What Are Lenses For?
What Are Lenses?
First we need to do a little (or more) scene setting, if we want to represent a bunch of achievements for a bunch of users in some games a super simple representation of that in JavaScript might involve some code like this:
users[userIDToUpdate].games['bestgameever']
.achievements['completedthegame']
.completedDate = new Date()
This code sets the completed date for the achievement 'completedthegame'
for the game 'bestgameever'
for the user with some identifier userIDToUpdate
. There’s two kinds of access going on here, that of known fields like .games
and .achievements
, as well as the object key indexing done with square brackets which might address values that don’t exist (for example ['bestgameever']
). That’s problem no. 1 though, if any of these ids are wrong the code blows out with a nasty looking exception.
This becomes a lot harder if the goal is to immutably update these values:
const user = users[userIDToUpdate]
const game = user.games['bestgameever']
const achievement = game.achievements['completedthegame']
const updatedAchievement = {
...achievement,
: new Date()
completedDate
}const updatedGame = {
...game,
: {
achievements...game.achievements,
'completedthegame']: updatedAchievement
[
}
}const updatedUser = {
...user,
: {
games...user.games,
'bestgameever']: updatedGame
[
}
}const updatedUsers = {
...users,
: updatedUser
[userIDToUpdate] }
We’ve still got the problem of invalid values throwing exceptions but also about 10 times the code.
Lets move this to Haskell where we still have the same kind of problem.
Here’s some setup for language options and imports upfront:
{-# LANGUAGE TypeApplications #-}
{-# LANGUAGE DataKinds #-}
{-# LANGUAGE DeriveGeneric #-}
{-# LANGUAGE GADTs #-}
{-# LANGUAGE FlexibleContexts #-}
import GHC.Generics
import qualified Data.HashMap.Strict as M
import Data.Time.Clock
import Control.Lens
import Data.Generics.Product
import Data.Generics.Sum
Our basic data types to represent everything:
data Achievement = Achievement { completedDate :: Maybe UTCTime } deriving (Eq, Show, Generic)
data Game = Game { achievements :: M.HashMap String Achievement } deriving (Eq, Show, Generic)
data User = User { games :: M.HashMap String Game } deriving (Eq, Show, Generic)
type Users = M.HashMap String User
Now the starting values:
= Achievement Nothing
defaultAchievement = Game (M.singleton "completedthegame" defaultAchievement)
defaultGame = User (M.singleton "bestgameever" defaultGame)
defaultUser = M.singleton "sean" defaultUser defaultUsers
Updating this is a whole mess of pain:
=
updateAchievement completedAt achievement = Just completedAt }
achievement { completedDate = game { achievements =
updateGame completedAt game "completedthegame" (achievements game) }
M.adjust (updateAchievement completedAt) = user { games =
updateUser completedAt user "bestgameever" (games user) }
M.adjust (updateGame completedAt) =
getManuallyUpdatedUsers completedAt "sean" defaultUsers M.adjust (updateUser completedAt)
That’s quite verbose and just not very pleasant to read, it wont throw any exceptions if a value is missing however and will instead return the original value. Which is a little bit better but not that great overall.
With lenses though it’s possible to split apart the “targeting” from the operation being performed, first we create a lens which handles the former:
= at "sean"
updateCompletedDateLens . _Just
. field @"games"
. at "bestgameever"
. _Just
. field @"achievements"
. at "completedthegame"
. _Just
. field @"completedDate"
In this case we’re chaining lenses and there’s a lot of clarity as to what is going on, at
is used to peer into the HashMap
values and field
peers into the the fields of the records.
To update this we use one of the utility functions from the lens
library itself:
= set updateCompletedDateLens (Just completedAt) defaultUsers getLensUpdatedUsers completedAt
This snippet runs the Haskell of this post:
= do
main print defaultUsers
<- getCurrentTime
completedAt print $ getManuallyUpdatedUsers completedAt
print $ getLensUpdatedUsers completedAt
Update Everything
The Problem.
In the project I’m working on there is some logic which assigns a UUID to a field in a bunch of values which make up an AST structure representing a user’s code. For testing purposes there are a series of functions which take a value in that structure and wipe out the unique ID so that we can use an equality check against it. Those are manually written functions which we have to update as the structure changes, leading to the odd “Wait, why has that still got a unique ID?” moment when we make changes. I had an inkling that if this was written in Haskell (the code I’m referring to is written in TypeScript) this manually written code, along with the maintenance cost, would be unnecessary.
First Steps.
I’ll be using nix-shell
to get myself a REPL to work with the code, using this command:
nix-shell
-p "haskellPackages.ghcWithPackages (p: with p; [ghc uniplate pretty-simple])"
--command ghci
Turn on an option that we’ll be needing for an easy life:
GHCi> :set -XDeriveDataTypeable
Add an import related to that option, don’t worry about these for now:
GHCi> import Data.Data
Next we’ll use the multi-line input support in GHCi to input this (put :{
on an empty line first and on an empty line after write :}
to finish the multi-line editing):
newtype UniqueID = UniqueID { uid :: String }
deriving (Eq, Show, Data)
data JSObject = JSObject { objectParts :: [(String, JSValue)] }
deriving (Eq, Show, Data)
data JSArray = JSArray { arrayParts :: [JSValue] }
deriving (Eq, Show, Data)
data JSString = JSString { stringValue :: String }
deriving (Eq, Show, Data)
data JSBool = JSBool { boolValue :: Bool }
deriving (Eq, Show, Data)
data JSNumber = JSNumber { numberValue :: Double }
deriving (Eq, Show, Data)
data JSNull = JSNull deriving (Eq, Show, Data)
data JSUndefined = JSUndefined deriving (Eq, Show, Data)
data JSCode = JSCode { code :: String, codeUniqueID :: UniqueID }
deriving (Eq, Show, Data)
data JSValue = JSValueObject JSObject
| JSValueArray JSArray
| JSValueString JSString
| JSValueBool JSBool
| JSValueNumber JSNumber
| JSValueNull JSNull
| JSValueUndefined JSUndefined
| JSValueCode JSCode
deriving (Eq, Show, Data)
What we have above is something that approximates the definition of JSON in most libraries but with the addition of the JSCode
type which represents a chunk of arbitrary JavaScript and is where we hold our UniqueID
value. As we can have one of those buried 9 layers down, we need to walk the tree to update those wherever we might find them.
Building The Parts.
We’ll start with a function to clear unique ID values, which just throws away the original value and gives us an empty one (use the multi-line support again for this):
clearUniqueID :: UniqueID -> UniqueID
= UniqueID "" clearUniqueID _
Now build up a value to play with:
GHCi> exampleCode = JSCode "5 + 10" (UniqueID "ABC")
GHCi> exampleArray = JSArray [JSValueCode exampleCode]
GHCi> exampleString = JSString "Good News"
GHCi> exampleFirstValue = ("first", JSValueString exampleString)
GHCi> exampleSecondValue = ("second", JSValueArray exampleArray)
GHCi> exampleObject = JSValueObject (JSObject [exampleFirstValue, exampleSecondValue])
We can use the pretty-simple
package to see the structure more clearly:
GHCi> import Text.Pretty.Simple
GHCi> pPrintNoColor exampleObject
JSValueObject
JSObject
( =
{ objectParts
["first"
( JSValueString
, JSString { stringValue = "Good News" } )
(
)
,"second"
( JSValueArray
, JSArray
( =
{ arrayParts JSValueCode
[ JSCode
( = "5 + 10"
{ code = UniqueID { uid = "ABC" }
, codeUniqueID
}
)
]
}
)
)
]
} )
The Good Stuff.
With the Uniplate library we just need a couple of imports:
GHCi> import Data.Generics.Uniplate.Data
GHCi> import Data.Generics.SYB
Then with the everywhere
function we apply our transformation from earlier:
GHCi> updatedExample = everywhere clearUniqueID exampleObject
Lets see the result.
GHCi> pPrintNoColor updatedExample
JSValueObject
JSObject
( =
{ objectParts
["first"
( JSValueString
, JSString { stringValue = "Good News" } )
(
)
,"second"
( JSValueArray
, JSArray
( =
{ arrayParts JSValueCode
[ JSCode
( = "5 + 10"
{ code = UniqueID { uid = "" }
, codeUniqueID
}
)
]
}
)
)
]
} )
…that’s it! All done, see you next time!
A Deeper Look.
First off, lets have a look at the type of that everywhere
function, because that was the most mysterious part.
everywhere :: Biplate b a => (a -> a) -> b -> b
So our clearUniqueID
function slotted into the first parameter and then we get a function that transforms from JSValue
to JSValue
. But this is only if we have an instance of Biplate JSValue UniqueID
.
Looking up Biplate
we want a Biplate JSValue UniqueID
, there’s an instance for (Data a, Data b, Uniplate b) => Biplate a b
. So a Data JSValue
, a Data UniqueID
and a Uniplate UniqueID
gives us that Biplate JSValue UniqueID
.
Following that chain along to Uniplate
, if we look at the instances for it we can see there’s this one: Data a => Uniplate a
. So Data a
gives us a Uniplate a
for every Data a
that exists.
Since everything ends up with Data JSValue
and Data UniqueID
, which GHC has graciously derived automatically for us, we don’t have to write all of that code for walking the various types.
So What Does This Give Us?
The benefits of this are as follows:
- No manually written pile of code that needs regular maintenance.
- Reflection or similar runtime introspection is avoided, which might drill into the wrong thing and possibly throw an exception.
- The types guide the behaviour, so we can guarantee that it’ll be applied to everything it should be and nothing else.
That last point is part of a bigger pattern, not using types as validation but as a building block for behaviour. Having the compiler do work which is repetitive and/or error prone and to the first point which then needs keeping up to date after it is first implemented.
Bonus Round.
There are other wonders to be found in the uniplate
library too, like childrenBi
which will get all the values matching a particular type from the hierarchy.
GHCi> childrenBi exampleObject :: [UniqueID]
[UniqueID "ABC"]
Maybe Versus Nullable
Select All The Things
Recently I was faced with the task of updating our product to support multiple selection, so that for example a user can move multiple elements together with one mouse drag. As the codebase is written in TypeScript it meant taking this field:
: ThingTarget | null; selectedThing
and changing it to this:
: Array<ThingTarget>; selectedThings
On the surface this isn’t a huge change, but the code operating on these values has to change quite dramatically because of their different concrete structures. Taking the display of the indicators that show something is selected in the designer as an example:
if (selectedThing == null) {
return [buildSelectedControl(selectedThing)];
else {
} return [];
}
Now becomes:
return selectedThings.map(buildSelectedControl);
Make All The Changes
These are the simplest examples but fundamentally the logic always needed changing along these lines. As this felt a little unsatisfactory to me I pondered how this would’ve panned out with Haskell (or PureScript/Scala/etc). So we would’ve started with code that looked like:
selectedThing :: Maybe ThingTarget
That would’ve turned into:
selectedThings :: [ThingTarget]
However, in either of these cases the logic is the same and looks like:
fmap buildSelectedControl selectedThings
-- The type of fmap being:
-- fmap :: Functor f => (a -> b) -> f a -> f b
Similarly to check if something has been selected we’d use:
elem thing selectedThings
-- The type of elem being:
-- elem :: (Foldable t, Eq a) => a -> t a -> Bool
When people talk of abstraction it’s often in terms of things close to concrete use cases like a SelectionManager
(let’s ignore that nonsense name) class which then doesn’t end up being very re-usable. However with abstractions like Foldable
or Functor
we have much less specific but just as if not more strict ones which tend more towards general purpose use cases. But also because the language lacks as many special case features (like null
and undefined
are in JavaScript, even if we don’t think of them traditionally as such) there are fewer hitches to the user of this kind of abstraction.
One Step Further
This extends beyond the functions that can be used to the functions that can be built too:
selectedControls :: (Functor f) = (Editor -> f Thing) -> Editor -> f Control
= fmap buildSelectedControl $ getSelected editor selectedControls getSelected editor
In this case the function is as general as fmap
allows, if the accessor function has a type of Editor -> Maybe Thing
, then the above function returns a Maybe Control
. If no type was supplied, this is what GHC would determined the type of selectedControls
would be.
GHCi> data Thing = Thing
GHCi> data Control = Control
GHCi> data Editor = Editor
GHCi> :{
Prelude| buildSelectedControl :: Thing -> Control
Prelude| buildSelectedControl _ = Control
Prelude| :}
GHCi> selectedControls getSelected editor = fmap buildSelectedControl $ getSelected editor
GHCi> :t selectedControls
selectedControls :: Functor f => (t -> f Thing) -> t -> f Control
It turns out this is more flexible that the version further above, as there’s no need (or even way) to constrain the function to the Editor
type the compiler makes it polymorphic to any value t
.
So What?
I posit that nullable values interfere with code reuse as they’re a disjoint union of the value type and null
(undefined
as well), but none of the common abstractions can be applied to a disjoint union. Simple types and simple abstractions lead to code reuse and less destructive refactors and if used, more flexible implementations.