LUMIERA.clone/doc/technical/library/DiffFramework.txt

Diff Handling Framework
=======================

Within the support library, in the namespace `lib::diff`, there is a collection of loosely coupled of tools
known as »the diff framework«. It revolves around generic representation and handling of structural differences.
Beyond some rather general assumptions, to avoid stipulating the usage of specific data elements or containers,
the framework is kept _generic_, cast in terms of *elements*, *sequences* and *strategies*
for access, indexing and traversal.

.motivation
**********************
Diff handling is multi purpose within Lumiera:
Diff representation is seen as a meta language and abstraction mechanism;
it enables tight collaboration without the need to tie and tangle the involved implementation data structures.
Used this way, diff representation reduces coupling and helps to cut down overall complexity — so to justify
the considerable amount of complexity seen within the diff framework implementation.
**********************

Definitions
-----------
element::
  the atomic unit treated in diff detection, representation and application. +
  Elements are considered to be

  - lightweight copyable values
  - equality comparable
  - bearing distinct identity
  - unique _as far as concerned_

sequence::
  data is delivered in the form of a sequence, which might or might not be _ordered,_
  but in any case will be traversed once only.

diff::
  the changes necessary to transform an input sequence (``old sequence'') into a target sequence (``new sequence'')

diff language::
  differences are spelled out in linearised form: as a sequence of constant-size diff actions, called »diff verbs«

diff verb::
  a single element within a diff. Diff verbs are conceived as operations, which,
  when applied consuming the input sequence, will produce the target sequence of the diff.

diff application::
  the process of consuming a diff (sequence of diff verbs), with the goal to produce some effect at the 
  _target_ of diff application. Typically we want to apply a diff to a data sequence, to mutate it
  into a new shape, conforming with the shape of the diff's ``target sequence''

diff generator::
  a facility producing a diff, which is a sequence of diff verbs.
  Typically, a diff generator works lazily, demand driven.

diff detector::
  special kind of diff generator, which takes two data sequences as input:
  an ``old sequence'' and a ``new sequence''. The diff detector traverses and compares
  these sequences to produce a diff, which describes the steps necessary to transform
  the ``old'' shape into the ``new'' shape of the data.


List Diff Algorithm
-------------------
While in general this is a well studied subject, in Lumiera we'll confine ourselves to a very
specific flavour of diff handling: we rely on _elementary atomic units_ with well established
object identity. And in addition, within the scope of one coherent diff handling task,
we require those elements to be 'unique'. The purpose of this design decision is to segregate
the notorious matching problem and treat diff handling in isolation.

Effectively this means that, for any given element, there can be at most _one_ matching
counterpart in the other sequence, and the presence of such can be detected by using an *index*.
In fact, we retrieve an index for every sequence involved into the diff detection task;
this is our trade-off for simplicity in the diff detection algorithm.footnote:[traditionally,
diff detection schemes, especially those geared at text diff detection, engage into great lengths
of producing an ``optimal'' diff, which effectively means to build specifically tuned pattern
or decision tables, from which the final diff can then be pulled or interpreted.
We acknowledge that in our case building a lookup table index with additional annotations can
be O(n^2^); we might well be able to do better, but likely for the price of turning the algorithm
into some kind of mental challenge.]
In case this turns out as a performance problem, we might consider integrating the index
maintenance into the data structure to be diffed, which shifts the additional impact of
indexing onto the data population phase.footnote:[in the general tree diff case this is far
from trivial, since we need an self-contained element index for every node, and we need the
ability to take a snapshot of the ``old'' state before mutating a node into ``new'' shape]

Element classification
~~~~~~~~~~~~~~~~~~~~~~
By using the indices of the old and the new sequence, we are able to _classify_ each element:

- elements only present in the new sequence are treated as *inserts*
- elements only present in the old sequence are treated as *deletes*
- elements present in both sequences form the *permutation*

Processing pattern
~~~~~~~~~~~~~~~~~~
We _consume both the old and the new sequence synchronously, while emitting the diff sequence_.

The diff describes a sequence of operations, which, when applied, consume a sequence congruent
to the old sequence, while emitting a sequence congruent to the new sequence. We use the
following *list diff language* here:

verb `ins(elm)`::
  insert the given argument element `elm` at the _current processing position_
  into the target sequence. This operation allows to inject new data
verb `del(elm)`::
  delete the _next_ element `elm` at _current position._
  For sake of verification, the element to be deleted is also included as argument (redundancy).
verb `pick(elm)`::
  accepts the _next_ element at _current position_ into the resulting altered sequence.
  The element is given redundantly as argument.
verb `find(elm)`::
  effect a re-ordering of the target list contents. This verb requires to search
  for the (next respective single occurrence of the) given element in the remainder
  of the datastructure, to bring it forward and insert it as the _next_ element.
verb `skip(elm)`::
  processing hint, emitted at the position where an element previously fetched by
  some `find` verb happened to sit in the old order. This allows an optimising
  implementation to ``fetch'' a copy and just drop or skip the original,
  thereby avoiding to shift other elements.

Since _inserts_ and _deletes_ can be detected and emitted right at the processing frontier,
for the remaining theoretical discussion, we consider the insert / delete part filtered
away conceptually, and concentrate on generating the permutation part.

Handling sequence permutation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Permutation handling turns out to be the turning point and tricky part of diff generation;
the handling of new and missing members can be built on top, once we manage to describe a
permutation diff. If we consider -- for a theoretical discussion -- the _inserts_ and _deletes_
to be filtered away, what remains is a permutation of index numbers to describe the re-ordering.
We may describe this re-ordering by the index numbers in the new sequence, given in the order
of the old sequence. For such a re-ordering permutation, the theoretically optimal result
can be achieved by http://wikipedia.org/wiki/Cycle_sort[Cycle Sort] in _linear time_.footnote:[
assuming random access by index is possible, *Cycle Sort* walks the sequence once. Whenever
encountering an element out-of order, i.e. new postion != current position, we leap to the
indicated new position, which necessarily will be out-of-order too, so we can leap from there
to the next indicated position, until we jump back to the original position eventually. Such
a closed cycle can then just be _rotated_ into proper position. Each permutation can be
decomposed into a finite number of closed cycles, which means, after rotating all cycles
out of order, the permutation will be sorted completely.] Starting from such ``cycle rotations'',
we possibly could work out a minimal set of moving operations.

But there is a twist: Our design avoids using index numbers, since we aim at _stream processing_
of diffs. We do not want to communicate index numbers to the consumer of the diff; rather we
want to communicate reference elements with our _diff verbs_. Thus we prefer the most simplistic
processing mechanism, which happens to be some variation of *Insertion Sort*.footnote:[to support
this choice, Insertion Sort -- in spite of being O(n^2^) -- turns out to be the best choice for
sorting small data sets for reasons of cache locality; even typical Quicksort implementations
switch to insertion sorting of small subsets for performance reasons]


Implementation and Complexity
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Based on these choices, we're able to consume two permutations of the same sequence simultaneously,
while emitting `find` and `pick` verbs to describe the re-ordering. Consider the two sequences
split into an already-processed part, and a part still-to-be-processed.

.Invariant
Matters are arranged such, that, in the to-be-processed part, each element appearing at the
front of the ``new'' sequence _can be consumed right away_.

Now, to arrive at that invariant, we use indices to determine

- if the element at head of the old sequence is not present in the new sequence, which means
  it has to be deleted
- while an element at head of the new sequence not present in the old sequence has to be inserted
- and especially a non-matching element at the old sequence prompts us to fetch the right
  element from further down in the sequence and insert it a current position

after that, the invariant is (re)established and we can consume the element and move the
processing point forward.

For generating the diff description, we need index tables of the ``old'' and the ``new'' sequence,
which causes a O(n log n) complexity and storage in the order of the sequence length. Application
of the diff is quadratic, due to the find-and-insert passes.footnote:[In the theoretical treatment
of diff problems it is common to introduce a *distance metric* to describe how _far apart_ the
two sequences are in terms of atomic changes. This helps to make the quadratic (or worse) complexity
of such algorithms look better: if we know the sequences are close, the nested sub-scans will be
shorter than the whole sequence (with n·d < n^2^). +
However, since our goal is to support permutations and we have to deal with arbitrary sequences, such
an argument looks somewhat pointless. Let's face it, structural diff computation is expensive; the only
way to keep matters under control is to keep the local sequences short, which means to exploit structural
knowledge instead of comparing the entire data as flat sequence]