[erlang-questions] trouble with erlang or erlang is a ghetto

[Joel Reymont]

I'll offer a datapoint… 

I hate Erlang, although it's been a great way for me to make money. OpenPoker, for example, keeps on giving and is being licensed by a major game development company.

To Tony's list of Erlang warts I would also add the complete lack of transparency during performance optimization. This has bit me during OpenPoker development and also during the "Hot wheels" optimization contest [1]. 

You do have the facility for call counting (eprof) as well as for per-process profiling with fprof. A regular Erlang system these days will likely have thousands of processes running, processing requests and returning a response. What you get with fprof is a useless profiling listing of thousands of processes. 

What you really want is to know what happens along the request processing path. It took N seconds to process the request. Where did this time go? You want to average this over hundreds of requests and see where the bottlenecks are. Do you really want this? You are out of luck, I'm sorry. No tools exist to help you.

That said, I recently came off a project where we build an ad serving platform with OCaml and ZeroMQ. I pushed for this technology stack as I wanted to see if this is a suitable replacement for Erlang. The short answer is no.

We ended up defining tons of Google Protocol Buffers messages for everything we needed and writing boring serialization and deserialization code. Also, what would normally be an Erlang process turned out to be a separate executable. There was some loss of compilation speed as the number of these executables multiplied. Still, it was quite manageable until we got to sharing data. 

We started by using plain Redis. This required a server to "front" Redis and process various requests, updating other processes of state changes via ZeroMQ. Serializing things to strings and back is an incredible pain in the ass!

Later, it turned out that Redis was not the ideal choice since a lot of the long-running processes required most of the data in Redis and it was too slow to suck it out upon startup of each of these processes. We ended up dumping data into "bootstrap files", easily and quickly loadable into OCaml. 

What I'm trying to say here is that our application clearly benefited from code running "on top" of the data, e.g. Erlang processes using ETS or Mnesia. Our scores of Protobuf messages would have become simple Erlang tuples or records. Data sharing would have become a non-issue. 

I wish Tony good luck with his search for the Erlang replacement. I haven't found one yet.

[1] https://groups.google.com/d/topic/erlang-programming/2pRrWneJwG8/overview

--------------------------------------------------------------------------
- for hire: mac osx device driver ninja, kernel extensions and usb drivers
---------------------+------------+---------------------------------------
http://wagerlabs.com | @wagerlabs | http://www.linkedin.com/in/joelreymont
---------------------+------------+---------------------------------------