Re: Records in Haskell

2012/1/14 Johan Tibell :

On Fri, Jan 13, 2012 at 3:52 PM, Simon Peyton-Jones wrote:

...

I know of no proposal that advocates only (A).  It seems that we are agreed that we must make use of types to disambiguate common cases.

I will try to make the case for (A), just so it has been put on the table.

Proposal =========

The proposal is to implement http://hackage.haskell.org/trac/ghc/wiki/Records/NameSpacing without any of the type resolution parts. I think this approach is dismissed a bit to easily on the wiki page above:

   "We have name-spaces, but it is hard to see how this is better than the current practice of adding prefixes to record fields: data Record = Record { recordA :: String }"

There are (at least) three benefits of using namespaces (e.g. 'Record.a') rather than ad-hoc prefixes (e.g. 'recordA'):

 * You can use a type synonym to abbreviate the namespace part (as shown on the wiki page.)

 * If there's no ambiguity you don't need to use a namespace (e.g. you can use 'a' instead of 'Record.a').

 * The namespace name is predictable (e.g. <Typename>.<fieldname>) while ad-hoc prefixes tend to use different conventions e.g. the whole record name (e.g. 'recordA') or some abbreviation thereof (e.g. 'rcrdA'.)

The main argument for this approach is its simplicity; it's simple to understand for users and (hopefully) simple to implement.

At work we have extended the OCaml compiler with the same mechanism, to solve the same problem. I find it quite annoying to use, as you tend to try without a namespace name first, and if you get a compile error you add one. I'd prefer a system which just infers the right selector - less time spent fighting with the compiler! David