The Cost of no Architecture

Like many others, I enjoy various reverse engineering and tear-down stories. Personally, I mean things like iFixit tear-downs and Ken Shirriff’s blog, so I started following this tweet thread by foone.

This continues with another tweet sequence about getting software running on the remote control. Having enjoyed these tweets, I started thinking.

The Harmony remotes are quite expensive in my mind. I can’t find any exact numbers for the number of sold devices, but I found this 2018 Q4 earnings report. Looking at the net sales, I guess the remotes are either “Tablets & Other Accessories” or “Smart Home”. They represent sales net sales of ~107 and ~89 MUSD over 12 months. Let’s pick the lower number and just look at magnitudes. The Harmony 900 seems to have retailed for ~350 USD back when it was new. So, if all the Smart Home stuff was harmonies, we’re looking at 250k units over a year. So I’m guessing the magnitude is around 10k – 100k units annually – but the Harmony 900 is from 2013, so I assume that it sold closer to the lower number, if not below. The market was new and so on.

Then we look at the tweets again. What have we got? Let’s put aside security issues, unencrypted communications, and other clear mistakes and just look at how the device is built.

Flash to drive the UI, double web servers on-board, Lua, QNX and what not. A 233 MHz CPU and ~64MB of FLASH – for a remote control. From an engineering perspective, this sounds like a fun system to work on – from an architecture perspective, it looks like a ball of mud.

Back in 2013, QNX might have been a good choice compared to Linux. Today, with Yocto and similar tools for developing embedded Linux systems, it feels like an odd choice to add a license cost to such a device. But no biggie. Back in the day this was not an unreasonable choice (and still isn’t for certain applications).

The Flash stuff. There were alternatives back in 2013, but sure, there were plenty of developers at hand and things like Qt QML was still probably a bit clunky (I can’t recall the state of it back then – it required OpenGL ES, which I guess was a big ask back then).

But the mix of techniques and tools. The on-board web servers. The complexity of a small system and the costs it brings to maintenance and testability. If this is the foundation for Harmony remotes and a platform that has been used for the better past of the past decade, I wonder if the added engineering costs for architecture the platform to be more optimized early on would not have paid off in lower maintenance costs, as well as lower hardware costs.

I know how it is when you’re in a project. The deadline is there in big writing on one of the walls. You can get something working by stringing what you have together with duktape and glue. The question I’m asking myself is more along the lines of how do we run embedded systems engineering projects? Where did we go wrong? Why don’t we prioritize the thinking and refactoring over the just-get-this-thing-out-of-the-door?

The answer is time to market and such, but over a decade of building on a ball of mud, the economical numbers start adding up in favour for the better engineered product. For continuous improvement. For spending time thinking about how to improve the system as a whole.