what's a useful definition of full text index on a repository?

Basically a "blue sky" question about full-text indexing git repositories.

A while back, whilst talking about overall git structure (see

http://marc.info/?l=git&m=118891945402778&w=2

), Jon Smirl raised the question of putting a full-text index on a repository. I doubt I full text index is of much use on a code repository because the question tends to be focussed around either released versions or immediate git-blame stuff. However, for repositories of things like evolving documents/presentations/notes where content is deleted because it doesn't fit the context rather than being superceded by better stuff, it might be more natural to search for "where's that paragraph I wrote on 'human' 'vision' and 'kinematic' 'feedback' ?". So I got to thinking about experimenting with a full text index and even started writing some code. However, I then realised that it's not obvious what the most useful definition of a full text index is on evolving files. (To be clear, I'm _not_ thinking about changing the database fundamentals as discussed in the referenced thread and indeed would put the full-text index into a different file that just references the existing git db stuff precisely because I doubt the text index will be of use to most people.)

A "classical" full-text index seems inappropriate because, if I've got a long text document that in a blob in commit n1 uses word 'x' in one section and the corresponding file in descendant commit n2 has the same text using word 'x' but has changes to a different section of the document, there's probably no point showing me both documents (and I can always track through the history once I've got one (commit-id,file pair)). So in the case of a single word, a "useful" definition would be the entry for word w in the full-text index should consist of those (commit-id,file) pairs whose diff with their parent contains an addition of text containing word w. (This will catch the creation of the text containing w and then precisely those files which are close modifications of it but ignore changes to other areas.) This seems to make sense for a single word. Let's call this the "appearance diff" definition of a full text index.

However, things become unclear if you consider a query with multiple words. Consider the simplest case of a linear history where commit n0 adds word "vision" to file p1.tex (with respect to its single parent), there are some intermediate commits and then commit n7 adds word "feedback" to p1.tex. Then there's no commit whose diff with its single parent contains both words "vision" and "feedback". In the linear history case you could imagine trying to find the first commit which is a child of _all_ the commits under the "appearance diff" definition. However, that clearly doesn't "obviously" extend to general full DAG histories, and in any case it's probably not fully correct even in the linear case. So maybe a different definition would be better. So I'm just throwing the question out to the list in case anyone has any better ideas for what a full-text index on an evolving set of files ought to be.

(One question is "why do you want to build a table rather than actively search the full git repo for each query (ie, combination of words) as you make it?" My primary motivation is that I might in the future like to do queries on some sort of low processor power UMPC-type thing, having built the file containing a "full text index" data structure for the index on a quite beefy desktop. The other point is that searching natural language text based on a fallible memory you're more likely to try different combinations of search terms iteratively to try and hit the right one, so there might be some point in trying to build an index.)

Anyway, it's currently an idle speculation,

On 10/1/07, David Tweed <david.tweed@gmail.com> wrote:

This is what full text is used for with code: http://lxr.mozilla.org/

It makes grep instant.

For source code you can take the full text concept further and store parse trees. This lets you instantly find the callers of a function, or all users of a variable.

Once you have parse trees in the database you can offer refactoring too. I have used powerful proprietary system that used parse trees to make complicated refactoring quite easy.

Note that a parse tree database doesn't have to be generated for all the old revisions, it is mainly usefully for the current HEAD. Same for the full text index. When you generate this data you end up with lots of tiny files that need to be kept in sync with the current HEAD. git is good for holding those files.

You want all this analysis coordinated with git so that when you commit a change the right parts get regenerated. Linux seems to be missing good, automated refactoring assistance. Instead the kernel janitorial work is being done manually.

An example I noticed today. Use of pr_debug() is chaotic in the kernel source. Many drivers have #if DEBUG their own versions of pr_debug(). Fixing everything to have consistent use of pr_debug is a refactoring that could probably be mostly automated.

Mozilla is undergoing some massive automated rewrites. http://blog.mozilla.com/tglek/category/decomtamination/

Full text indexing can also achieve high levels of compression as stated in the earlier threads. It is full scale dictionary compression. When it is being used for compression you want to apply it to all revisions.

You would full text index the expanded source text for each revision, not the delta. There are forms of full text indexes that record the words position in the document. They let you search for "vision NEAR feedback"

A good feature of this is finding when a variable or function was first added.

I do admit that these indexes are used to make functions that can be done with brute force faster. As computers get faster the need for these decrease. Right now the size of the kernel repo is not growing faster than the progress of hardware. If you went back are tried to do these things on a 386 you'd be shouting for indexes tomorrow.

On 10/1/07, Jon Smirl <jonsmirl@gmail.com> wrote:

I'd thought that keeping a full-text index of all my program files was my dirty little secret that shows I'm not a "pro" programmer ;-)

[details snipped]

This sounds interesting in principle but is beyond what I'm thinking in practice (particularly since I'm not in the "C is the only language worth ever using" camp).

Well, as I say I'm not convinced it makes sense to integrate this with existing pack stuff precisely because I don't think it's universally useful. So you seem to end up with all the usual tricks, eg, Golomb coding inverted indexes, etc, _if_ you treat each blob as completely independent. I was wondering if there was anything else you can do given the special structure that might be both more useful and more compact?

Well, the kind of question I was thinking was "clearly you can use the existing sort of full text indexing (eg, the stuff covered in Cleary, Witten & Bell's covered Managing Gigabytes), but is that the most useful way of doing things in the context of an evolving database?" If you treat every blob as essentially a different document there are indexing tools out there already you can use. What I was wondering was if it's really that useful to a human user to report every revision of a document containing those keywords even if the differences are in other parts of the text far removed from the text containing the keywords. I don't know the answer.

The other point is that direct searching is easier because you know exactly what the query is at the point you have access to the full text, whereas building an index you want to extract no more and no less information to be able to answer all allowed queries. But I still like the idea of getting a UMPC type thing if they become affordable.

On 10/2/07, David Tweed <david.tweed@gmail.com> wrote:

Dictionary compression can be used without full-text indexes. It is just really easy to build the full-text index if the data is already dictionary compressed. Dictionary compression works for everything except binary or random data.

Git is already using a small scale dictionary compressor via zip. I suspect doing a full scale dictionary for a pack file and then using arithmetic encoding of the tokens would provide substantially more compression. The big win is having a single dictionary instead of a new dictionary each time zip is used.

When we were working on Mozilla, Mozilla changed licenses three times. The license text ended up taking about 30MB in the current scheme. With full dictionary compression this would reduce down to a few kb.

More compression is good for git. It means we can keep more data in RAM and reduce download times. With current hardware it is almost always better to trade off CPU to reduce IO.