[KPhotoAlbum] Quick search proof of concept patch

Shawn Willden shawn-kimdaba at willden.org
Sun May 13 08:51:45 CEST 2007


A lot of ideas related to my "quick search" notion were kicked around, but I 
felt like I needed to simplify it some to test out some ideas.  Attached is 
the not-so-quick-but-quite-dirty result.

I'd like to get feedback and performance results (the code is instrumented and 
prints some timing data to stdout).

One of my biggest concerns for the idea was performance.  Would it be possible 
to make it fast enough for interactive search-as-you-type usage, so I decided 
to focus on that first.  Besides, I had some ideas for a nifty approach to 
searching that would be fun to play with ;-)

This patch implements the basic idea with the following constraints:

1.  It only works for XML DB.
2.  It doesn't really manage the UI correctly; some stuff acts funny.
3.  It is slow in the "home" view, though performance seems good to me in the 
thumbnail view.
4.  It doesn't support quoting for searching for keywords with embedded spaces
5.  It doesn't support "OR" -- all search terms are ANDed together
6.  It does integrate with "drilldown" searches and date bar searches, by the 
simple expedient of applying the quick search first and then testing the 
remaining images against the drilldown and data bar criteria.
7.  It doesn't search on date or EXIF fields, only 'tags'.  It does, however, 
search all tags, including supercategories, case insensitively.
8.  It searches only by tag "prefix".  That is "foo" will 
match "foo", "foobar" and "footh", but not "myfoo".

In addition to all of that, the code is somewhat ugly, doesn't follow 
kphotoalbum style standards and generally needs a lot of work.

That said, my goal was to test the performance of my in-memory approach, and 
I'm quite happy with the results, at least with my database on my system, and 
I think it will also do well even on larger databases and slower machines.

In my case (Core Duo 2.2 GHz, 12.6K images, 1109 unique tags) building the 
ternary search tree used for the lookups takes about 700 ms, and the 
worst-case searches take under 40 ms.  Most searches take under 2 ms.

Please give it a try, especially those of you with big image databases and 
lots of tags.  To stress it a little, pick some single-letter searches that 
have large result sets.  "i" matches "Image" from the media type category (it 
probably doesn't make sense to index that one), so that's a big result set.  
Then put them together.  For example, I tried "i j c m e".  Those other 
letters are the first letters of my kids' names, so they have large result 
sets.

Oh, and do all this from the thumbnail view, with your full database showing.  
The search tree works just as well in the "home" view, but the process used 
to recalculate the tags per category is slow.

Thanks,

	Shawn.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: quicksearch_v_0_1.patch
Type: text/x-diff
Size: 29464 bytes
Desc: not available
Url : /mailman/pipermail/kphotoalbum/attachments/20070513/45cf4f58/attachment-0001.bin 


More information about the KPhotoAlbum mailing list