Top K Quickselect in Clojure

This is my implementation of a Top K quickselect in Clojure. The beauty of this algorithm is that on average it runs in O(n) time, unlike quicksort's O(n log n). Note the :keys destructuring.

(defn top-k [coll k]
  (let [pivot (rand-nth coll)
        {:keys [left middle right]} (group-by
                                      #(cond (= % pivot) :middle (< % pivot) :left :else :right) coll)
        left-cnt (count left)]
    (cond
      (= left-cnt k) left
      (= (inc left-cnt) k) (concat left middle)
      (> left-cnt k) (top-k left k)
      :else ; left-cnt < k
      (concat left middle (top-k right (- k left-cnt 1))))))

(eval)

;; Example

(top-k [4 3 2 1 0] 3) ; => (0 1 2)
(top-k [4 5 3 2 8 3 8 9 0 9] 3) ; => (0 2 3)

(eval)

;; How does it work?

The top-k function chooses a random pivot element in a collection coll and groups smaller elements into left and larger elements into right. Based on the length of left, it recursively applies top-k to the right side until it has k values and returns left.

This function could be rewritten with recur tail-call optimization to prevent blowing the stack for massive collections.