Awesome

diffit

A Clojure(Script) diff and patch implementation for vectors and maps.

Include the dependency as shown above in your project.clj

Diff calculates the edit-distance and a minimal edit-script to transform the first into the second collection via patch. The vector based diff implementation follows An O(NP) Sequence Comparison Algorithm by Wu, Manber, Myers and Miller.

Why

The obvious application of the vector based implementation is diff and patch of text files represented by seqs of strings.

Another could be the desire to synchronize the final result of a data transformation (composed of side-effect free functions) with state wrapped by an expensive mutation-based API, for example a bound JavaFX ObservableList. A fast diff with a configurable patch is one way to tackle this.

A 3rd scenario could be the requirement to transmit a minimal amount of data changes to a remote process for further processing.

There is clojure.data/diff but it does not exactly what I had in mind, as the following REPL output shows:

(clojure.data/diff [1 2 3 4 5 6 7 8] [1 3 4 5 6 9])
;= [[nil 2 3 4 5 6 7 8] [nil 3 4 5 6 9] [1]]

The docstring says that the resulting triplet contains [things-only-in-a things-only-in-b things-in-both]. However, the common sequence [3 4 5 6] is part of the first two, so the result is not a minimal edit-script.

Next stop was clj-diff, but unfortunately the good work never made it into a non-snapshot release. My implementation is from scratch, but - of course - contains insights from existing open source work such as clj-diff and others.

Usage

There are two namespaces: diffit.vec and diffit.map. Each contains a diff and a patch function.

diffit.vec/diff and diffit.vec/patch work for sequential things which will be treated as vectors.

diffit.map/diff and diffit.map/patch work for associative things.

Let x, y both be sequential or associative, respectively:

(diff x y) produces a diff-result which is a pair of [edit-distance edit-script].
(patch x diff-result) applies the edit-script in diff-result to x in order to reproduce y.
It holds that (= y (patch x (diff x y)).

In the REPL:

(require '[diffit.vec :as v])
;= nil
(v/diff [1 2 3 4] [1 2 7 8 4])
;= [3 [[:+ 2 [7 8]] [:- 4 1]]]

The diff-result is a pair: the first part is the edit-distance, the second part is a sequence of edits (called an edit-script) that is needed to produce the second input collection from the first.

The edit-script for vectors allows sequential processing in the patch function with insert and remove operations, where insert takes a sequence, a position and a sequence to insert, and remove takes a sequence, a position and a number of items to remove. Positions are zero-based.

The edit-script in the output above can be read like this:

First, add the sequence [7 8] at position 2 (this will shift the items 3 4 to the right).
Then, remove 1 item at position 4.

Once you have the result from diff you can apply its edit-script with patch:

(v/patch [1 2 3 4] diff-result)
;= [1 2 7 8 4]

The patch function can be used with two or three, respectively, additional arguments to hand in functions that actually do the insert, remove, assoc, dissoc or replace operation. This lets you adapt patch to an API based on mutation.

For a Java ArrayList here's a corresponding snippet:

(patch (fn [l index item]
		 (.addAll l index item)
		 l)
	   (fn [l index n]
		 (dotimes [_ n] (.remove l index))
		 l)
	   mutable-list	diff-result)

For Java HashMap this looks like this

(patch (fn [m k v]
         (.put m k v)
         m)
       (fn [m k]
	     (.remove m k)
         m)
       (fn [m k v]
         (.put m k v)
         m)
       mutable-map diff-result)

Performance for vector based diff

diffit.vec/diff is a slight bit better than the clj-diff implementation which uses the same algorithm and was also built with performance as key requirement. Both clearly outperform the 1.3.0 version of an open source Java library called diffutils.

(So far, I did no extensive research for other Java alternatives. If you find a better candidate that handles Java lists, and not only text, drop me a mail.)

Setup

I tested it for 2000 items where the sink bs was derived from the source as by randomly adding or removing items. Add/remove took place with 10% probability each, so 80% of the sink is the same as the source.

This snippet defines source and sink sequences:

(do (def as (range 2000))
    (def bs (rand-alter 80 10 10 as))))

Here's the piece of code that creates the sink from the source vector:

(defn rand-alter
  [pass-prob remove-prob add-prob xs]
  (let [ops (vec (concat (repeat pass-prob :=)
                         (repeat remove-prob :-)
                         (repeat add-prob :+)))]
    (reduce (fn [xs x]
              (case (rand-nth ops)
                :+ (conj xs x "-")
                :- xs
                := (conj xs x)))
            []
            xs)))

I used criterium bench to gather times on a JDK 1.8.0_5 with Clojure 1.6.0 and 4 cores of Intel(R) Core(TM) i5 CPU M 560 @ 2.67GHz. Here are the results:

diffit

diffit.vec-test> (>bench (diff as bs))
Evaluation count : 2340 in 60 samples of 39 calls.
             Execution time mean : 28.204187 ms
    Execution time std-deviation : 1.357085 ms
   Execution time lower quantile : 26.592218 ms ( 2.5%)
   Execution time upper quantile : 30.952525 ms (97.5%)
                   Overhead used : 2.049385 ns

Found 2 outliers in 60 samples (3.3333 %)
	low-severe	 1 (1.6667 %)
	low-mild	 1 (1.6667 %)
 Variance from outliers : 33.6221 % Variance is moderately inflated by outliers

clj-diff

diffit.vec-test> (>bench (clj-diff.core/diff as bs))
Evaluation count : 1860 in 60 samples of 31 calls.
             Execution time mean : 33.078036 ms
    Execution time std-deviation : 1.514409 ms
   Execution time lower quantile : 31.606479 ms ( 2.5%)
   Execution time upper quantile : 35.715329 ms (97.5%)
                   Overhead used : 2.049385 ns

Found 3 outliers in 60 samples (5.0000 %)
	low-severe	 2 (3.3333 %)
	low-mild	 1 (1.6667 %)
 Variance from outliers : 31.9529 % Variance is moderately inflated by outliers

java-diff-utils

diffit.vec-test> (>bench (DiffUtils/diff as bs))
Evaluation count : 120 in 60 samples of 2 calls.
             Execution time mean : 1.014824 sec
    Execution time std-deviation : 16.119216 ms
   Execution time lower quantile : 990.193018 ms ( 2.5%)
   Execution time upper quantile : 1.054244 sec (97.5%)
                   Overhead used : 2.049385 ns

Found 4 outliers in 60 samples (6.6667 %)
	low-severe	 4 (6.6667 %)
 Variance from outliers : 1.6389 % Variance is slightly inflated by outliers

Performance for map based diff

Unsurprisingly, the diff for maps is an order of magnitude faster due to the performance characteristics of the underlying datastructures.

I did no comparison between different libraries (are there any?) or implementations.

License

Distributed under the Eclipse Public License version 1.0.