Get the Lay of the (Lucene) Land

Elastic's Adrien Grand presents all the news you need to know about Apache Lucene at Elastic{ON}16 in February 18, 2016 in San Francisco.

Transcript

1. ‹#› Adrien Grand @jpountz Get the Lay of the (Lucene) Land
2. Working with Lucene since 2010 Lucene committer since 2012 Lucene PMC since 2013 Elastic employee since 2013 19
3. ‹#› Apache Lucene is a free and open-source information retrieval software library Wikipedia
4. Many use-cases 4 Full-text search
5. Many use-cases 5 Full-text search Structured search Analytics Geo search Suggestions Data store
6. Segment 6 Index structure inverted index column store document store norms Since 4.0
7. ‹#› Faster structured search Powered by k-d trees
8. Segment 8 Index structure inverted index column store document store norms k-d tree New in 6.0 Since 4.0
9. 9 Efficient structured search on an inverted index 000 001 010 011 100 101 110 111 000 001010 011100 101110 111 00 01 10 11 00 01 10 11 0 1 0 1 • Index data with several levels of precision • (1,0) • (10, 01) • (101, 010)
10. 10 Efficient structured search on an inverted index 000 001 010 011 100 101 110 111 000 001010 011100 101110 111 00 01 10 11 00 01 10 11 0 1 0 1 • Search by visiting as few cells as possible • (01,01) • (10,01) • (010,100) • (011,100) • (100,100) • (101,100)
11. 11 Efficient structured search on an inverted index 000 001 010 011 100 101 110 111 000 001010 011100 101110 111 00 01 10 11 00 01 10 11 0 1 0 1 • Search by visiting as few cells as possible • (01,01) • (10,01) • (010,100) • (011,100) • (100,100) • (101,100)
12. 12 Efficient structured search on an inverted index 000 001 010 011 100 101 110 111 000 001010 011100 101110 111 00 01 10 11 00 01 10 11 0 1 0 1 • Search by visiting as few cells as possible • (01,01) • (10,01) • (010,100) • (011,100) • (100,100) • (101,100)
13. 13 Efficient structured search on an inverted index 000 001 010 011 100 101 110 111 000 001010 011100 101110 111 00 01 10 11 00 01 10 11 0 1 0 1 • Search by visiting as few cells as possible • (01,01) • (10,01) • (010,100) • (011,100) • (100,100) • (101,100)
14. 14 Efficient structured search on an inverted index 000 001 010 011 100 101 110 111 000 001010 011100 101110 111 00 01 10 11 00 01 10 11 0 1 0 1 • Search by visiting as few cells as possible • (01,01) • (10,01) • (010,100) • (011,100) • (100,100) • (101,100)
15. 15 Efficient structured search on an inverted index 000 001 010 011 100 101 110 111 000 001010 011100 101110 111 00 01 10 11 00 01 10 11 0 1 0 1 • Search by visiting as few cells as possible • (01,01) • (10,01) • (010,100) • (011,100) • (100,100) • (101,100)
16. 16 Efficient structured search on an inverted index 000 001 010 011 100 101 110 111 000 001010 011100 101110 111 00 01 10 11 00 01 10 11 0 1 0 1 • Search by visiting as few cells as possible • (01,01) • (10,01) • (010,100) • (011,100) • (100,100) • (101,100)
17. K-D trees 17
18. K-D trees 18
19. K-D trees 19
20. K-D trees 20
21. K-D trees 21
22. K-D trees 22
23. K-D trees 23
24. K-D trees 24
25. K-D trees 25
26. K-D trees 26
27. K-D trees 27
28. K-D trees 28
29. K-D trees 29
32. 32 More information tomorrow at 11am Geospatial Data Structures in Elasticsearch and Apache Lucene by Nick Knize
33. ‹#› Better relevance BM25 by default
34. 34 Doc freq contribution • Common words are naturally discriminated • No more need to exclude stop words TF-IDF BM25
35. 35 Term freq contribution • Saturated tf contribution TF-IDF BM25
36. 36 More information tomorrow at 12:45 pm Improved Text Scoring with BM25 by Britta Weber
37. ‹#› Slow queries got faster thanks to the match cost API
38. 38 Two-phase iteration • Query can be divided into • a fast approximation • a (slower) match • Phrase • Approximation = conjunction: • Match = position check • Geo polygon query • Approximation = points in cells that are in or cross the polygon • Match = check the point against the polygon
39. Match cost API 39 description:”search engine” AND body:”postings list” (description:search AND description:engine) AND (body:postings AND body:list) 1. Approximate 2. Match description:”search engine” body:”postings list”
40. Match cost API 40 description:”search engine” AND body:”postings list” Field Value df ttf description search 200k 420k description engine 15k 18k body postings 1k 13k body list 370k 5920k
41. Match cost API 41 description:”search engine” AND body:”postings list” Field Value df ttf description search 200k 420k description engine 15k 18k body postings 1k 13k body list 370k 5920k
42. Match cost API 42 description:”search engine” AND body:”postings list” 1. Approximate 2. Match description:”search engine” body:”postings list” 1. Iterate body:postings (1k) 2. Check description:engine (15k) 3. Check description:search (200k) 4. Check body:list (370k)
43. Match cost API 43 description:”search engine” AND body:”postings list” Field Value df ttf ttf/df description search 200k 420k 2.1 description engine 15k 18k 1.2 body postings 1k 13k 13 body list 370k 5920k 16
44. Match cost API 44 description:”search engine” AND body:”postings list” 1. Iterate body:postings (1k) 2. Check description:engine (15k) 3. Check description:search (200k) 4. Check body:list (370k) 1. Approximate 2. Match 1. description:”search engine” (2.1+1.2=3.3) 2. body:”postings list” (16+13=29)
45. 45 Other changes • Better query-time synonyms • Disk-based norms • More memory-efficient doc values • More disk-efficient sparse doc values • BooleanQuery simplification in rewrite • Bulk scorer specialization for MatchAllDocsQuery and MUST_NOT clauses • Improved file truncation detection
46. oh, and we moved from svn to git! 46
47. ‹#› Thank you! Questions?
48. ‹#› Please attribute Elastic with a link to elastic.co Except where otherwise noted, this work is licensed under http://creativecommons.org/licenses/by-nd/4.0/ Creative Commons and the double C in a circle are registered trademarks of Creative Commons in the United States and other countries. Third party marks and brands are the property of their respective holders. 48