IMVU’s backend had been written in PHP from the beginning. Startup code is always gross, but after years of learning PHP’s ins and outs, we bent it to our will. Part of me looks back and says “It wasn’t that bad”, and indeed, it has a lot of great properties. But then I remember the horror.
— Chad Austin (@chadaustin) March 12, 2014
Back in the early days of IMVU, we actually had to write a PHP test called test_pi_is_about_3. And it _occasionally failed_.
— Chad Austin (@chadaustin) April 7, 2016
The terrible straight-line performance, the lack of concurrent IO, the insane memory usage… Once the codebase reached about a million lines, the latency on our REST service response times was over half a second and in the worst case several seconds. For some types of services, that’s okay, but IMVU strives to be a real-time, delightful experience, and service latency on interactive things like dressing up, browsing the catalog, sending messages to your friends, and shopping is really important.
Some of the performance-minded of us tried to get PHP to be fast, but it was basically intractable, and we couldn’t afford to fork PHP with a high-performance JIT like Facebook ended up doing years later.
The combination of PHP’s awful performance, terrifying pitfalls, and the fact that millions of lines of code in a dynamic language (even with copious test coverage) makes it really hard to refactor, led us to evaluate alternatives. Someone suggested Java (which would have been totally fine in hindsight) but the infrastructure team decided to adopt this newfangled Node.js thing. Node sucked. This predated ES6, promises, and the rich ecosystem of npm, so we spent far too long rebuilding basic, basic, infrastructure (like propagating stack traces across async operations) in Node that we already had in PHP, and it still wasn’t as good. In the end we decided to drop Node and stick with PHP.
Fast forward a year or two…
There were some people at IMVU who had dabbled with Haskell, but it always seemed too idealistic and not very practical. In fact, one weekend, Andy went home with the explicit goal of proving that Haskell was stupid and not good for writing production software.
He came back on Monday glowing. “Haskell is perfect for web services! We should use it! It’s fast and safe and concurrent!” We smiled and nodded (sometimes Andy gets enthusiastic about new things) and went on with our daily work. But in the evenings and weekends he plugged away, and built a prototype web service that was 1) dramatically faster than the corresponding PHP services 2) actually quite short in implementation and 3) surprisingly clear. Huh, maybe this Haskell thing did have legs.
OK, so let’s consider the possibility that Haskell could be something IMVU could adopt. How we would vet it?
We started by, during the next hack week, forming a small team of everyday engineers to learn Haskell and work on a feature. We learned two things: teaching Haskell was a nontrivial problem. BUT, Haskell newcomers hacking on the codebase only caused one regression, and that regression actually exposed a bug in some of our infrastructure. Well that’s pretty cool. For years, IMVU had a tradition of throwing new engineers into the deep end, after which they’d invariably take down the site, and then we’d demonstrate that we’re egoless and all failures are an opportunity to improve our process with a postmortem. What if Haskell could help us avoid some of these problems?
Over time we slowly ramped up Haskell within the company. At first we had growing pains. The infrastructure in our PHP stack was extremely mature and well-baked. The corresponding infrastructure in our Haskell (error reporting, DB access, unit testing) was not. It took us a while to sort out exactly what our Haskell best practices would be, and by that point a few people started to get a negative vibe about Haskell. But we kept pushing forward.
(By the way, I keep saying “we”, but at this point I hadn’t written any Haskell yet, and was not particularly for or against it, even though I did think PHP was terrible.)
By this point we’d shipped at least two sets of mobile services on Haskell, and many people considered it a success. However… Tensions about Haskell had really started to heat up, with some people saying things like “I never want to even look at Haskell again” or “at this stage in my career I don’t want another programming language” or “I think we’re making a horrible mistake by building another stack, especially in Haskell.” People were complaining to their managers, and the managers were coming to me. Some people loved Haskell, some people hated it, but very few were in the middle.
As a neutral third party, and as someone with a lot of respect in the organization, the engineering managers sent me out to gather ground truth. I took my little black notebook and a pen and interviewed something like half the engineering team on their opinions on Haskell, whether they were fans or not. I spent an hour with each person, asking what they liked, what they didn’t like, what they thought we should do, etc.
I then collated all the feedback and presented back to the engineering managers and leadership team (of which I was a part too, but since I had no direct reports and had no experience with Haskell at IMVU, I was fairly unbiased). I don’t have my report handy, but it went approximately like this:
75% of the people I interviewed were positive about or fine with Haskell.
25% of the people had major concerns. The concerns, in roughly decreasing priority, were:
- Haskell strings suck
- Specifically there’s a lot of cognitive overhead in switching between the five common string representations as needed
- Syntactic concerns
- Haskell’s syntax is iconoclastic: parens,
- some people in the Haskell community love point-free style
- cognitive overhead in switching between pure and monadic style
- let vs. where
- Haskell’s syntax is iconoclastic: parens,
- concerns about a second backend stack
- duplicate infrastructure across PHP and Haskell
- concerns about it not being a mainstream language
- teaching the whole organization
- interaction between Haskell experts and novices sometimes leaves the novices feeling dumb or inadequate
- bus factor to organization on key infrastructure
- Haskell is less debuggable by ops engineers
The wins were:
- GHC’s support for safe concurrent programming is world-class
- Sound type safety led to refactoring being a breeze
- Fun, stretches the mind (for some people)
Ultimately, we settled on a middle ground. We would not deprecate PHP, but we would continue to support Haskell as a technology stack. We had specific services in mind that needed high throughput and low latency (<100 ms) and we knew that PHP could not achieve those. That said, we respected the fact that several engineers, even those in leadership roles, put their foot down and said they didn’t want to touch Haskell. (From my perspective it was kind of a weird experience. Highly-skilled people that I respect just got really angry and irrational and basically refused to even try to learn Haskell, which resulted in a culture divide in the company.)
In the end, I never did manage to convince some of my peers to try it. That said, there were many engineers who loved Haskell and we built a great deal of critical infrastructure in it. The Haskell infrastructure basically hums like a well-oiled machine. Perfectly reliable unit tests. Concurrent batched IO. Abstractions for things like timeouts and running tasks on other threads while correctly bubbling failures. Fantastic support for safely encoding and decoding JSON. One of the best benchmarking tools I’ve ever used.
Since one of the biggest concerns about Haskell was education, I taught a sequence of four or five classes for the engineering team with my take on how it should be taught. Traditional Haskell materials tend to talk about laziness and currying and all this FP junk that is certainly interesting and different, but mostly irrelevant to people trying to come in hot and get the job done. Instead, I focused on specifics first, and then generalizations.
“You know how you can add two integers and get a new integer back? Well, you can also add two floating point numbers. Int and Float are instances of Num.”
“You know how you can concatenate strings? And concatenating with the empty string gives you the same value? It turns out many things satisfy these properties. They’re called Monoids.”
“You know how you can map across a list? We do it all the time in PHP or Python. Well you can also map across Maybe. Think of Maybe as a list that can have either zero or one element. Things that can be mapped are called Functors.”
We started with specific examples of how to do stuff with Haskell, then talked about type classes, and the final optional class was about how Haskell is compiled down to the machine. You’d have to ask the attendees how much they remember, but the classes were well-reviewed at the time. :)
In summary, I definitely wouldn’t say that Haskell was a failure. GHC is an amazing piece of technology, and IMVU built infrastructure in it that wouldn’t have been possible in PHP. I also can’t say Haskell was a resounding success. I was truly disturbed by the culture drama that it brought. And, frankly, saddened and confused by how some people closed their minds. Personally, I would have happily traded a lot of Haskell’s safety for avoiding drama. I sometimes wonder what would have happened had we chosen Java instead of Node.js back in 2010. We could have hit our performance goals in Java too.
Oh well. Many of us still love Haskell. It’s not going away. I left IMVU last year (needed some more experience elsewhere) but I personally miss Haskell a great deal. If I were starting a new project with people that I trusted, I’d definitely reach for Haskell first. The expressiveness, brevity, and safety let you move extremely quickly, and Haskell is definitely mature enough for production use.