NAME
KinoSearch::Docs::FileFormat - overview of invindex file format
OVERVIEW
It is not necessary to understand the guts of the Lucene-derived "invindex" file format in order to use KinoSearch, but it may be helpful if you are interested in tweaking for high performance, exotic usage, or debugging and development.
On a file system, all the files in an invindex exist in one, flat directory. Conceptually, the files have a hierarchical relationship.
[invindex]--|
|-"segments" file
|
|-[segments]------|
|--[seg _0]--|
| |--[postings]
| |--[stored fields]
| |--[term vectors]
| |--[deletions]
|
|--[seg _1]--|
| |--[postings]
| |--[stored fields]
| |--[term vectors]
| |--[deletions]
|
|--[ ... ]---|
The "segments" file keeps a list of the segments that make up an invindex. When a new segment is being written, KinoSearch may put files into the directory, but until the segments file is updated, a Searcher reading the index won't know about them.
Each segment is an independent inverted index. All the files which belong to a given segment share a common alphanumeric prefix which begins with an underscore: _0, _a8, _rj. A fully optimized index has only a single segment.
In theory there are many files which make up each segment. However, when you look inside an invindex not in the process of being updated, you'll probably see only the segments file and files with either a .cfs or .del extension. The .cfs file, a "compound" file which is consolidated when a segment is finalized, "contains" all the other per-segment files.
Segments are written once, and with the exception of the deletions file, are never modified once written. They are deleted when their data is written to new segments during the process of optimization.
A segment's component parts
Each segment can be said to have four logical parts: postings, stored fields, the deletions file, and the term vectors data.
Stored fields
The stored fields are organized into two files.
[seg_name].fdx - Field inDeX - pointers to field data
[seg_name].fdt - Field DaTa - the actual stored fields
When a document turns up as a hit in a search and must be retrieved, KinoSearch looks at the Field inDeX file to see where in the data file the document's stored fields start, then retrieves all of them in one lump.
_1.fdx--|
|--[doc#0 => 0]----->_1.fdt--|
| |--[bodytext]
| |--[title]
| |--[url]
|--[doc#1 => 305]----->_1.fdt--| # byte 305
| |--[bodytext]
| |--[title]
| |--[url]
|--[...]--------------->_1.fdt--|--[...]Postings
"Posting" is a technical term from the field of Information Retrieval which refers to an 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.
The postings data is spread out over 4 main files (not including field normalization data, which we'll get to in a moment). From lowest to highest in the hierarchy, they are...
[seg_name].prx - PRoXimity data. A list of the positions at which terms appear in any given document. The .prx file is just a raw stream of VInts; the document numbers and terms are implicitly indicated by files higher up the hierarchy.
[seg_name].frq - FReQuency data for terms. If a term has a frequency of 5 in a given document, that implies that there will be 5 entries in the .prx file. The terms themselves are implicitly specified by the .tis file.
_1.frq--|
|--[doc#40 => 2]----->_1.prx--|--[54,107]
|--[doc#0 => 1]----->_1.prx--|--[6]
|--[doc#6 => 1]----->_1.prx--|--[504]
|--[doc#36 => 3]----->_1.prx--|--[2,33,747]
|--[...]------------->_1.frq--|--[...][seg_name].tis - TermInfoS. Among the items stored here is the term's doc_freq, which is the number of documents the term appears in. If a term has a doc_freq of 22 in a given collection, that implies that there will be 22 corresponding entries in the .frq file. Terms are ordered lexically, first by field, then by term text.
_1.tis--|
|--[...]----------------------->_1.frq--|--[...]
|--[bodytext:mule => 1]-->_1.frq--|--[doc#40 => 2]
|--[bodytext:multitude => 3]-->_1.frq--|--[doc#0 => 1]
| |--[doc#6 => 1]
| |--[doc#36 => 3]
|--[bodytext:navigate => 1]-->_1.frq--|--[doc#21 => 1]
|--[...]----------------------->_1.frq--|--[...]
|--[title:amendment => 27]-->_1.frq--|--[doc#21 => 1]
| |--[doc#22 => 1]
|--[...]----------------------->_1.frq--|--[...][seg_name].tii - TermInfos Index. This file, which is decompressed and loaded into RAM as soon as the IndexReader is initialized, contains a small subset of the .tis data, with pointers to locations in the .tis file. It is used to locate the right general vicinity in the .tis file as quickly as possible.
_1.tii--|
|--[bodytext:a => 20]---------->_1.tis--|--[bodytext:a] # byte 20
| |--[bodytext:about]
| |--[bodytext:absolute]
| |--[...]
|--[bodytext:mule => 27065]---->_1.tis--|--[bodytext:mule]
| |--[bodytext:multitude]
| |--[...]
|--[title:amendment => 56992]-->_1.tis--|--[title:amendment]
|--[...]Here's a simplified version of how a search for "freedom" against a given segment plays out:
The searcher asks the .tii file, "Do you know anything about 'freedom'?" The .tii file replies, "Can't say for sure, but if the .tis file does, 'freedom' is probably somewhere around byte 21008".
The .tis file tells the searcher "Yes, we have 2 documents which contain 'freedom'. You'll find them in the .frq file starting at byte 66991."
The .frq file says "document number 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', take a look at the .prx file starting at..."
If the searcher is only looking for 'freedom' in isolation, that's where it stops. It already knows enough to assign the documents scores against "freedom", with the 8-freedom document scoring higher than the single-freedom document.
Term vectors
KinoSearch's Term Vector support is in development, and this section is a placeholder.
Deletions
When a document is "deleted" from a segment, it is not actually purged from the postings data and the stored fields data right away; it is merely marked as "deleted", via the .del file. The .del file contains a bit vector with one bit for each document in the segment; if bit #254 is set then document 254 is deleted, and if it 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 optimization process that deleted documents truly go away.
Field Normalization Files
For the sake of simplicity, the example search scenario above omits the role played the field normalization files, or "fieldnorms" for short. These files have the (theoretical) suffix of .f followed by an integer -- .f0, .f1, etc. Each segment contains one such file for every indexed field.
By default, the fieldnorms' job is to make sure that a field which is 100 terms long and contains 10 mentions of the word 'freedom' scores higher than a field which also contains 10 mentions of the word 'freedom', but is 1000 terms in length. The idea is that the higher the density of the desired term, the more relevant the document.
The fieldnorms files contain one byte per document per indexed field, and all of them must be loaded into RAM before a search can be executed.
Document Numbers
Document numbers in Lucene/KinoSearch are ephemeral. They change every time a document gets moved from one segment to a new one during optimization. If you need to assign a primary key to each document, you need to create a field and populate it with an externally generated unique identifier.
Compatibility with Java Lucene
KinoSearch's file format is almost, but not quite, a Lucene ~1.9 compound index. If your data is 100% pure ASCII with no null bytes, files generated by Lucene and KinoSearch will be mutually compatible. However, if your data does not follow this strict standard to the letter, indexes written by Lucene will cause KinoSearch to crash and vice-versa.
The sample code in KinoSearch::Docs::Tutorial produces a Lucene-compatible index.
SEE ALSO
The technical specification of the Lucene index format is distributed along with Lucene. As of this writing, it is also located at http://lucene.apache.org/java/docs/fileformats.html.
See the section on Strings in the KinoSearch::Docs::DevGuide for a technical discussion of Lucene compatibility issues and efforts being made to resolve them.
COPYRIGHT
Copyright 2005-2006 Marvin Humphrey
LICENSE, DISCLAIMER, BUGS, etc.
See KinoSearch version 0.05_03.