[erlang-questions] multiple behaviours

[Ulf Wiger]

On 6 Jun 2012, at 15:33, Tony Rogvall wrote:

>> Or, as you also suggest, write a new behaviour combining those two. If
>> this situation is not common in your design I would suggest this
>> solution as prevents you from going against what I think is the
>> intended use of behaviours. I guess the original idea was to have at
>> most a behaviour per module, but more authoritative voices may have a
>> word about this.
>> 
> Any takers on this topic (meaning people involved in the design of course)

This has been debated.

Since my part-time job is to fill the list with old gossip about 
the time before discussions moved to the web, this is the story
of behaviors (at least as I recall it). For my current opinion,
skip to the last paragraph.

The first behavior-style components were actually built in 
BOS, a middleware layer designed for, and partly by, the 
first real commercial erlang-based product: Mobility Server.

BOS was the predecessor of OTP. When designing OTP,
the concept of behaviors was evangelized by Joe Armstrong
as a more generalized version of the support libraries in BOS.

Back in those days, it wasn't easy at all to define your own
behavior; the linter had hard-coded knowledge of the ones in
OTP, and would not help with any 'fake' behaviors.

After having read a nice thesis (can't recall where to find it)
about behaviors, which also described how to hack your 
own behaviors - painful, but rewarding if you could prevail.

I wrote a prototype that introduced the behaviour_info/1
callback. It was accepted by OTP, with typos and all 
(I had used the American spelling, and it took a while 
before someone pointed this out; by this time, the 
easiest remedy was to accept that the erlang linter 
had become bi-lingual).

There were discussions now and then about whether it
was ok to mix behaviors in a single module. The OTP 
stance was that it was frowned upon, but their case was
weakened by the fact that this is done in a number of 
places in OTP (e.g. kernel.erl). Nowadays, the linter
accepts multiple behaviors, as long as they don't lay
claim to the same callbacks (checking Fun/Arity only).

With the new -callback() types, it should be possible to
accept overlapping callbacks, as long as their type 
signatures are compatible (preferrably the same).

Ulf Wiger, Co-founder & Developer Advocate, Feuerlabs Inc.
http://feuerlabs.com