# Building persistent data structures in Clojure

## How to build your custom data structures in Clojure

Finn Völkel
Software Engineer

Clojure ships by default with a really good set of data structures, so it's worth thinking hard before building your own. There are however cases when doing so is warranted. Let's look at two approaches that can be taken when building custom data structures.

## Bags as a primitive example

For the purpose of this post we are going to look at bags (also known as multisets). A bag extends the concept of a set to a set with a multiplicity for each item. So whenever we check for an item in a bag, the result will be the number of times the item was added to the bag. nil if it is not present. The natural way to represent a bag in Clojure is via a map.

(defn bag [] {:size 0 :content {}})

We also add a size field so we can answer the question of the total size of the bag in $O(1)$.

The implementations for add and rmv then come naturally.

(defn add [bag ele]
(-> bag
(update :size inc)
(update-in [:content ele] (fnil + 0) 1)))

(defn rmv [{content :content :as bag} ele]
(let [cnt (get content ele)]
(-> bag
(update :size dec)
(assoc :content (cond
(nil? cnt) (throw (ex-info "No such element" {:ele ele}))
(= cnt 1) (dissoc content ele)
:else (update content ele - 1))))))

In add we simply increase the size of the whole bag and increment the counter of the element if existent (defaulting to a value of 0 if not). When removing an element from the bag we first decrease the total size of the bag and then decrement the counter of the appropriate element. Two special cases are if the item to be removed is not present in the bag and if the item is only present with multiplicity 1. In the former we throw an error as the operation is not valid, in the latter we simply remove any reference to the element from the content map.

Checking for an item in a bag

(defn contains?' [bag ele]
(some? (get-in bag [:content ele])))

(defn count' [bag ele]
(get-in bag [:content ele]))

(defn size [bag] (:size bag))

contains?' and count' are quoted so that we are not overriding Clojure's core functions.

Using what we have just defined.

(-> (bag) (add "foo") (add "foo") (rmv "foo") (add "bar"))
;; => {:size 2, :content {"foo" 1, "bar" 1}}
(contains?' *1 "foo")
;; => true
(count' *2 "foo")
;; => 1

## Extending Clojure’s core library

The second approach is to extend Clojure's core interfaces. In this case clojure.lang.IPersistentSet. The implementation details remain the same except that we need to create a new object when modifying the bag.

(deftype Bag [size content]
clojure.lang.IPersistentSet

(seq [this] (seq content))

(count [this] size)

(cons [this o]
(Bag. (inc size) (update content o (fnil + 0) 1)))

(empty [this] (Bag. 0 {}))

(equiv [this other] (= this other))

(disjoin [this key]
(let [cnt (get content key)]
(Bag. (dec size)
(cond
(nil? cnt) (throw (ex-info "No such key" {:key key}))
(= cnt 1) (dissoc content key)
:else (update content key - 1)))))

(contains [this key] (some? (get content key)))

(get [this key] (get content key)))

(defn bag [] (Bag. 0 {}))

With this implementation we can use Clojure’s core functions to manipulate our bag.

(-> (bag) (conj "foo") (conj "foo") (disj "foo") (conj "bar"))
;; => #{["foo" 1] ["bar" 1]}
(contains? *1 "foo")
;; => true
(get *2 "foo")
;; => 1

The trade-off between the two approaches is one between explicit and implicit. Extending Clojure’s core interfaces gives the flexibility of using core functions. If the people manipulating that part of the code base are very familiar with it, go for it. The gotcha is that people not familiar with it might make wrong assumptions about incoming/outgoing data (i.e. we can only disjoin on a standard set). In that case it might be more prudent to create an extra namespace with dedicated functions.

## Choosing the interfaces

In case one decides to extend Clojure’s core interfaces one will probably select a subset of some common interfaces. Some examples are

• java.lang.Object - for implementing equality
• clojure.lang.IHasheq - if the type needs to support Clojure’s hash function
• clojure.lang.IPersistentMap - in case your data structure fits a map type
• clojure.lang.IFn - if the type should support a function-like style (similar to (#{1 2} 1) => 1) )
• clojure.lang.IObj - for metadata support
• clojure.lang.Counted - so that count works for the type
• clojure.lang.Associative - if the type fits an associative interface
• clojure.lang.Seqable - if the type implements the Seq interface

In certain cases an interface extends a more primitive interface. For example maps implement clojure.lang.Associative.

We could extend our Bag type with clojure.lang.IFn to get the same function semantics as Clojure core sets.

clojure.lang.IFn
(invoke [this k]
(get this k))

(applyTo [this args]
(let [n (clojure.lang.RT/boundedLength args 2)]
(if (= 1 n)
(.invoke this (first args))
(throw (clojure.lang.ArityException. n (.. this (getClass) (getSimpleName)))))))
(def b (-> (bag) (conj "foo") (conj "foo") (conj "bar")))
(b "foo") ;; => 2

If we are creating a new type there is also the possibility to modify the printer.

(defmethod print-method Bag [bag ^java.io.Writer w]
(.write w "#Bag{")
(doseq [kv (seq bag)]
(.write w (pr-str kv)))
(.write w "}"))
(-> (bag) (conj "foo") (conj "foo") (disj "foo") (conj "bar"))
;; => #Bag{["foo" 1]["bar" 1]}

When building a new type on top of Clojure’s persistent data structures you get history for free. This is also true for the standard core data structures, but it’s still worth reiterating.

(def bag-with-history (let [b (bag)] {:current b :history [b]}))

(defn conj-history [{:keys [current] :as ds} v]
(let [new-ds (conj current v)]
(-> ds
(update :history conj new-ds)
(assoc :current new-ds))))

(-> bag-with-history
(conj-history "foo")
(conj-history "foo")
(conj-history "bar")
:history)
;; => [#Bag{} #Bag{["foo" 1]} #Bag{["foo" 2]} #Bag{["foo" 2]["bar" 1]}]

We hope this post has been useful to you and that you will be hacking on your own data structures soon.

### References

Recommended Resources