Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
知っておくと便利な Bloom Filter / Bloom Filter
Search
Yoshiaki Yoshida
October 14, 2016
Technology
2
6.4k
知っておくと便利な Bloom Filter / Bloom Filter
Blog /
https://kakakakakku.hatenablog.com/entry/2016/10/14/220055
Yoshiaki Yoshida
October 14, 2016
Tweet
Share
More Decks by Yoshiaki Yoshida
See All by Yoshiaki Yoshida
技術ブロガーを育てる!ブログメンタリングで何を教えているのか / Passion for Blog Mentoring
kakakakakku
8
37k
プログラミング初心者に教えるときは「身近な比喩」が重要なのだ! / Metaphor is Important for Beginner Programmer
kakakakakku
2
5.6k
プロジェクトの成功を支える ZenHub と モブプログラミング / ZenHub and Mob Programming
kakakakakku
1
5.8k
楽しく!アウトプットを習慣化しよう / Let's Enjoy Output
kakakakakku
3
6.7k
さぁ!今すぐプロジェクトリーダーに立候補しよう / Be a Project Leader
kakakakakku
3
9k
プロジェクトをリードする技術 (Kyash 社 再演) / Project Leading is Skill for Kyash
kakakakakku
4
2.1k
プロジェクトをリードする技術 / Project Leading is Skill
kakakakakku
43
46k
Mackerel で ECS をどこまでモニタリングできるのか / Monitoring ECS with Mackerel
kakakakakku
0
13k
[2018/01/30] Redash 初心者向けハンズオン / Redash Meetup #0.1
kakakakakku
0
2.4k
Other Decks in Technology
See All in Technology
GitHub MCP Serverを使って Pull Requestを作る、レビューする
hiyokose
2
600
Medmain FACTBOOK
akinaootani
0
170
2025年春に見直したい、リソース最適化の基本
sogaoh
PRO
0
270
マルチアカウント管理で必須!AWS Organizationsの機能とユースケース解説
nrinetcom
PRO
1
120
Agile TPIを活用した品質改善事例
tomasagi
0
600
3/26 クラウド食堂LT #2 GenU案件を通して学んだ教訓 登壇資料
ymae
1
240
AWSエンジニアがSAPのデータ抽出してみた
mayumi_hirano
0
110
Langchain4j y Ollama - Integrando LLMs con programas Java @ Commit Conf 2025
deors
0
100
やさしいMCP入門
minorun365
PRO
140
81k
OPENLOGI Company Profile for engineer
hr01
1
23k
近年の PyCon 情勢から見た PyCon APAC のまとめ
terapyon
0
270
ゆるくVPC Latticeについてまとめてみたら、意外と奥深い件
masakiokuda
2
200
Featured
See All Featured
The Art of Programming - Codeland 2020
erikaheidi
53
13k
Statistics for Hackers
jakevdp
798
220k
Large-scale JavaScript Application Architecture
addyosmani
511
110k
Gamification - CAS2011
davidbonilla
81
5.2k
We Have a Design System, Now What?
morganepeng
51
7.5k
A better future with KSS
kneath
239
17k
How GitHub (no longer) Works
holman
314
140k
Into the Great Unknown - MozCon
thekraken
36
1.7k
Scaling GitHub
holman
459
140k
The Cult of Friendly URLs
andyhume
78
6.3k
Building Flexible Design Systems
yeseniaperezcruz
328
38k
Making the Leap to Tech Lead
cromwellryan
133
9.2k
Transcript
͓ͬͯ͘ͱศརͳ Bloom Filter 2016-10-14 ࣾษڧձ @kakakakakku
Bloom Filter • 1970ʹߟҊ͞Εͨ • ߟҊऀ Burton Howard Bloom ࢯ
• ۭؒޮͷྑ͍֬తσʔλߏ • ཁૉ͕ू߹ͷதʹؚ·ΕΔ͔Λఆ͢ΔͨΊʹ͏ • σʔλྔʹґଘͤͣ O(k) ͷܭࢉྔͰߴʹఆͰ͖Δ
ू߹ ؚ·ΕΔ? ؚ·Εͳ͍? ؚ·ΕΔ? ؚ·Εͳ͍?
׆༻ྫΛΕ Bloom Filter Λ ͬͱۙʹײ͡ΒΕΔͣʂ
• Cassandra • Key Λݕࡧ͢Δͱ͖ͷ I/O Λݮ͢ΔͨΊ SSTable ʹ Bloom
Filter Λॻ͖ࠐΜͰ͍Δ • HBase • HFile ʹಛఆͷσʔλؚ͕·Ε͍ͯͳ͍͜ͱΛ ݕࡧ͢ΔͨΊʹ Bloom Filter Λ׆༻͍ͯ͠Δ Bloom Filter ׆༻ྫ 1
Bloom Filter ׆༻ྫ 2 • H2O • ແବͳαʔόϓογϡΛૹ৴͠ͳ͍ͨΊʹ ϒϥβΩϟογϡใͷ Bloom
Filter Λ ѹॖͯ͠ HTTP Ͱฦ͍ͯ͠Δ CASPER (Cache-Aware Server PushER) • Bitcoin • τϥϯβΫγϣϯͷݕࡧʹ׆༻͍ͯ͠Δ? ʢৄ֬͘͠ೝͰ͖ͯͳ͍ʣ
Bloom Filter ׆༻ྫ 3 • pixiv • ࡞ʹ͍ͨλάͱ pixiv ඦՊࣄయͷλάू߹ͷ
ଘࡏ֬ೝʹ׆༻͍ͯ͠Δ? http://inside.pixiv.net/entry/2014/07/22/132103
ʘ Bloom Filter ͷڍಈ ʗ
0 0 1 0 2 0 3 0 4 0
5 0 6 0 7 0 8 0 9 0 10 0 11 0 12 0 13 0 14 0 15 0 • m bit ͷྻΛ༻ҙ͢Δ • ࠓճ ྻۭؒ = m = 16 ͱ͢Δ • શͯ 0 ͰॳظԽ͓ͯ͘͠
0 0 1 0 2 0 3 0 4 0
5 0 6 0 7 0 8 0 9 0 10 0 11 0 12 0 13 0 14 0 15 0 • ҙͷϋογϡؔΛ༻ҙ͢Δ • ࠓճ k = 2 ݸͷؔΛ͏ • h1(key) = (key * 1) mod m • h2(key) = (key * 2) mod m
0 1 1 0 2 0 3 0 4 0
5 0 6 0 7 0 8 1 9 0 10 0 11 0 12 0 13 0 14 0 15 0 • key = 1000 ΛՃ͢Δ • h1(1000) = 8 • h2(1000) = 0 [ 1000 ]
0 1 1 0 2 1 3 0 4 0
5 0 6 0 7 0 8 1 9 1 10 0 11 0 12 0 13 0 14 0 15 0 • key = 1001 ΛՃ͢Δ • h1(1001) = 9 • h2(1001) = 2 [ 1000, 1001 ]
0 1 1 0 2 1 3 0 4 0
5 0 6 0 7 0 8 1 9 1 10 0 11 0 12 1 13 0 14 0 15 0 • key = 1004 ΛՃ͢Δ • h1(1004) = 12 • h2(1004) = 8 • h1(1000) = 8 ͱॏෳ͍ͯ͠Δ • ϑϥά 1 ͷ··ʹ͢Δ [ 1000, 1001, 1004 ]
0 1 1 0 2 1 3 0 4 0
5 0 6 0 7 0 8 1 9 1 10 0 11 0 12 1 13 0 14 0 15 0 • Query : key = 1005 ଘࡏ͢Δ ? • h1(1005) = 13 • h2(1005) = 10 • h1(1005) = h2(1005) = 0 • ʮଘࡏ͠ͳ͍ʯͱஅݴͰ͖Δ [ 1000, 1001, 1004 ]
0 1 1 0 2 1 3 0 4 0
5 0 6 0 7 0 8 1 9 1 10 0 11 0 12 1 13 0 14 0 15 0 • Query : key = 1000 ଘࡏ͢Δ ? • h1(1000) = 8 • h2(1000) = 0 • h1(1000) = h2(1000) = 1 • ʮଘࡏ͢Δʯ͔͠Εͳ͍ • 1000 ࣮ࡍʹଘࡏ͢Δ [ 1000, 1001, 1004 ]
0 1 1 0 2 1 3 0 4 0
5 0 6 0 7 0 8 1 9 1 10 0 11 0 12 1 13 0 14 0 15 0 • Query : key = 1020 ଘࡏ͢Δ ? • h1(1020) = 12 • h2(1020) = 8 • h1(1000) = h2(1000) = 1 • ʮଘࡏ͢Δʯ͔͠Εͳ͍ • 1020 ࣮ࡍʹଘࡏ͠ͳ͍ [ 1000, 1001, 1004 ]
(ƅшƅ) Űō? ޡఆͯ͠Δ͚Ͳ…?
False Positive ِཅੑ False Negative ِӄੑ ʮଘࡏ͠ͳ͍ʯͱ͖ʹʮଘࡏ͢Δʯͱఆͯ͠͠·͏͜ͱ ʮଘࡏ͢Δʯͱ͖ʹʮଘࡏ͠ͳ͍ʯͱఆͯ͠͠·͏͜ͱ
False Positive ِཅੑ False Negative ِӄੑ ʮଘࡏ͠ͳ͍ʯͱ͖ʹʮଘࡏ͢Δʯͱఆͯ͠͠·͏͜ͱ ʮଘࡏ͢Δʯͱ͖ʹʮଘࡏ͠ͳ͍ʯͱఆͯ͠͠·͏͜ͱ ↑ Bloom
Filter ʹ False Positive ͷՄೳੑ͕͋Δ
False Positive ͷՄೳੑ • O(k) ͰߴʹఆͰ͖Δঈͱͯ͠ False Positive ͷՄೳੑ͕͋Δ •
Αͬͯ key = 1020 ͷΑ͏ʹ ʮଘࡏ͢Δʯͱޡݕͯ͠͠·͏߹͕͋Δ • ͨͩ͠ False Negative 100% ͋Γಘͳ͍
ϝϦοτ • ܭࢉྔ O(k) • ઢܗ୳ࡧͩͱ O(N) • ೋ୳ࡧͩͱ O(log
N) • ϋογϡςʔϒϧͳΒ O(1) ͩͬ͠ͱߴ? • k = 1 ͳΒ Bloom Filter O(1) ʹͳΔ • σʔλΛอ࣋͢Δඞཁ͕ͳۭؒ͘ޮ͕ྑ͍
ʘ Bloom Filter ཁૉ͕আͰ͖ͳ͍ ʗ
0 0 1 0 2 1 3 0 4 0
5 0 6 0 7 0 8 0 9 1 10 0 11 0 12 1 13 0 14 0 15 0 • 1000 Λআ͢Δͱ • h1(1005) = 8 • h2(1005) = 0 • 1004 আ͞Εͯ͠·͏ !!! • h1(1005) = 12 • h2(1005) = 8 [ 1000, 1001, 1004 ]
ʘ ཁૉΛআ͢ΔͳΒ Counting Filter ʗ Bloom Filter Λ֦ுͨ͠ΞϧΰϦζϜ
0 1 1 0 2 0 3 0 4 0
5 0 6 0 7 0 8 1 9 0 10 0 11 0 12 0 13 0 14 0 15 0 • key = 1000 ΛՃ͢Δ • h1(1000) = 8 • h2(1000) = 0 [ 1000 ]
0 1 1 0 2 1 3 0 4 0
5 0 6 0 7 0 8 1 9 1 10 0 11 0 12 0 13 0 14 0 15 0 • key = 1001 ΛՃ͢Δ • h1(1001) = 9 • h2(1001) = 2 [ 1000, 1001 ]
0 1 1 0 2 1 3 0 4 0
5 0 6 0 7 0 8 2 9 1 10 0 11 0 12 1 13 0 14 0 15 0 • key = 1004 ΛՃ͢Δ • h1(1004) = 12 • h2(1004) = 8 • ॏෳͨ͠ΒΠϯΫϦϝϯτ͢Δ [ 1000, 1001, 1004 ] ϏοτͰͳ͘ ΧϯλʔͰදݱ͢Δ͕ Bloom Filter ͱҟͳΔ
0 0 1 0 2 1 3 0 4 0
5 0 6 0 7 0 8 1 9 1 10 0 11 0 12 1 13 0 14 0 15 0 • key = 1000 Λআ͢Δ • h1(1004) = 8 • h2(1004) = 0 • σΫϦϝϯτ͢Δ [ 1000, 1001, 1004 ]
ʘ False Positive ֬ ʗ
Bloom Filter ެࣜͰࢉग़ m : ྻۭؒ (bit) n : ొཁૉ
ৄ͘͠ Wikipedia ʹʂ https://ja.wikipedia.org/wiki/ϒϧʔϜϑΟϧλ False Positive Λ࠷খʹ͢Δ ࠷దͳϋογϡؔͷۙࣅ ࠷దͳ k Λͬͨ߹ͷ False Positive ֬
ʘ ৺ແ༻ ʗ ࠷దͳ k Λ͑ False Positive ΛݶΓͳ͘͘Ͱ͖Δ
ʘ ·ͱΊ ʗ False Positive ͷՄೳੑ͋Δ͠ ཁૉͷআͰ͖ͳ͍͚Ͳ τϨʔυΦϑΛ࠷େݶ׆༻ͯ͠ ߴ &
ۭؒޮͷྑ͍ॲཧ͕Ͱ͖Δʂ
ʘ ͓ͬͯ͘ͱศརͳ Bloom Filter ʗ