A quest for a "standard" library with uniform, composable abstractions.
Originally motivated by a desire for an error monad and generic option type operations, and stood for Higher Order Programming in Erlang. Soon after, I wished all standard containers used consistent conventions and protocols (such as consistent accessor names, argument positioning rules and expression of semantics with option and result types).
Here lies an experiment to see what something like that could look like. As all proper experiments should, this one is used daily in production projects (hence the high-ish version number, 'cause semver).
I entertain any forward-thinking library design ideas, but more than anything else, these are influenced by Jane Street's Core of the OCaml world.
- A module per data type implementation
- Name of the module is the name of the type
- Inside the module, the type it implements is always named t(..), such as:
hope_foo:t()
, nothope_foo:foo()
- t(..) is always the first argument
- Names of private records may be short, such as:
#foo{}
or#t{}
(Though I'm second-guessing this idea, since seeing{t, ..}
in stack traces is less than helpful. I'm considering requiring fully-qualified names for all record definitions and maybe short-handing what would've been#t{..}
as-define(T, ?MODULE). -record(?T, {..}).
, which may be a bit ugly. Still thinking...) - Names of public records must be fully qualified, such as:
#hope_module_record{}
- Names of all modules must be fully qualified, such as:
hope_module
(this should go without saying, but just to be sure...) - Keep the number of (anonymous) arguments "reasonably" low:
- up to 3 is normal
- 4 is suspicious but may be reasonable
- 5 is very suspicious and probably unnecessary
- more than 5 is unacceptable, so consider reducing by:
- revising abstractions, or, if not practical
- creating a public record specifically for the purpose of passing many arguents, which simulates labeled arguments. For an example see https://github.com/xandkar/oauth1_core where I used that technique extensively (especially in oauth1_server.erl)
A class of burritos, used for expressing sequences of operations on some data
type. Defined in hope_gen_monad
, implemented as:
hope_result
: for composition of common functions returning{ok, Val} | {error, Reason}
. An alternative to exceptions, which makes the error conditions apparent in the spec/signature. Analogous to Haskell'sData.Either a b
, Jane Street Core's (OCaml)('a, 'b) Result.t
, Rust'sResult<T, E>
hope_option
: for expressing and composing the intention that the value may or may not be available. An alternative to the commonundefined
(which is equivalent to the dreadednull
). Analogous to ML's (SML, OCaml, etc)'a Option.t
, Rust'sOption<T>
and Haskell'sData.Maybe a
1.
A class of abstract data types to which we have exclusive access and can put things in and take them out. See issue #9
- Operations on all abstract types of containers should share a common lexicon
- Concrete implementations of an abstract data type must be swapable
Defined in hope_gen_dictionary
, implemented as:
hope_kv_list
. Equivalent to orddict/proplist. Operations implemented with BIFs fromlists
module, where possible
TBD:
hope_hash_tbl
. API around stdlib'sdict
hope_gb_dict
. API around stdlib'sgb_trees
TBD:
hope_hash_set
. API around stdlib'ssets
hope_gb_set
. API around stdlib'sgb_sets
TBD
Should include both FIFO (queue) and LIFO (stack), so that user can swap if a different order is desired.
Should we attempt to include priority queues or make them a separate abstract type?
TBD
Not yet defined and only partially implemented as:
hope_list
A class of abstract systems to which we share access with an unknown number of users and can make requests to perform operations which may not get done for any number of reasons.
TBD
See issue #11