Taco Bell Programming
by Ted Dziuba on Thursday, October 21, 2010
Every item on the menu at Taco Bell is just a different configuration of roughly eight ingredients. With this simple periodic table of meat and produce, the company pulled down $1.9 billion last year.
The more I write code and design systems, the more I understand that many times, you can achieve the desired functionality simply with clever reconfigurations of the basic Unix tool set. After all, functionality is an asset, but code is a liability. This is the opposite of a trend of nonsense called DevOps, where system administrators start writing unit tests and other things to help the developers warm up to them - Taco Bell Programming is about developers knowing enough about Ops (and Unix in general) so that they don't overthink things, and arrive at simple, scalable solutions.
Here's a concrete example: suppose you have millions of web pages that you want to download and save to disk for later processing. How do you do it? The cool-kids answer is to write a distributed crawler in Clojure and run it on EC2, handing out jobs with a message queue like SQS or ZeroMQ.
The Taco Bell answer? xargs and wget. In the rare case that you saturate the network connection, add some split and rsync. A "distributed crawler" is really only like 10 lines of shell script.
Moving on, once you have these millions of pages (or even tens of millions), how do you process them? Surely, Hadoop MapReduce is necessary, after all, that's what Google uses to parse the web, right?
Pfft, fuck that noise:
find crawl_dir/ -type f -print0 | xargs -n1 -0 -P32 ./process
32 concurrent parallel parsing processes and zero bullshit to manage. Requirement satisfied.
Every time you write code or introduce third-party services, you are introducing the possibility of failure into your system. I have far more faith in xargs than I do in Hadoop. Hell, I trust xargs more than I trust myself to write a simple multithreaded processor. I trust syslog to handle asynchronous message recording far more than I trust a message queue service.
Taco Bell programming is one of the steps on the path to Unix Zen. This is a path that I am personally just beginning, but it's already starting to pay dividends. To really get into it, you need to throw away a lot of your ideas about how systems are designed: I made most of a SOAP server using static files and Apache's mod_rewrite. I could have done the whole thing Taco Bell style if I had only manned up and broken out sed, but I pussied out and wrote some Python.
If you don't want to think of it from a Zen perspective, be capitalist: you are writing software to put food on the table. You can minimize risk by using the well-proven tool set, or you can step into the land of the unknown. It may not get you invited to speak at conferences, but it will get the job done, and help keep your pager from going off at night.
This is great until your job outgrows a single computer or you want to have redundancy. Also, chains of bash tools don't have the best error management when something chokes in one of the middle steps in the pipe. You can still leverage simple bash tools for a lot of the under the hood stuff, but you start needing something more to act as the glue petty quickly when you scale. KISS should still apply.
I think that's the point. Don't jump to the complex right away. Keep it simple and compose the capabilities you have readily available until you need to become more complex. When the task requires it, yeah, do the complex thing, but keep the simplicity mandate in mind and only add the new complexity that you need. You can get pretty far with the simple, and what about all of the situations where that future pivot or growth never happens?
The philosophy strikes a cord with me - I'm often contending with teams that are building for the future complexities that they think might come up, and we realize later that we did get complexity in the problem later, but not the kind we had planned for, so all of that infrastructure and planning was wasted on an imaginary problem that no only didn't help us but often actually make our task harder. The trick is to keep the solution set composable and flexible so that if complexity shows up later, we can reconfigure and build the new capabilities that we need rather than having to maneuver a large complicated system that we built on a white board before we really knew what the problem looked like.
It's more complex to have 10 different ways to do the same thing. Like, just take a week to teach your ops team how to use Docker and Kubernetes, so everything can simplified to just one Kubernetes cluster instead of 20 bespoke EC2 instances.