Hi,

I have a proof of concept inverted index based accelerator for guix
package search. The attached proof of concept script searches for
packages whose description contains a certain query string, and achieves
at least a 10x speedup (maybe more!) over brute force search.

Example proof of concept script run below:

Building index
clock utime stime cutime cstime gctime
 9.15 11.62  0.16   0.00   0.00   5.63

Brute force search
clock utime stime cutime cstime gctime
 0.12  0.12  0.00   0.00   0.00   0.00

Inverted index search
clock utime stime cutime cstime gctime
 0.00  0.00  0.00   0.00   0.00   0.00

* Method

We iterate over all package descriptions, decompose them into trigrams
(substrings of length 3), and build an "inverted index" -- a hash table
mapping trigrams to package names. Building this index needs to be done
only once, say during 'guix pull'. When the user provides a search
string, we decompose the search string into trigrams and use the
inverted index to look up packages whose descriptions contain those
trigrams. For typical search strings, this only has an average
complexity of O(1) whereas brute force search over all package
descriptions scales as O(n).

* Extension to regexp search

The attached proof of concept script only works for exact string
search. However, I can extend this to work for arbitrary regular
expressions (see references) if there is sufficient interest to have
this in Guix.

I think performance of guix package search will become even more
important as the number of packages in Guix grows. And, even as it is,
guix package search is relatively slow. Having blazing fast package
search could do wonders for user experience and user perception of
Guix's speed! :-)

WDYT?

* References

- https://en.wikipedia.org/wiki/Inverted_index
- http://wonconsulting.com/~cho/papers/cho-regex.pdf
- https://swtch.com/~rsc/regexp/regexp4.html

Regards,
Arun