How Facebook Ships Code

Sometimes, when software has been around for 10+ years, stuff does get too old. It happens. For real.

I know people seem to forget that here with startups having been around for less than a year being "old players" and all that, but common.

Software may have been written with a bunch of presumptions which was valid when the project was initially started, but no longer are that. Maybe it was written on top of a platform which is no longer as productive as it was when the project was initiated. Nothing kills enthusiasm and productivity like working on a platform no longer deemed modern.

And then you have technical limitations and pre-conditions. Once, servers was expensive and you wanted your solution to distribute it's load to clients so that you could save on servers. Once servers was the only machines powerful enough to complete time-critical tasks within reasonable time-frames, so you architected the software accordingly.

Now, we suddenly have HTML5 and webworkers and shit, so you probably want your solution to be a web-solution with a distributed computation model. I.e. full-circle. Full rewrite. Twice. What's next, I don't know.

Sometimes a rewrite is just right. Saying that every rewrite is based in incompetence is simplistic.

Sorry, but I'm really not. In the case of Chrome, for example, I would routinely test new work on web projects in all the major browsers. I am writing this with some empirical data in front of me, because I've just checked the bug tracking systems for a couple of projects I work on to be sure: for both projects, excluding mobile browsers, Chrome has been responsible for a clear majority of all defects where the root cause was found to be a browser bug over the past two years.

I don't doubt that there are bugs in Docs or Chrome, but they are relatively obscure.

Respectfully, if they were that obscure, my colleagues and I wouldn't keep running into them on multiple projects. I'd agree that the particular symptoms of any particular bug are usually a corner case: set this option to A and that option to B and it breaks, but other combinations work OK. It's the way that several of these bugs have recurring themes that betray both an underlying architectural weakness and a lack of effective quality control that I find most unfortunate.

Using Chrome as an example again, it is clearly aggressive with caching and conservative with repainting, but sometimes that means it is simply not behaving properly at all. If I set a part of the DOM to display instead of being hidden and then send an AJAX request, I want my "please wait" message displayed while the request is running, not afterwards, or indeed not at all since it probably gets hidden again as soon as the response arrives.

That's the nature of software, not every bug is caught by a unit test and users end up seeing them and reporting them.

And as I mentioned, several of those bugs in Chrome had been reported, and subsequently closed without the root cause ever being identified and fixed.

Expecting low-cost web apps to be as reliable as airplane control systems is unrealistic.

I don't expect Google's software to be as reliable as airplane control systems, but somewhere close to as reliable as everyone else's software would be nice. I appreciate that you're having trouble believing it or reconciling it with your own experience, and I've already acknowledged that my experience might be a complete outlier, but I'm looking at several years of empirical data across multiple completely different projects and development teams and it is quite clear that the Google products I'm looking at here haven't been keeping up lately.

The issue was closed almost immediately, with some obviously hastily written comments, apparently because no-one could reproduce it in a different version of Chrome on Linux or Mac. As far as I can see, no-one even tried to reproduce the other reported failing case on Windows, and there was no attempt at all to investigate the original bug and determine how it happened and why it was no longer observable on the platforms tested.

The issue was simply marked "fixed", despite no actual fix having been identified, rather than giving it a more specific "no longer reproducible" status.

There are still serious problems with that combination of attributes today.