確率的データ構造を Java で扱いたい！ #JJUG

Transcript

1. ֬཰తσʔλߏ଄Λ Java Ͱѻ͍͍ͨʂ 2017-08-23 JJUG night seminar LT KOMIYA Atsushi

2. @komiya_atsushi

3. Today’s topic

4. ֬཰తσʔλߏ଄

5. ֬཰తσʔλߏ଄ͱ͸ʁ • ֬཰࿦తಛੑΛར༻ͨ͠σʔλߏ଄ • ͋Δ໰୊Λɺ࣌ؒత΋͘͠͸ۭؒతʹޮ཰Α͘
 (≅লϝϞϦͰ) ղ͘͜ͱΛ໨తͱ͢Δ • ࠓճ͸ʮۭؒޮ཰ͷΑ͍σʔλߏ଄ʯʹண໨ •

   σʔλߏ଄ʹΑͬͯ͸ɺݫີղͰ͸ͳۙ͘ࣅղ ͕ಘΒΕΔ͜ͱ͕͋Δ • ਫ਼౓ͱۭؒޮ཰͸τϨʔυΦϑͷؔ܎

6. ͲΜͳͱ͖ʹ࢖͏ͷ͔ʁ

7. ͲΜͳͱ͖ʹ࢖͏ͷ͔ʁ • ϦΞϧλΠϜ͔ͭେྔʹൃੜ͢ΔσʔλΛ
 ΦϯϥΠϯͰॲཧ͍ͨ͠ • ϝϞϦʹऩ·Γ͖Βͳ͍େن໛ͳσʔλΛ
 ඇྗͳ PC Ͱॲཧ͍ͨ͠ •

   ෼ࢄॲཧͰ͖Δ؀ڥ͕͋ΔͳΒɺ͋͑ͯ
 ֬཰తσʔλߏ଄Λ࢖͏ඞཁ͸ͳ͍

8. Java Ͱ ֬཰తσʔλߏ଄Λѻ͏

9. ࣗલ࣮૷ʁ ϥΠϒϥϦ࢖͏ʁ • ଟ͘ͷ֬཰తσʔλߏ଄͸ɺͦͷ࿦จ͕͙͙ Ε͹ӾཡՄೳͳঢ়ଶͰ͙͢ʹݟ͔ͭΔ • ͦΕΛಡΜͰࣗલ࣮૷͢Δͷ΋Α͠ • ҰํͰ Maven

   central ʹ͸͍ͭ͘΋ͷطଘ࣮ ૷͕ଘࡏ͍ͯ͠Δ • ڊਓͷݞͷ্ʹཱͭͷ͕ݡ͍΍Γํ

10. ֬཰తσʔλߏ଄ͷ Java ࣮૷ • stream-lib ‘com.addthis:stream-lib’ • Membership query /

    cardinality estimation / frequency counting / quantile estimation • Google Guava ‘com.google.guava:guava’ • Membership query • java-hll ‘net.agkn:hll’ • Cardinality estimation • t-digest ‘com.tdunning:t-digest’ • Quantile estimation

11. ֬཰తσʔλߏ଄ͷ Java ࣮૷ • stream-lib ‘com.addthis:stream-lib’ • Membership query /

    cardinality estimation / frequency counting / quantile estimation • Google Guava ‘com.google.guava:guava’ • Membership query • java-hll ‘net.agkn:hll’ • Cardinality estimation • t-digest ‘com.tdunning:t-digest’ • Quantile estimation

12. stream-lib ʹΑΔ ֬཰తσʔλߏ଄ͷར༻ํ๏

13. http://bit.ly/JJUG-2017-08- probds-code

14. Membership query

15. ཁૉ͕ू߹ʹଐ͢Δ͔൱͔Λ൑ఆ͢Δ

16. ཁૉ͕ू߹ʹଐ͢Δ͔൱͔Λ൑ఆ͢Δ Set<T> Λ༻ҙͯ͠ Set#contains(T) Ͱଘ൱Λ൑ఆ͠ Set#add(T) Ͱू߹ʹཁૉΛ௥Ճ͢Δ

17. Bloom filter • ֬཰తʹؒҧͬͨ౴͑ʢଘ൱݁ՌʣΛฦ͢ • ِཅੑ (ଘࡏ͠ͳ͍΋ͷΛଘࡏ͢Δͱޡೝ͢ Δࣄ৅) ͸ੜ͡Δ͕ɺِӄੑ͸ੜ͡ͳ͍ •

    ʮ૝ఆ͞ΕΔཁૉͷछྨ਺ʯ΍ʮڐ༰Ͱ͖Δِ ཅੑͷ֬཰ʯΛࢦఆͯ͠ɺώʔϓ࢖༻ྔΛ੍ޚ Ͱ͖Δ • ཁૉͷ௥Ճ͸Ͱ͖Δ͕ɺ࡟আ͸೉͍͠

18. stream-lib ͷ Bloom filter

19. stream-lib ͷ Bloom filter ཁૉ਺ͱِཅੑ֬཰Λࢦఆͯ͠ BloomFilter Λ༻ҙ͠ BloomFilter#isPresent(String) Ͱଘ൱Λ൑ఆ Set

    ͱಉ༷ʹ add() ͢Δ

20. ώʔϓ࢖༻ྔΛ֬ೝͯ͠ΈΔ • “Lorem ipsum” ͷςΩετΛྫʹɺJOL (Java Object Layout) Ͱώʔϓ࢖༻ྔΛଌఆ •

    http://openjdk.java.net/projects/code- tools/jol/ • Set: 6,032 bytes • stream-lib BloomFilter: 136 bytes 97.8% smaller !

21. Cardinality estimation

22. ҟͳΓ਺ΛٻΊΔ

23. ҟͳΓ਺ΛٻΊΔ Set<T> Λ༻ҙ͠ɺ Set#add() Ͱͻͨ͢ΒಥͬࠐΉ Set#size() ͰҟͳΓ਺͕ಘΒΕΔ

24. HyperLogLog++ (1/2) • ҟͳΓ਺Λਪఆ͢Δσʔλߏ଄ • ಘΒΕΔਪఆ஋͸ɺຊདྷͷҟͳΓ਺ʹର্ͯ͠ৼΕɾԼৼ Εͱ΋ʹى͜Γ͏Δ • Redshift /

    BigQuery / Presto ͳͲͰ΋ɺCOUNT(DISTINCT x) Λۙࣅ͢Δखஈͱͯ͠࢖ΘΕ͍ͯΔ • https://aws.amazon.com/jp/about-aws/whats-new/ 2013/11/11/amazon-redshift-new-performance-data- loading-security-features/ • https://cloud.google.com/blog/big-data/2017/07/ counting-uniques-faster-in-bigquery-with-hyperloglog

25. HyperLogLog++ (2/2) • ʮਪఆ஋ͷਫ਼౓ pʯΛௐ੔͢Δ͜ͱͰɺώʔϓ࢖༻ྔΛ੍ ޚ͢Δ͜ͱ͕Ͱ͖Δ • ஋Λେ͖͘͢Δͱਫ਼౓͕ߴ͘ͳΔ & ۭؒޮ཰͸ѱԽ͢Δ

    • ૝ఆ͞ΕΔҟͳΓ਺΍ඞཁͱ͞ΕΔਫ਼౓ɺώʔϓͷ੍໿ Λߟྀͯ͠ p Λܾఆ͢Δ • HyperLogLog ͷ࢓૊ΈΛཧղ͢Δʹ͸ɺҎԼͷϒϩάΤϯ τϦ͕͓͢͢Ί • http://blog.brainpad.co.jp/entry/2016/06/27/110000

26. stream-lib ͷ HyperLogLog++

27. stream-lib ͷ HyperLogLog++ ਫ਼౓Λࢦఆͯ͠ HyperLogLogPlus() Λ༻ҙ͢Δ HyperLogLogPlus#offer() ͰཁૉΛ௥Ճ͍ͯ͘͠ HyperLogLogPlus#cardinality() ͰҟͳΓ਺͕ಘΒΕΔ

28. Frequency counting

29. ཁૉͷස౓Λ਺্͑͛Δ

30. ཁૉͷස౓Λ਺্͑͛Δ Map Ͱཁૉ͝ͱͷΧ΢ϯλΛදݱ͢Δ ͻͨ͢Βཁૉ͝ͱʹ਺্͑͛Δ

31. Count-min sketch (1/2) • ཁૉͷස౓Λਪఆ͢Δσʔλߏ଄ͷҰͭ • ࣮ࡍͷස౓ΑΓ΋େ͖͍ਪఆ஋Λฦ͢͜ͱ͕ ͋ΔҰํͰɺখ͍͞ਪఆ஋Λฦ͢͜ͱ͸ͳ͍ • ස౓͕খ͍͞ཁૉ΄Ͳɺ͜ͷόΠΞεͷӨ

    ڹΛड͚΍͘͢ͳΔ

32. Count-min sketch (2/2) • width ͱ depth ͷೋͭͷύϥϝʔλͰɺۭؒ ޮ཰΍ਫ਼౓Λ੍ޚ͢Δ •

    width * depth ͷݸ਺ͷΧ΢ϯλ͕࡞ΒΕΔ • Χ΢ϯλ͸ 2࣍ݩ഑ྻͰදݱ • depth ͷ਺͚ͩϋογϡؔ਺͕࣮ߦ͞ΕΔͷ Ͱɺ଎౓తͳύϑΥʔϚϯεʹӨڹΛ༩͑Δ

33. stream-lib ͷ Count-min sketch

34. stream-lib ͷ Count-min sketch width:10 * depth:30 ͷΧ΢ϯλʹΑΔ Count-Min sketch

    Λ༻ҙ͢Δ CountMinSketch#add(String, int) ͰΧ΢ϯτ͍ͯ͘͠

35. Quantile estimation

36. ύʔηϯλΠϧ஋ΛٻΊΔ

37. ύʔηϯλΠϧ஋ΛٻΊΔ ιʔτ͞Εͨঢ়ଶͰ഑ྻԽ͢Δ ͋ͱ͸ n ύʔηϯλΠϧΛࢀর͢Δ͚ͩ

38. t-digest • ਺஋ྻͷ෼Ґ਺Λਪఆ͢Δσʔλߏ଄ • ܦݧ෼෍Λۙࣅతʹදݱ͢Δ • ύʔηϯλΠϧ஋͸ɺ͜ͷܦݧ෼෍ͷۙࣅදݱ͔Β಺ૠ Λ༻͍ͯࢉग़͞ΕΔ • ʮѹॖύϥϝʔλʯʹΑͬͯɺਫ਼౓ͱۭؒޮ཰ͷτϨʔυ

    ΦϑΛௐ੔͢Δ • ஋Λେ͖͘͢Δ͜ͱͰɺਫ਼౓ΛߴΊΔ͜ͱ͕Ͱ͖Δ

39. stream-lib ͷ t-digest

40. stream-lib ͷ t-digest ѹॖύϥϝʔλΛࢦఆͯ͠ TDigest Λ༻ҙ͢Δ TDigest#add(double) Ͱ਺஋Λ௥Ճ͍ͯ͘͠ TDigest#quantile(double) ͰύʔηϯλΠϧ஋ΛಘΔ

41. ·ͱΊ

42. ·ͱΊ • ֬཰తσʔλߏ଄Λ༻͍Δ͜ͱͰɺେن໛σʔλॲཧ΍ ΦϯϥΠϯॲཧΛޮ཰తʹ࣮ݱͰ͖Δʢ͔΋ʣ • Java Ͱ֬཰తσʔλߏ଄Λ͓खܰʹѻ͍͍ͨͳΒɺ
 ·ͣ͸stream-lib ͷར༻Λݕ౼ͯ͠ΈΔ •

    ਪఆਫ਼౓ͱۭؒޮ཰ͷτϨʔυΦϑΛ੍ޚ͢Δ
 ύϥϝʔλͷௐ੔͸ɺ৬ਓܳʹͳΓ͕ͪ • JOL ΍ JMH Λ༻͍ͯɺ࣮ࡍͷۭؒޮ཰ͱ࣌ؒޮ཰Λ ͖ͪΜͱଌఆ͠ͳ͕Βௐ੔͢Δ͜ͱΛ͓͢͢Ί͍ͨ͠

43. Thank you!