dep-t-advice
This package is a companion to
dep-t. It provides a mechanism for
handling cross-cutting concerns in your application by adding "advices" to the
functions in your record-of-functions, in a way that is composable and
independent of each function's particular number of arguments.
Rationale
So, you have decided to structure your program in a record-of-functions style,
using dep-t. Good choice!
You have already selected your functions, decided which base monad use for
DepT
, and now you are ready to construct the environment record, which serves
as your composition
root.
Now seems like a good moment to handle some of those pesky "croscutting
concerns", don't you
think?
Stuff like:
- Logging
- Caching
- Monitoring
- Validation
- Setting up transaction boundaries.
- Setting up exception handlers for uncaught exceptions.
But how will you go about it?
A perfectly simple and reasonable solution
Imagine that you want to make this function print its argument to stdout:
foo :: Int -> DepT e_ IO ()
Easy enough:
foo' :: Int -> DepT e_ IO ()
foo' arg1 = do
liftIO $ putStrLn (show arg1)
foo arg1
You can even write your own general "printArgs" combinator:
printArgs :: Show a => (a -> DepT e_ IO ()) -> (a -> DepT e_ IO ())
printArgs f arg1 = do
liftIO $ putStrLn (show arg1)
f arg1
You could wrap foo
in printArgs
when constructing the record-of-functions,
or perhaps you could modify the corresponding field after the record had been
constructed.
This solution works, and is easy to understand. There's an annoyance though:
you need a different version of printArgs
for each number of arguments a
function might have.
And if you want to compose different combinators (say, printArgs
and
printResult
) before applying them to functions, you need a composition
combinator specific for each number of arguments.
The solution using "advices"
The Advice
datatype provided by this package encapsulates a transformation on
DepT
-effectful functions, in a way that is polymorphic over the number of
arguments. The same advice will work for functions with 0
, 1
or N
arguments.
Advices can't change the type of a function, but they might:
-
Analyze and change the values of the function's arguments.
-
Add additional effects to the function, either effects from the base monad, or effects from handlers found in the environment.
-
Change the result value of the function.
-
Sidestep the execution of the function altogether, providing al alternative result.
Here's how a printArgs
advice might be defined:
printArgs :: forall e_ m r. MonadIO m => Handle -> String -> Advice Show e_ m r
printArgs h prefix =
makeArgsAdvice
( \args -> do
liftIO $ hPutStr h $ prefix ++ ":"
hctraverse_ (Proxy @Show) (\(I a) -> liftIO (hPutStr h (" " ++ show a))) args
liftIO $ hPutStrLn h "\n"
liftIO $ hFlush h
pure args
)
The advice receives the arguments of the function in the form of an n-ary
product
from sop-core. But it
must be polymorphic on the shape of the type-level list which indexes the
product. This makes the advice work for any number of parameters.
The advice would be applied like this:
advise (printArgs stdout "foo args: ") foo
Advices should be applied at the composition root
It's worth emphasizing that advices should be applied at the "composition
root",
the place in our application in which all the disparate functions are assembled
and we commit to a concrete monad, namely DepT
.
Before being brought into the composition root, the functions need not be aware
that DepT
exists. They might be working in some generic MonadReader
environment, plus some constraints on that environment.
Once we decide to use DepT
, we can apply the advice, because advice only
works on functions that end on a DepT
action. Also, advice might depend on
the full gamut of functionality stored in the environment.
What about Dep.ReaderAdvice
and Dep.IOAdvice
?
Advice
s from Dep.Advice
require us to work with DepT
, but DepT
is kind
of weird. Instead, we may might to use more common monads for our effects, like
ReaderT
o plain old IO
.
That's why Dep.ReaderAdvice
and Dep.IOAdvice
exist: they provide alternative
versions of the Advice
type which work with those monads.
Ain't that a lot of code duplication? Why not have a single Advice
type which
works with all monads? That leads us to...
What about Dep.SimpleAdvice
?
Dep.SimpleAdvice
provides a version of the Advice
type that can be used with
different concrete monads like ReaderT
or IO
.
See this
thread
in the Haskell Discourse for more info.
There's a catch, however. Dep.SimpleAdvice
depends on the coerce
mechanism
of Haskell, and it can sometimes be finicky, for example when some required
constructor hasn't been imported, or when there are polymorphic functions
involved.
That's the reason Dep.ReaderAdvice
and Dep.IOAdvice
are still necessary. For
their particular monads, they work in more cases.
Historical aside
According to Wikipedia, the term "advice" in the sense of aspect-oriented
programming goes back to 1966. Quoting from PILOT: A Step Toward Man-Computer
Symbiosis:
There are two ways a user can modify programs in this subjective model of
programming: he can modify the interface between procedures, or he can modify
the procedure itself. (Since procedures are themselves made up of procedures,
modifying a procedure at one level may correspond to modifying the interface
between procedures at a lower level.) Modifying the interface between
procedures is called advising. Modifying a procedure itself is editing.
Advising is the basic innovation in the model, and in the PILOT system.
Advising consists of inserting new procedures at any or all of the entry or
exit points to a particular procedure (or class of procedures). The
procedures inserted are called "advice procedures" or simply "advice".
Since each piece of advice is itself a procedure, it has its own entries and
exits. In particular, this means that the execution of advice can cause the
procedure that it modifies to be bypassed completely, e.g., by specifying as
an exit from the advice one of the exits from the original procedure; or the
advice may change essential variables and continue with the computation so
that the original procedure is executed, but with modified variables.
Finally, the advice may not alter the execution or affect the original
procedure at all, e.g., it may merely perform some additional computation
such as printing a message or recording history. Since advice can be
conditional, the decision as to what is to be done can depend on the results
of the computation up to that point.
The principal advantage of advising is that the user need not be concerned
about the details of the actual changes in his program, nor the internal
representation of advice. He can treat the procedure to be advised as a
unit, a single block, and make changes to it without concern for the
particulars of this block. This may be contrasted with editing in which the
programmer must be cognizant of the internal structure of the procedure.
tweet.
Links