gives several code examples where you would expect one piece of code to be optimal, but in fact, it performed poorly (see chapters 25 and 26). One of his key points is that you should always test a performance optimization. If it isn't worth testing, it isn't worth writing the "highly performant" code in the first place. McConnell's premise doesn't just apply to his low level coding examples, but to high level decisions such as preloading images above the fold as well.

I can also attest to the importance of A/B testing at a professional level. I used to work on Amazon's SEO team and we A/B tested everything. The fact of the matter is that you never really know how customers will respond to something. Nobody can predict customer behavior - not even Jeff Bezos - and you really need to back up your hypothesis with real data to prove or disprove the validity of what you're doing.

Even though you can find multiple blog articles and online sources discussing whether preloading is better or not, you don't really know whether that will work for you until you've done it. Different people have different servers with different performance characteristics and different network topologies, etc. You just don't know which way is better for you until you have the data. If you launch your A/B test and find that find that your repel rate goes up when preloading, then you know that you have to dial back your treatment and return to your control. If however, you find that customers don't get bored waiting and click through at a higher rate than before, then you have a winner and you dial up the treatment - deleting the control code entirely after a period.

I hope that helps.