Awesome
microlens
A tiny part of the lens library with no dependencies.
If you're completely new to this whole lenses thing
Read this tutorial. It's for lens, but it applies to microlens just as well (except for module names).
What is microlens?
microlens is a lens library, just like lens, but smaller. It provides
essential lenses and traversals (like _1
and _Just
), as well as ones
which are simply nice to have (like each
, at
, and ix
), and some
combinators (like failing
and singular
), but everything else is
stripped. As the result, microlens has no dependencies. However, there are
also separate packages (microlens-ghc and microlens-platform) which
provide additional instances and let you use each
and friends with various
container types.
If you're writing an app, you should probably use microlens-platform
and not microlens. You'll get additional functions, instances,
makeLenses
, and other useful things. microlens is mostly for library
writers and for toying with lenses. See Which microlens library should I use? for more.
Here are the build times for all libraries in the family:
Package | Build time with dependencies | Pure build time |
---|---|---|
microlens | 3.5s | 3.5s |
microlens-th | 7.2s | 4.5s |
microlens-ghc | 5.7s | 3.3s |
microlens-mtl | 8.8s | 3.7s |
microlens-platform | 1m47s | 4.9s |
microlens-contra | 1m12s | 2.1s |
microlens-aeson | 3m47s | 9.2s |
microlens-process | 9.8s | 3.8s |
lens | 4m10s | 1m12s |
Other features:
-
Nicer documentation.
-
Compatibility with lens. If you want to define a
Lens
or aTraversal
in your package, you can depend on this package without fear. -
No awkward renamed functions or any of such nonsense. You can at any moment replace
Lens.Micro
withControl.Lens
and get the full power of lens. There are also no unique to microlens functions which you would have to rewrite when switching to lens (even though I was tempted to add some). -
No Template Haskell dependency. There is a separate package for generating (lens-compatible) record lenses, which is called microlens-th.
-
All
INLINE
pragmas sprinkled through lens were preserved. Performance shouldn't suffer; if it does, it's a bug.
The reason microlens exists is that lens is a huge library with lots of dependencies, but lenses are very useful and it's not nice to limit them to applications and bigger packages. (I'm not talking about exporting lenses, I'm talking about using lenses to write code.) microlens attempts to be a library that would be a nearly unquestionable win for some people.
Which microlens library should I use?
If you don't know where to start, start with microlens-platform and change later if necessary.
If you need Prism
s or Iso
s, use microlens-pro instead.
If you're a library author or otherwise want a low dependency footprint, start with microlens and only add others when needed (likely you will want microlens-th to conveniently generate lenses for your datatypes).
Migrate from lens
-
If you use
ALens
, indexed traversals, orWrapped
, you won't be able to migrate (although some indexed functions are available elsewhere – containers and vector provide them, and ilist provides indexed functions for lists). -
If you have your own instances of
Each
,At
,Ix
,Zoomed
, orField*
, and you don't export them, it's okay. Otherwise you should keep using lens, since those classes are incompatible with classes defined in lens. Similarly, if you export any functions withAt
/Zoom
/etc constraints, don't migrate. -
If you export
Getter
s orFold
s, you would have to use microlens-contra for full compatibility, and it has more heavy dependencies (but still much less heavy than lens). “Full compatibility” here means that some lens functions (such astakingWhile
) don't work withSimpleGetter
andSimpleFold
available from the main microlens package. -
In the very rare case of using
makeLensesWith
and havinggenerateUpdateableOptics
disabled, you'd have to applyfromSimpleFold
andfromSimpleGetter
to folds/getters you export. Same with fields that have aforall.
in them.
Otherwise, everything should work just fine without any code changes – the microlens API mirrors the lens API. The license is the same, too.
(The list might look big, but in reality it isn't and in the majority of cases you'll be able to migrate just fine. “If it compiles and you didn't have to change any type signatures, it works.”)
If you're unsure, just open an issue in your project, mention me (@neongreen), and I'll look at your code and tell you whether it'll work or not.
All packages in the family
- microlens – all basic functionality, plus
each
/at
/ix
- microlens-mtl –
+=
and friends,use
,zoom
/magnify
- microlens-th –
makeLenses
andmakeFields
- microlens-ghc – everything in microlens + instances to make
each
/at
/ix
usable with arrays,ByteString
, and containers - microlens-platform – microlens-ghc + microlens-mtl + microlens-th +
instances for
Text
,Vector
, andHashMap
- microlens-pro –
Prism
s andIso
s along with related definitions (e.g.only
,non
,review
, etc.) - microlens-contra –
Fold
andGetter
that are copies of types in lens (the reason they're in a separate library is that those types depend on contravariant)
Unofficial:
- microlens-aeson – a port of lens-aeson
- microlens-process - a port of lens-process
If you're writing a library, use microlens and other packages as needed; if you're writing an application, perhaps use microlens-platform.
Versions of microlens-ghc and microlens-platform are incremented whenever versions of their dependencies are incremented, so if you're using these packages it's always enough to specify just their versions and nothing else. In other words, there's no risk of the following happening:
- a new version of microlens is released, with several functions removed
- version of microlens-platform stays the same
- your code silently stops compiling as the result
Competitors
-
basic-lens – the smallest library ever, containing only
Lens
,view
,set
, andover
(and no lenses whatsoever). Uses only 1 extension –RankNTypes
– and thus can be used with e.g. JHC and really old GHCs. -
reasonable-lens – a bigger library which has
Lens
, some utilities (likeview
,use
,+=
),makeLenses
even, but little else – no lenses (except for_1
..._4
), noTraversal
, no documentation. Overall it looks like something slapped together in a hurry by someone who simply needed to get rid of a lens dependency in one of nir projects. -
lens-simple – a single module reexporting parts of lens-family. It's the most feature-complete library on this list, with both
Lens
andTraversal
available, as well as a number of lenses, traversals, and utilities. However, it has some annoyances – noeach
,_1
and_2
work only on pairs,ix
doesn't work on lists or arrays and is thus useless,at
only works onMap
, etc. I don't think these will ever be fixed, as they require defining some ad-hoc typeclasses, and the current absence of any such typeclasses in lens-family seems to suggest that the authors consider it a bad idea. -
data-lens-light – a library which uses a different formulation of lenses and is thus incompatible with lens (it uses different names, too). Doesn't actually provide any lenses.
So, I recommend:
-
lens-simple if you specifically want a library with a clean, understandable implementation, even if it's sometimes more cumbersome to use and can be a bit slower.
-
lens-family if you like lens-simple but don't want the Template Haskell dependency.
-
microlens otherwise.
What's bad about this package
I hate it when people advertise things without also describing their disadvantages, so I'll list the ones I can think of here.
-
No indexed traversals.
-
This package doesn't actually let you write everything full lens-style (for instance, there are few operators, myriads of functions generalised for lenses by adding the
Of
suffix aren't included, etc). On the other hand, I guess some would actually consider it an advantage. Anyway, if you want to use lens as a language instead of as a tool, you probably can afford depending on the full package. -
There are orphan instances, e.g. in the microlens-ghc package. (However, the only way someone can actually break things is by using
Lens.Micro.Internal
and ignoring the warnings there, so I think it's not a huge danger.) -
There are
set
andover
in the basic module (i.e.Lens.Micro
), butview
lives inLens.Micro.Extras
and it doesn't work inMonadReader
(and the version that does is in microlens-mtl). -
makeLenses
can generateSimpleFold
andSimpleGetter
which are sli-ightly less general thatFold
andGetter
in lens. (If you're a lens user, you still can convert from those versions to fully general versions, so you're not doomed or anything – it's just a minor nuisance / opportunity for confusion. Also, microlens-contra provides trueFold
andGetter
.) -
The implementation is as cryptic/complicated as lens's (performance has its costs).
Design decisions
microlens doesn't include anything lens doesn't include, even though sometimes I'm very tempted to improve something in microlens just because I have control over it.
I don't mind adding new functions from lens to the package,
even when done in an inconsistent way (e.g. I added mapAccumLOf
just
because someone needed it, but I haven't added mapAccumROf
even though
that would've been more consistent). However, I am only able to add
functions as long as microlens stays small, so if you plan to adopt
microlens first and make dozens of requests for function additions later,
this package is not for you.
Most *Of
functions aren't included. If you don't know, those are sumOf
,
lengthOf
, setOf
, etc., and they are roughly equivalent to following:
sumOf l s = sum (s ^.. l)
lengthOf l s = length (s ^.. l)
setOf l s = Set.fromList (s ^.. l)
(Where ^..
takes something which extracts several targets, and returns a
list of those targets. E.g. (1, 2) ^.. both
is [1, 2]
).
I guess the reason for including them all into lens
(and there's an awful
lot of them) is somewhere between
- “they are faster than going through intermediate lists”
- “there are some rare cases when you can use a SomeSpecialisedMonoid but
can't use
Endo [a]
” - “it's nice to be able to say
sumOf (each._1) [(1,"x"),(2,"y")]
instead of clumsysum . (^.. each._1) $ [(1,"x"),(2,"y")]
”
I suspect that the last reason is the most important one. The last reason is also the one I dislike most.
There are lots of functions which work on lists; lists are something like
“the basic collection/stream type” in Haskell. GHC tries a lot to optimise
code which produces and consumes lists; admittedly, it doesn't always
succeed. lens
seems to be trying to sidestep this whole list machinery.
-
With lists: one function traverses something and extracts a list of results, another function does something to those results.
-
With lenses: one function traverses something and takes another function as a parameter (to know what to do with results). Note that here
each._1
is the traversing function; it seems likesumOf
takes it as a parameter, but in realitysumOf
merely gives “summation” as the parameter to the traversing function.
The latter way is theoretically nicer, but not when you've got the rest of
huge ecosystem using lists as the preferred way of information flow,
otherwise you're bound to keep rewriting all functions and adding Of
to
them. lens
is good for creating functions which extract data, and for
creating functions which update structures (nested records, etc.), but it's
probably not good enough to make the whole world want to switch to writing
lens-compatible consumers of data.
Nothing indexed is included since it's impossible to
get Conjoined
without adding a pile of dependencies:
class ( Choice p, Corepresentable p
, Comonad (Corep p), Traversable (Corep p)
, Strong p, Representable p
, Monad (Rep p), MonadFix (Rep p)
, Distributive (Rep p)
, ArrowLoop p, ArrowApply p, ArrowChoice p
)
=> Conjoined p
class Conjoined p => Indexable i p
Instances of Ixed
, Each
, At
, etc are all split off into separate
packages, which is understandable, because otherwise we'd have to have
vector as a dependency (the alternative is having orphan instances,
which I'm not particularly afraid of). However, even instances for libraries
shipped with GHC (such as array) are in their own
package. There are 2 reasons for this:
- I really want to be able to say “this library has no dependencies”.
- All those instances actually take quite some time to build (for the same reason not all instances for tuples are included in the main package).
What about lens-family?
lens-family is another small lenses library which is mostly compatible
with lens (unless I decide to nitpick and say that its makeLensesBy
and
intAt
aren't present in lens at all), which has few dependencies, and
which provides Template Haskell in a separate package as well.
It looks like lens-family values cleanness and simplicity, which
unfortunately means that it might've been hard for me (if possible at all)
to convince its maintainer to make changes which would bring it closer to
lens (INLINE
pragmas, using unsafe #.
operator, adding each
, etc). I
actually like cleanness and dislike excessive optimisation (especially of
the kind that is used in lens) too, but making a library I would like
wasn't my goal. The goal was to push people who aren't using a lens library
towards using one.