juji-io.editscript

https://github.com/juji-io/editscript.git

git clone 'https://github.com/juji-io/editscript.git'

(ql:quickload :juji-io.editscript)
160

editscript

A library designed to extract the differences between two Clojure/Clojurescript data structures as an “editscript”, which represents the minimal modification necessary to transform one to another. Currently, the library can diff and patch any nested Clojure/Clojurescript data structures consisting of regular maps, vectors, lists, sets and values.

At Juji, we need to take snapshots of our AI agents' states and later restore them. Such a use case requires a good diffing library for nested Clojure data structures to avoid overwhelming our storage systems. I have not found such a library in Clojure ecosystem, so I implemented my own. Hopefully this little library could be of some use to further enhance the Clojure's unique strength of Data-Oriented Programming.

Usage

The library is available at clojars and npm

npm version

API Documentation is available on cljdoc

cljdoc badge

Here is a usage example:

(use 'editscript.core)
(use 'editscript.edit)

;; Here are two pieces of data, a and b
(def a ["abc" 24 22 {:a [1 2 3]} 1 3 #{1 2}])
(def b [24 23 {:a [2 3]} 1 3 #{1 2 3}])

;; compute the editscript between a and b
(def d (diff a b))

d
;;==>
;;[[[0] :-]
;; [[1] :r 23]
;; [[2 :a 0] :-]
;; [[5 3] :+ 3]]

;; get the edit distance, i.e. number of edits
(edit-distance d)
;;==> 4

;; get the size of the editscript
(get-size d)
;;==> 22

;; patch a with the editscript to get back b, so that
(= b (patch a d))
;;==> true

An EditScript contains a vector of edits, where each edit is a vector of two or three elements.

The first element of an edit is the path, similar to the path vector in the function call update-in. However, update-in only works for associative data structures (map and vector), whereas the editscript works for map, vector, list and set alike.

The second element of an edit is a keyword representing the edit operation, which is one of :- (deletion), :+ (addition), and :r(replacement).

For addition and replacement operation, the third element is the value of new data.


;; get the edits as a plain Clojure vector
(def v (get-edits d))

v
;;==>
;;[[[0] :-]
;; [[1] :r 23]
;; [[2 :a 0] :-]
;; [[5 3] :+ 3]]

;; the plain Clojure vector can be passed around, stored, or modified as usual,
;; then be loaded back as a new EditScript
(def d' (edits->script v))

;; the new EditScript works the same as the old one
(= b (patch a d'))
;;==> true

Diffing Algorithms

The library currently implements two diffing algorithms. The default algorithm produces diffs that are optimal in the number of editing operations and the resulting script size. A quick algorithm is also provided, which does not guarantee optimal results but is very fast.

A* diffing

This A* algorithm aims to achieve optimal diffing in term of minimal size of resulting editscript, useful for storage, query and restoration. This is an original algorithm that has some unique properties: unlike many other general tree differing algorithms such as Zhang & Shasha 1989, our algorithm is structure preserving.

Roughly speaking, the edit distance is defined on sub-trees rather than nodes, such that the ancestor-descendant relationship and tree traversal order are preserved, and nodes in the original tree does not split or merge. These properties are useful for diffing and patching Clojure's immutable data structures because we want to leverage structure sharing and use identical? reference checks. The additional constraints also yield algorithms with better run time performance than the general ones. Finally, these constraints feel natural for a Clojure programmer.

The structure preserving properties were proposed in Lu 1979 and Tanaka 1995. These papers describe diffing algorithms with O(|a||b|) time and space complexity. We designed an A* based algorithm to achieve some speedup. Instead of searching the whole editing graph, we typically search a portion of it along the diagonal.

The implementation is optimized for speed. Currently the algorithm spent most of its running time calculating the cost of next steps, perhaps due to the use of a very generic heuristic. A more specialized heuristic for our case should reduce the number of steps considered. For special cases of vectors and lists consisting of leaves only, we also use the quick algorithm below to enhance the speed.

Although much slower than the non-optimizing quick algorithm below, the algorithm is practical for common Clojure data that include lots of maps. Maps and sets do not incur the penalty of a large search space in the cases of vectors and lists. For a drawing data set, the diffing time of the algorithm is in the range of 2ms to 4ms on a 2014 2.8 GHz Core i5 16GB MacBook Pro.

Quick diffing

This quick diffing algorithm simply does an one pass comparison of two trees so it is very fast. For sequence (vector and list) comparison, we implement Wu et al. 1990, an algorithm with O(NP) time complexity, where P is the number of deletions if b is longer than a. The same sequence diffing algorithm is also implemented in diffit. Using their benchmark, our implementation has slightly better performance due to more optimizations. Keep in mind that our algorithm also handles nested Clojure data structures. Compared with our A* algorithm, our quick algorithm is up to two orders of magnitude faster.

The Wu algorithm does not have replacement operations, and assumes each edit has a unit cost. These do not work well for tree diffing. Consequently, the quick algorithm does not produce optimal results in term of script size. In principle, simply changing a pointer to point to b instead of a produces the fastest “diffing” algorithm of the world, but that is not very useful. The quick algorithm has a similar problem.

For instances, when consecutive deletions involving nested elements occur in a sequence, the generated editscript can be large. For example:

(def a [2 {:a 42} 3 {:b 4} {:c 29}])
(def b [{:a 5} {:b 5}])

(editscript.diff.quick/diff a b)
;;==>
;;[[[0] :-]
;; [[0] :-]
;; [[0] :-]
;; [[0 :b] :-]
;; [[0 :a] :+ 5]
;; [[1 :c] :-]
;; [[1 :b] :+ 5]]

(editscript.diff.a-star/diff a b)
;;==>
;;[[[0] :-]
;; [[0 :a] :r 5]
;; [[1] :-]
;; [[1 :b] :r 5]
;; [[2] :-]]

In this case, the quick algorithm seems to delete the original and then add new ones back. The reason is that the quick algorithm does not drill down (i.e. do replacement) at the correct places. It currently drills down wherever it can. An optimizing algorithm is needed if minimal diffs are desired.

Platform

The library supports JVM Clojure and Clojurescript. The later has been tested with node, nashorn, chrome, safari, firefox and lumo. E.g. run our test suite:

# Run Clojure tests
lein test

# Run Clojurescript tests on node.js
lein doo node

# Run Clojurescript tests on chrome
lein doo chrome browser

Rationale

Editscript is designed with stream processing in mind. An editscript should be conceptualized as a chunk in a potentially endless stream of changes. Individual editscripts can combine (concatenate) into a larger edistscript. I consider editscript as a part of a larger data-oriented effort, that tries to elevate the level of abstraction of data from the granularity of characters, bytes or lines to that of maps, sets, vectors, and lists. So instead of talking about change streams in bytes, we can talk about change streams in term of higher level data structures.

Roadmap

There are a few things I have some interest in exploring with this library. Of course, ideas, suggestions and contributions are very welcome.

References

License

Copyright © 2018 Juji, Inc.

Distributed under the Eclipse Public License either version 1.0 or (at your option) any later version.