Awesome
New: The Clojure/north 2020 talk is up!
<img src="https://img.youtube.com/vi/If3GT8zSHfE/maxresdefault.jpg" width="40%">
Methodical
Methodical is a library that provides drop-in replacements for Clojure multimethods and adds several advanced features.
(require '[methodical.core :as m])
(m/defmulti my-multimethod
:type)
(m/defmethod my-multimethod Object
[m]
(assoc m :object? true))
(my-multimethod {:type Object})
;; -> {:type java.lang.Object, :object? true}
Calling the next-most-specific method with next-method
Inspired by the Common Lisp Object System (CLOS), Methodical methods can call the next-most-specific method, if they
should so desire, by calling next-method
:
(m/defmethod my-multimethod String
[m]
(next-method (assoc m :string? true)))
(my-multimethod {:type String})
;; -> {:type java.lang.String, :string? true, :object? true}
This makes it easy to reuse shared parent implementations of methods without having to know the exact dispatch value of the next method. In vanilla Clojure multimethods, you'd have to do something like this:
((get-method my-multimethod Object) (assoc m :string? true))
If you're not sure whether a next-method
exists, you can check whether it's nil
before calling it.
Methodical exports custom clj-kondo configuration and hooks for defmulti
and defmethod
; with the exported configuration it will even tell you if you call next-method
with the wrong number
of args:
Auxiliary Methods: :before
, :after
, and :around
Inspired by the CLOS, Methodical multimethods support both primary methods and auxiliary methods. Primary methods
are the main methods that are invoked for a given dispatch value, such as the implementations for String
or Object
in the examples above; they are the same type of method vanilla defmethod
supports. Auxiliary methods are additional
methods that are invoked :before
, :after
, or :around
the primary methods:
(m/defmethod my-multimethod :before String
[m]
(assoc m :before? true))
(m/defmethod my-multimethod :around String
[m]
(next-method (assoc m :around? true)))
(my-multimethod {:type String})
;; -> {:type java.lang.String, :around? true, :before? true, :string? true, :object? true}
:before
methods
All applicable :before
methods are invoked before the primary method, in order from most-specific (String before
Object) to least-specific. Unlike the CLOS, which ignores the results of :before
and :after
auxiliary methods, by
default Methodical threads the result of each :before
method into the next method as its last argument. This better
supports Clojure's functional programming style.
(m/defmulti before-example
(fn [x acc]
(:type x)))
(m/defmethod before-example :before String
[x acc]
(conj acc :string))
(m/defmethod before-example :before Object
[x acc]
(conj acc :object))
(m/defmethod before-example :default
[x acc]
(conj acc :default))
(before-example {:type String} [])
;; -> [:string :object :default]
:before
methods unlock a whole new range of solutions that would be tedious with vanilla Clojure multimethods:
suppose you wanted add logging to all invocations of a multimethod. With vanilla multimethods, you'd have to add an
individual log statement to every method! With Methodical, just add a new :default
:before
method:
(m/defmethod my-multimethod :before :default
[& args]
(log/debugf "my-multimethod called with args: %s" args)
;; return last arg so it is threaded thru for next method
(last args))
:after
methods
All applicable :after
methods are invoked after the primary method, in order from least-specific (Object before
String) to most-specific. Like :before
methods, (by default) the result of the previous method is threaded thru as
the last argument of the next function:
(m/defmulti after-example
(fn [x acc]
(:type x)))
(m/defmethod after-example :after String
[x acc]
(conj acc :string))
(m/defmethod after-example :after Object
[x acc]
(conj acc :object))
(m/defmethod after-example :default
[x acc]
(conj acc :default))
(after-example {:type String} [])
;; -> [:default :object :string]
An example usecase for :after
is chess. When looking up legal moves, you
implement how each piece can move, then in :after :default
limit it to only
spaces on the board.
:around
methods
:around
methods are called around all other methods and give you the power to choose how or when to invoke those
methods, and modify any arguments passed to them, or their result, as needed. Like primary methods (but unlike
:before
and :after
methods), :around
methods have an implicit next-method
argument; you'll need to call this to invoke
the next method. :around
methods are invoked from least-specific to most-specific (Object before String):
(m/defmulti around-example
(fn [x acc]
(:type x)))
(m/defmethod around-example :around String
[x acc]
(as-> acc acc
(conj acc :string-before)
(next-method x acc)
(conj acc :string-after)))
(m/defmethod around-example :around Object
[x acc]
(as-> acc acc
(conj acc :object-before)
(next-method x acc)
(conj acc :object-after)))
(m/defmethod around-example :default
[x acc]
(conj acc :default))
(around-example {:type String} [])
;; -> [:object-before :string-before :default :string-after :object-after]
Around methods give you amazing power: you can decider whether to skip invoking next-method
altogether, or even
invoke it more than once; you can acquire resources for the duration of the method invocation with with-open
or the
like.
Method combinations are discussed more in detail below.
Defining multiple auxiliary methods for the same dispatch value
Unlike primary methods, you can have multiple auxiliary methods for the same dispatch value. However, adding an
additional duplicate auxiliary method every time you reload a namespace would be annoying, so the defmethod
macro
automatically replaces existing auxiliary methods for the same multimethod and dispatch value in the same namespace:
(m/defmulti after-example
(fn [x acc]
(:type x)))
(m/defmethod after-example :after String
[x acc]
(conj acc :string))
;; replaces the aux method above
(m/defmethod after-example :after String
[x acc]
(conj acc :string-2))
(m/defmethod after-example :default
[x acc]
(conj acc :default))
(after-example {:type String} [])
;; -> [:default :string-2]
In most cases, this is what you want, and the least-annoying behavior. If you actually do want to define multiple aux methods of the same type for the same multimethod and dispatch value, you can give each method a unique key:
(m/defmulti after-example
(fn [x acc]
(:type x)))
(m/defmethod after-example :after String "first String :after method"
[x acc]
(conj acc :string))
(m/defmethod after-example :after String "another String :after method"
[x acc]
(conj acc :string-2))
(m/defmethod after-example :default
[x acc]
(conj acc :default))
(after-example {:type String} [])
;; -> [:default :string-2 :string]
You can also use this key to remove specific auxiliary methods.
Getting the "effective method"
The effective method is the method that is ultimately invoked when you invoke a multimethod for a given dispatch
value. With vanilla Clojure multimethods, get-method
returns this "effective method" (which is nothing more than a
single function); in Methodical, you can use effective-method
to build an effective method that combines all auxiliary
methods and primary methods into a single composed function. By default, this effective method is cached.
Constructing and composing multimethods programmatically
Perhaps one of the biggest limitations of vanilla multimethods is that they can't be passed around and modified
on-the-fly like normal functions or other Clojure datatypes -- they're defined statically by defmulti
, and methods
can only be added destructively, by altering the original object. Methodical multimethods are implemented entirely as
immutable Clojure objects (with the exception of caching).
(let [dispatch-fn :type
multifn (-> (m/default-multifn dispatch-fn)
(m/add-primary-method Object (fn [next-method m]
:object)))
multifn' (m/add-primary-method multifn String (fn [next-method m]
:string))]
((juxt multifn multifn') {:type String}))
;; -> [:object :string]
Note that when using these programmatic functions, primary and :around
methods are each passed an implicit
next-method
arg as their first arg. The defmethod
macro binds this automatically, but you'll need to handle it
yourself when using these functions.
Every operation available for Clojure multimethods, and quite a few more, are available with programmatic functions like
add-primary-method
.
Advanced Customization
Clojure's multimethods, while quite powerful, are somewhat limited in the ways you can customize their behavior. Here's a quick list of some of the things you can do with Methodical multimethods, all of which are simply impossible with vanilla Clojure mulitmethods:
-
Dispatch with multiple hierarchies (e.g., one for each arg)
-
Change the strategy used to cache effective methods (the compiled function that is ultimately invoked for a set of args)
-
Invoke all applicable primary methods, and return a sequence of their results
-
Dynamically compute new primary or auxiliary methods without users manually adding them
-
Support default values for part of a dispatch value, e.g. when dispatching off a pair of classes, e.g. support
[String String]
,[:default String]
, or[String :default]
-
Combine multiple multimethods into a single multimethod; that, when invoked, tries invoking each constituent multimethod in turn until it finds one with a matching method implementation
To enable such advanced functionality, Methodical multimethods are divided into four components, and two that manage them:
-
The method combination, which defines the way applicable primary and auxiliary methods are combined into a single effective method. The default method combination,
thread-last-method-combination
, binds implicitnext-method
args for primary and:around
methods, and implements logic to thread the result of each method into the last argument of the next. Method combinations also specify which auxiliary method qualifiers (e.g.:before
or:around
) are allowed, and howdefmethod
macro forms using those qualifiers are expanded (e.g., whether they get an implicitnext-method
arg). Method combinations implement theMethodCombination
interface. -
The method table stores primary and auxiliary methods, and returns them when asked. The default implementation,
standard-method-table
, uses simple Clojure immutable maps, but there is nothing stopping you from creating an implementation that ignores requests to store new methods, or dynamically generates and returns a set of methods based on outside factors. Method tables implement theMethodTable
interface. -
The dispatcher decides which dispatch value should be used for a given set of arguments, which primary and auxiliary methods from the method table are applicable for that dispatch value, and the order those methods should be applied in -- which methods are most specific, and which are the least specific (e.g.,
String
is more-specific thanObject
.) The default implementation,multi-default-dispatcher
, mostly mimics the behavior of Clojure multimethods, using a dispatch function to determine dispatch values, and a single hierarchy andprefers
map to determine which methods are applicable, but supports partial-default methods, e.g,[:default String]
. (See this blog post for more information about partial-default dispatch.) You could easily create your own implementation that uses multiple hierarchies, or one that uses no hierarchies at all. Dispatchers implement theDispatcher
interface. -
A cache, if present, implements a caching strategy for effective methods, so that they need not be recomputed on every invocation. Caches implement the
Cache
interface. Depending on whether you create a multimethod viadefmulti
or with the programmatic functions, the cache is either awatching-cache
, which watches the hierarchy referenced by the dispatcher (by default,#'clojure.core/global-hierarchy
), clearing the cache when it changes; orsimple-cache
, a bare-bones cache. You could easily implement alternative caching strategies, such as TTL or LRU caches, or ones that better optimize memory and locality.
The method combination, method table, and dispatcher are managed by an object called the multifn impl, which
implements MultiFnImpl
. If this impl supports caching, it manages a cache as well, albeit indirectly (thru its
implementation of the method effective-method
.) The default implementation is actually a combination of two multifn
impls: cached-multifn-impl
manages a cache and wraps standard-multifn-impl
, which itself retains the other three
components.
Finally, the multifn impl is wrapped in StandardMultiFn
, which implements a variety of interfaces, such as
clojure.lang.IObj
, clojure.lang.Named
, clojure.lang.IFn
, as well as MethodCombination
, MethodTable
,
Dispatcher
, and MultiFnImpl
.
You can use alternative components directly in the defmulti
macro by passing :combo
, :method-table
,
dispatcher
, or :cache
:
(m/defmulti custom-multifn
some-dispatch-fn
:combo (m/thread-first-method-combination))
When constructing multimethods programmatically, you can use standard-multifn-impl
and multifn
to create a multimethod with the desired
combination of components:
(m/multifn
(m/standard-multifn-impl
(m/thread-last-method-combination)
(m/standard-dispatcher some-dispatch-fn)
(m/standard-method-table))
nil
(m/simple-cache))
Component implementations that ship with Methodical
As previously mentioned, Methodical ships with a variety of alternative implementations of these constituent components of multimethods. The following summarizes all component implementations that currently ship with Methodical:
Method Combinations
-
clojure-method-combination
- mimics behavior of vanilla Clojure multimethods. Disallows auxiliary methods; primary methods do not get an implicitnext-method
arg. -
clos-method-combination
- mimics behavior of the CLOS standard method combination. Supports:before
,:after
, and:around
auxiliary methods. Return values of:before
and:after
methods are ignored.:after
methods are only called with the result of the primary method. Primary and:around
methods are given an implicitnext-method
argument. -
thread-last-method-combination
: the default method combination. Similar toclos-method-combination
, but the result of:before
methods, the primary method, and:after
methods are threaded thru to the next method as the last argument.:after
methods are passed the full set of arguments the multimethod as a whole was invoked with. -
thread-first-method-combination
: Likethread-last-method-combination
, but results of each method are threaded into the next method as its first arg. -
Operator method combinations. The following method combinations are inspired by CLOS operator method combinations; each combination behaves similarly, in that it invokes all applicable primary methods, from most-specific to least-specific (String before Object), combining results with the operator for which they are named. Generally, the result is of this form:
(operator (primary-method-1 args) (primary-method-2 args) (primary-method-3 args)))
Operator method combinations support
:around
methods, but not:before
or:after
; primary methods do not supportnext-method
, but:around
methods do.The following operator method combinations ship with Methodical:
-
do-method-combination
-- executes all primary methods sequentially, as if bydo
, returning the result of the least-specific method. The classic use case for this combination is to implement the equivalent of hooks in Emacs Lisp -- you could, for example, define a system shutdown multimethod, and various implementations can be added as needed to to define additional shutdown actions:;; This example uses the `everything-dispatcher`, see below ;; ;; defmulti always expects a dispatch fn, but since it's not used by the everything dispatcher we can pass ;; anything (m/defmulti ^:private shutdown! :none :dispatcher (m/everything-dispatcher) :combo (m/do-method-combination)) (m/defmethod shutdown! :task-scheduler [] (println "Shutting down task scheduler...")) (m/defmethod shutdown! :web-server [] (println "Shutting down web server...")) (m/prefer-method! #'shutdown! :web-server :task-scheduler) (m/defmethod shutdown! :around :initiate [] (println "Initiating shutdown...") (next-method)) (shutdown!) ;; -> Initiating shutdown... ;; -> Shutting down web server... ;; -> Shutting down task scheduler...
-
min-method-combination
-- returns the minimum value returned by all primary methods. -
max-method-combination
-- returns the maximum value returned by all primary methods. -
+-method-combination
-- returns the sum of all values returned by all primary methods. The classic example use case is calculating total electricity usage from a variety of sources. -
seq-method-combination
-- returns a lazy sequence of all values returned by all primary methods. -
concat-method-combination
-- returns a lazy concatenated sequence of all values returned by all primary methods.seq-method-combination : map :: concat-method-combination : mapcat
-
and-method-combination
-- invokes all primary methods until one returns a non-truthy value, at which point it short-circuts. -
or-method-combination
-- invokes all primary methods until one returns a truthy value, at which points it short-circuts and returns that value. You could use this method combination to implement a chain-of-responsibility pattern.
-
Dispatchers
-
multi-default-dispatcher
-- The default. Similar to the behavior of vanilla Clojure multimethods, but also supports "partial default" methods like[String :default]
or[:default :some-key]
. See this blog post for a more detailed explanation. -
standard-dispatcher
-- Dispatcher that mimics behavior of vanilla Clojure multimethods. Uses a single hierarchy, dispatch function, default dispatch value, and map of preferences defined byprefer-method
. -
everything-dispatcher
-- Dispatcher that always considers all primary and auxiliary methods to be matches. Does not calculate dispatch values, but can sort methods from most- to least-specific using a hierarchy and map of preferences. Particularly useful with the operator method combinations.
Method Tables
-
standard-method-table
-- The default. A simple method table based on Clojure immutable maps. -
clojure-method-table
-- Likestandard-method-table
, but disallows auxiliary methods.
Caches
-
simple-cache
-- Default for multimethods constructed programmatically. Simple cache that maintains a map of dispatch value -> effective method. -
watching-cache
-- Default for multimethods constructed viadefmulti
. Wraps another cache (by default,simple-cache
) and watches one or more Vars (by default,#'clojure.core/global-hierarchy
), clearing the cache when the watched Vars change. Clears watches when cache is garbage-collected.
Multifn Impls
-
standard-multifn-impl
-- Basic impl that manages a method combination, dispatcher, and method table. -
cached-multifn-impl
-- wraps another multifn impl and an instance ofCache
to implement caching.
Validation
Methodical offers a few opportunities for validation above and beyond what normal Clojure multimethods offer.
:dispatch-value-spec
If you include a :dispatch-value-spec
in the metadata of a defmulti
, it will automatically be used to validate the
dispatch value form of any defmethod
forms at macroexpansion time:
(m/defmulti mfx
{:arglists '([x y]), :dispatch-value-spec (s/cat :x keyword?, :y int?)}
(fn [x y] [x y]))
(m/defmethod mfx [:x 1]
[x y]
{:x x, :y y})
;; => #'methodical.macros-test/mfx
(m/defmethod mfx [:x]
[x y]
{:x x, :y y})
;; failed: Insufficient input in: [0] at: [:args-for-method-type :primary :dispatch-value :y] [:x]
This is a great way to make sure people use your multimethods correctly and catch errors right away.
:defmethod-arities
A set of allowed/required arities that defmethod
forms are allowed to have. defmethod
forms must have arities that
match all of the specified :defmethod-arities
, and all of its arities must be allowed by :defmethod-arities
:
(m/defmulti ^:private mf
{:arglists '([x]), :defmethod-arities #{1}}
keyword)
(m/defmethod mf :x [x] x)
;; => ok
(m/defmethod mf :x ([x] x) ([x y] x y))
;; => error: {:arities {:disallowed #{2}}}
(m/defmethod mf :x [x y] x y)
;; => error: {:required #{1}, :disallowed #{2}}
:defmethod-arities
must be a set of either integers or [:> n]
forms to represent arities with &
rest
arguments, e.g. [:>= 3]
to mean an arity of three or-more arguments:
;; methods must have both a 1-arity and a 3+-arity
(m/defmulti ^:private mf
{:arglists '([x] [x y z & more]), :defmethod-arities #{1 [:>= 3]}}
keyword)
(m/defmethod mf :x ([x] x) ([x y z & more] x))
;; => ok
(m/defmethod mf :x [x y] x)
;; => error: {:arities {:required #{1 [:>= 3]}, :disallowed #{2}}}
When rest-argument arities are used, Methodical is smart enough to allow them when appropriate even if they do not
specifically match an arity specified in :defmethod-arities
:
(m/defmulti ^:private mf
{:arglists '([x y z & more]), :defmethod-arities #{[:>= 3]}}
keyword)
(m/defmethod mf :x
([a b c] x)
([a b c d] x)
([a b c d & more] x))
;; => ok, because everything required by [:>= 3] is covered, and everything present is allowed by [:>= 3]
Debugging
Methodical offers debugging facilities so you can see what's going on under the hood, such as the trace
utility:
and the describe
utility, which outputs Markdown-formatted documentation, for human-friendly viewing in tools like
CIDER:
This extra information is automatically generated and appended to a multimethod's docstring whenever methods or preferences are added or removed.
Methodical multimethods also implement datafy
:
(clojure.datafy/datafy mf)
=>
{:ns 'methodical.datafy-test
:name 'methodical.datafy-test/mf
:file "methodical/datafy_test.clj"
:line 11
:column 1
:arglists '([x y])
:class methodical.impl.standard.StandardMultiFn
:combo {:class methodical.impl.combo.threaded.ThreadingMethodCombination
:threading-type :thread-last}
:dispatcher {:class methodical.impl.dispatcher.multi_default.MultiDefaultDispatcher
:dispatch-fn methodical.datafy-test/dispatch-first
:default-value :default
:hierarchy #'clojure.core/global-hierarchy
:prefs {:x #{:y}}}
:method-table {:class methodical.impl.method_table.standard.StandardMethodTable
:primary {:default
{:ns 'methodical.datafy-test
:name 'methodical.datafy-test/mf-primary-method-default
:doc "Here is a docstring."
:file "methodical/datafy_test.clj"
:line 15
:column 1
:arglists '([next-method x y])}}
:aux {:before {[:x :default] [{:ns 'methodical.datafy-test
:name 'methodical.datafy-test/mf-before-method-x-default
:doc "Another docstring."
:file "methodical/datafy_test.clj"
:column 1
:line 20
:arglists '([_x y])
:methodical/unique-key 'methodical.datafy-test}]}
:around {[:x :y] [{:ns 'methodical.datafy-test
:name 'methodical.datafy-test/mf-around-method-x-y
:file "methodical/datafy_test.clj"
:column 1
:line 25
:arglists '([next-method x y])
:methodical/unique-key 'methodical.datafy-test}]}}}
:cache {:class methodical.impl.cache.watching.WatchingCache
:cache {:class methodical.impl.cache.simple.SimpleCache
:cache {}}
:refs #{#'clojure.core/global-hierarchy}}}
Performance
Methodical is built with performance in mind. Although it is written entirely in Clojure, and supports many more features, its performance is similar or better to vanilla Clojure multimethods in many cases. Profiling results with Criterium show Methodical performing up to 20% faster in some cases:
;;; Vanilla clojure
Evaluation count : 1133167380 in 60 samples of 18886123 calls.
Execution time mean : 43.643309 ns
Execution time std-deviation : 0.733846 ns
Execution time lower quantile : 42.421811 ns ( 2.5%)
Execution time upper quantile : 44.646005 ns (97.5%)
Overhead used : 8.836747 ns
;;; Methodical
Evaluation count : 1359687900 in 60 samples of 22661465 calls.
Execution time mean : 35.327155 ns
Execution time std-deviation : 0.067655 ns
Execution time lower quantile : 35.219823 ns ( 2.5%)
Execution time upper quantile : 35.449303 ns (97.5%)
Overhead used : 8.836747 ns
There is still room for even more performance improvement!
License
Code, documentation, and artwork copyright © 2019-2023 Cam Saul.
Distributed under the Eclipse Public License, same as Clojure.