I can’t think of a single large software company that doesn’t regularly draw internet comments of the form “What do all the employees do? I could build their product myself.” Benjamin Pollack and Jeff Atwood called out people who do that with Stack Overflow. But Stack Overflow is relatively obviously lean, so the general response is something like “oh, sure maybe Stack Overflow is lean, but FooCorp must really be bloated”. And since most people have relatively little visibility into FooCorp, for any given value of FooCorp, that sounds like a plausible statement. After all, what product could possible require hundreds, or even thousands of engineers?
A few years ago, in the wake of the rapgenius SEO controversy, a number of folks called for someone to write a better Google. Alex Clemmer responded that maybe building a better Google is a non-trivial problem. Considering how much of Google’s $500B market cap comes from search, and how much money has been spent by tens (hundreds?) of competitors in an attempt to capture some of that value, it seems plausible to me that search isn’t a trivial problem. But in the comments on Alex’s posts, multiple people respond and say that Lucene basically does the same thing Google does and that Lucene is poised to surpass Google’s capabilities in the next few years. It’s been long enough since then that we can look back and say that Lucene hasn’t improved so much that Google is in danger from a startup that puts together a Lucene cluster. If anything, the cost of creating a viable competitor to Google search has gone up.
Businesses that actually care about turning a profit will spend a lot of time (hence, a lot of engineers) working on optimizing systems, even if an MVP for the system could have been built in a weekend. There’s also a wide body of research that’s found that decreasing latency has a roughly linear effect on revenue over a pretty wide range of latencies for some businesses. Increasing performance also has the benefit of reducing costs. Businesses should keep adding engineers to work on optimization until the cost of adding an engineer equals the revenue gain plus the cost savings at the margin. This is often many more engineers than people realize.
And that’s just performance. Features also matter: when I talk to engineers working on basically any product at any company, they’ll often find that there are seemingly trivial individual features that can add integer percentage points to revenue. Just as with performance, people underestimate how many engineers you can add to a product before engineers stop paying for themselves.
Additionally, features are often much more complex than outsiders realize. If we look at search, how do we make sure that different forms of dates and phone numbers give the same results? How about internationalization? Each language has unique quirks that have to be accounted for. In french, “l’foo” should often match “un foo” and vice versa, but American search engines from the 90s didn’t actually handle that correctly. How about tokenizing Chinese queries, where words don’t have spaces between them, and sentences don’t have unique tokenizations? How about Japanese, where queries can easily contain four different alphabets? How about handling Arabic, which is mostly read right-to-left, except for the bits that are read left-to-right? And that’s not even the most complicated part of handling Arabic! It’s fine to ignore this stuff for a weekend-project MVP, but ignoring it in a real business means ignoring the majority of the market! Some of these are handled ok by open source projects, but many of the problems involve open research problems.
There’s also security! If you don’t “bloat” your company by hiring security people, you’ll end up like hotmail or yahoo, where your product is better known for how often it’s hacked than for any of its other features.
Everything we’ve looked at so far is a technical problem. Compared to organizational problems, technical problems are straightforward. Distributed systems are considered hard because real systems might drop something like 0.1% of messages, corrupt an even smaller percentage of messages, and see latencies in the microsecond to millisecond range. When I talk to higher-ups and compare what they think they’re saying to what my coworkers think they’re saying, I find that the rate of lost messages is well over 50%, every message gets corrupted, and latency can be months or years1. When people imagine how long it should take to build something, they’re often imagining a team that works perfectly and spends 100% of its time coding. But that’s impossible to scale up. The question isn’t whether or not there will inefficiencies, but how much inefficiency. A company that could eliminate organizational inefficiency would be a larger innovation than any tech startup, ever. But when doing the math on how many employees a company “should” have, people usually assume that the company is an efficient organization.
This post happens to use search as an example because I ran across some people who claimed that Lucene was going to surpass Google’s capabilities any day now, but there’s nothing about this post that’s unique to search. If you talk to people in almost any field, you’ll hear stories about how people wildly underestimate the complexity of the problems in the field. The point here isn’t that it would be impossible for a small team to build something better than Google search. It’s entirely plausible that someone will have an innovation as great as PageRank, and that a small team could turn that into a viable company. But once that company is past the VC-funded hyper growth phase and wants to maximize its profits, it will end up with a multi-thousand person platforms org, just like Google’s, unless the company wants to leave hundreds of millions or billions of dollars a year on the table due to hardware and software inefficiency. And the company will want to handle languages like Thai, Arabic, Chinese, and Japanese, each of which is non-trivial. And the company will want to have relatively good security. And there are the hundreds of little features that users don’t even realize that are there, each of which provides a noticeable increase in revenue. It’s “obvious” that companies should outsource their billing, except that when you talk to companies that handle their own billing, they can point to individual features that increase conversion by single or double digit percentages that they can’t get from Stripe or Braintree. That fifty person billing team is totally worth it, beyond a certain size. And then there’s sales, which most engineers don’t even think of2; the exact same line of reasoning that applies to optimization also applies to sales – as long as marginal benefit of adding another salesperson exceeds the cost, you should expect the company to keep adding salespeople, which can often result in a sales force that’s larger than the engineering team. There’s also research which, almost by definition, involves a lot of bets that don’t pan out!
It’s not that all of those things are necessary to run a service at all; it’s that almost every large service is leaving money on the table if they don’t seriously address those things. This reminds me of a common fallacy we see in unreliable systems, where people build the happy path with the idea that the happy path is the “real” work, and that error handling can be tacked on later. For reliable systems, error handling is more work than the happy path. The same thing is true for large services – all of this stuff that people don’t think of as “real” work is more work than the core service3.
I often make minor tweaks and add new information without comment, but the original version of this post had an error and removing the error was a large enough change that I believe it’s worth pointing out the change. I had a back of the envelope calculation on the cost of indexing the web with Lucene, but the numbers were based on benchmarks results from some papers and comments from people who work on a commercial search engine. When I tried to reproduce the results from the papers, I found that it was trivial to get orders of magnitude better performance than reported in one paper and when I tried to track down the underlying source for the comments by people who work on a commercial search engine, I found that there was no experimental evidence underlying the comments.
I’m experimenting with writing blog posts stream-of-consciousness, without much editing. Both this post and my last post were written that way. Let me know what you think of these posts relative to my “normal” posts!
Thanks to Leah Hanson, Joel Wilder, Kay Rhodes, and Ivar Refsdal for corrections.