=head1 NAME

KinoSearch::Docs::FileFormat - Overview of index file format.


It is not necessary to understand the current implementation details of the
index file format in order to use KinoSearch effectively, but it may be
helpful if you are interested in tweaking for high performance, exotic usage,
or debugging and development.  

On a file system, an index is a directory.  The files inside have a
hierarchical relationship: an index is made up of "segments", each of which is
an independent inverted index with its own subdirectory; each segment is made
up of several component parts.

             |         |--segmeta.json
             |         |--cfmeta.json
             |         |--cf.dat-------|
             |                         |--[lexicon]
             |                         |--[postings]
             |                         |--[documents]
             |                         |--[highlight]
             |                         |--[deletions]
             |         |--segmeta.json
             |         |--cfmeta.json
             |         |--cf.dat-------|
             |                         |--[lexicon]
             |                         |--[postings]
             |                         |--[documents]
             |                         |--[highlight]
             |                         |--[deletions]

=head1 Write-once philosophy

All segment directory names consist of the string "seg_" followed by a number
in base 36: seg_1, seg_5m, seg_p9s2 and so on, with higher numbers indicating
more recent segments.  Once a segment is finished and committed, its name is
never re-used and its files are never modified.

Old segments become obsolete and can be removed when their data has been
consolidated into new segments during the process of segment merging and
optimization.  A fully-optimized index has only one segment.

=head1 Top-level files

There are a handful of "top-level" files which belong to the entire index
rather than to a particular segment.

=head2 snapshot_XXX.json

A "snapshot" file, e.g. C<snapshot_m7p.json>, is list of index files.  Because
index files, once written, are never modified, the list of files in a snapshot
defines a point-in-time view of the data in an index.

Like segment directories, snapshot files also utilize the
unique-base-36-number naming convention; the higher the number, the more
recent the file.  The appearance of a new snapshot file within the index
directory constitutes an index update.  While a new segment is being written
new files may be added to the index directory, but until a new snapshot file
gets written, a Searcher opening the index for reading won't know about them.

=head2 schema_XXX.json

The schema file is a Schema object describing the index's format, serialized
as JSON.  It, too, is versioned, and a given snapshot file will reference one
and only one schema file.

=head2 write.lock

By default, only one indexing process may safely modify the index at any given
time.  Processes reserve an index by laying claim to the C<write.lock> file.
A smattering of other lock files may be used from time to time, as well.

=head1 A segment's component parts

By default, each segment has up to five logical components: lexicon, postings,
document storage, highlight data, and deletions.  Binary data from these
components gets stored in virtual files within the "cf.dat" compound file;
metadata is stored in a shared "segmeta.json" file.

=head2 segmeta.json

The segmeta.json file is a central repository for segment metadata.  In
addition to information such as document counts and field numbers, it also
warehouses arbitrary metadata on behalf of individual index components.

=head2 Lexicon 

Each indexed field gets its own lexicon in each segment.  The exact files
involved depend on the field's type, but generally speaking there will be two
parts.  First, there's a primary C<lexicon-XXX.dat> file which houses a
complete term list associating terms with corpus frequency statistics,
postings file locations, etc.  Second, one or more "lexicon index" files may
be present which contain periodic samples from the primary lexicon file to
facilitate fast lookups.

=head2 Postings

L<Posting|KinoSearch::Posting> is a technical term from the field of
L<information retrieval|KinoSearch::Docs::IRTheory>, defined as a single
instance of a one term indexing one document.  If you are looking at the index
in the back of a book, and you see that "freedom" is referenced on pages 8,
86, and 240, that would be three postings, which taken together form a
"posting list".  The same terminology applies to an index in electronic form.

Each segment has one postings file per indexed field.  When a search is
performed for a single term, first that term is looked up in the lexicon.  If
the term exists in the segment, the record in the lexicon will contain
information about which postings file to look at and where to look.

The first thing any posting record tells you is a document id.  By iterating
over all the postings associated with a term, you can find all the documents
that match that term, a process which is analogous to looking up page numbers
in a book's index.  However, each posting record typically contains other
information in addition to document id, e.g. the positions at which the term
occurs within the field.

=head2 Documents

The document storage section is a simple database, organized into two files:


=item * 

B<ocuments.dat> - Serialized documents.

=item *

B<documents.ix> - Document storage index, a solid array of 64-bit integers
where each integer location corresponds to a document id, and the value at
that location points at a file position in the documents.dat file.


=head2 Highlight data 

The files which store data used for excerpting and highlighting are organized
similarly to the files used to store documents.


=item * 

B<highlight.dat> - Chunks of serialized highlight data, one per doc id.

=item *

B<highlight.ix> - Highlight data index -- as with the C<documents.ix> file, a
solid array of 64-bit file pointers.


=head2 Deletions

When a document is "deleted" from a segment, it is not actually purged right
away; it is merely marked as "deleted" via a deletions file.  Deletions files
contains bit vectors with one bit for each document in the segment; if bit
#254 is set then document 254 is deleted, and if that document turns up in a
search it will be masked out.

It is only when a segment's contents are rewritten to a new segment during the
segment-merging process that deleted documents truly go away.

=head1 Compound Files

If you peer inside an index directory, you won't actually find any files named
"documents.dat", "highlight.ix", etc. unless there is an indexing process
underway.  What you will find instead is one "cf.dat" and one "cfmeta.json"
file per segment.

To minimize the need for file descriptors at search-time, all per-segment
binary data files are concatenated together in "cf.dat" at the close of each
indexing session.  Information about where each file begins and ends is stored
in C<cfmeta.json>.  When the segment is opened for reading, a single file
descriptor per "cf.dat" file can be shared among several readers.

=head1 A Typical Search

Here's a simplified narrative, dramatizing how a search for "freedom" against
a given segment plays out:


=item 1

The searcher asks the relevant Lexicon Index, "Do you know anything about
'freedom'?"  Lexicon Index replies, "Can't say for sure, but if the main
Lexicon file does, 'freedom' is probably somewhere around byte 21008".  

=item 2

The main Lexicon tells the searcher "One moment, let me scan our records...
Yes, we have 2 documents which contain 'freedom'.  You'll find them in
seg_6/postings-4.dat starting at byte 66991."

=item 3

The Postings file says "Yep, we have 'freedom', all right!  Document id 40
has 1 'freedom', and document 44 has 8.  If you need to know more, like if any
'freedom' is part of the phrase 'freedom of speech', ask me about positions!

=item 4

If the searcher is only looking for 'freedom' in isolation, that's where it
stops.  It now knows enough to assign the documents scores against "freedom",
with the 8-freedom document likely ranking higher than the single-freedom



Copyright 2005-2009 Marvin Humphrey


See L<KinoSearch> version 0.30.