Learn Ractor

Transcript

1. Learn Ractor
   
   
   [email protected]
   

   View Slide

2. E-book version still available
   The dRuby Book is out of print, but...
   スタンプラリー対象外です。Not eligible for stamp rally
   2
   

   View Slide

3. Agenda
   Concurrency is everywhere
   
   
   Introduction to Ractor
   
   
   case study
   3
   

   View Slide

4. Agenda
   Concurrency is everywhere
   
   
   About me
   
   
   Concurrency is everywhere
   
   
   Introduction to Ractor
   
   
   case study
   4
   

   View Slide

5. About me
   Masatoshi Seki, Ruby Core Committer (dRuby, Rinda, ERB), Programmer
   
   
   I started programming 40 years ago
   
   
   writing complex embedded systems for 30 years
   
   
   system with concurrency
   40年くらいコード書いてて、30年くらいは複雑なシステム書いてるよ
   5
   

   View Slide

6. Complex embedded systems
   dRuby
   in the real-world embedded systems.
   ΩϠϊϯϝσΟΧϧγεςϜζגࣜձࣾ
   ҩ༻ίϯϙʔωϯτٕज़։ൃηϯλʔɹؔকढ़ɹԂ઒ཾ໵
   4FQ ɹ
   - ref. RubyKaigi takeout 2021
   ミニマム構成の製品を使って若者をRubyistにした事例発表
   About us
   Masatoshi Seki, Ruby Core Committer (dRuby, Rinda, ERB), Programmer
   Tatsuya Sonokawa, Programmer
   Miwa Fukaya, Tochigi RubyKaigi Content Committee Chair, Software
   tester
   ਓؒυοΫޙͷ࠶ݕࠪͰࡱͬͯ΋ΒͬͨࣸਅͰ͢ɻେਓͳͷͰɻ
   9
   404
   

   View Slide

7. Complex embedded systems
   dRuby
   in the real-world embedded systems.
   ΩϠϊϯϝσΟΧϧγεςϜζגࣜձࣾ
   ҩ༻ίϯϙʔωϯτٕज़։ൃηϯλʔɹؔকढ़ɹԂ઒ཾ໵
   4FQ ɹ
   - ref. RubyKaigi takeout 2021
   X-ray CT image of my gallstones
   About us
   Masatoshi Seki, Ruby Core Committer (dRuby, Rinda, ERB), Programmer
   Tatsuya Sonokawa, Programmer
   Miwa Fukaya, Tochigi RubyKaigi Content Committee Chair, Software
   tester
   ਓؒυοΫޙͷ࠶ݕࠪͰࡱͬͯ΋ΒͬͨࣸਅͰ͢ɻେਓͳͷͰɻ
   9
   404
   Gallstones
   

   View Slide

8. Complex embedded systems
   Lots of processes, and explicit and implicit IPC
   無数のプロセスとプロセス間通信がある。明示的だったり暗黙的だったりするよ
   8
   

   View Slide

9. Concurrency is everywhere
   Concurrency exists at various levels of abstraction
   そこら中で時間的にオーバーラップして実行してる。ということはつまり...
   9
   

   View Slide

10. Concurrency is everywhere
    Sequential execution is a special case
    むしろ、逐次実行は特殊なケースである
    10
    

    View Slide

11. RubyKaigi takeout 2020
    Rinda
    in the real-world embedded systems.
    [email protected]
    First printing from 2005 (Still available)
    dRuby
    ʹΑΔ
    ؔকढ़ஶ
    ෼ࢄ
    ɾ
    Web
    ϓϩάϥϛϯά
    dRubyとRindaによる天文台のロボット望遠鏡（ガチ）の話
    - 🤖🔭
    

    View Slide

12. RubyKaigi takeout 2020
    自律して動作する装置／センサー群を統合するシステム。抽象度が高い感じする
    - Linda coordination model
    software block diagram
    1
    psz117
    X
    f/2.5
    20151207_yanagisawa.pdf ΑΓ
    OAOWFC䛾እほ
    2015/12/07 奛㷔塭伖㈨埻WS 16
    X
    

    View Slide

13. OpenMP
    比較的低いレベルの例
    13
    int main(int argc, char *argv[])
    
    
    {
    
    
    int i;
    
    
    #pragma omp parallel for
    
    
    for (i = 0; i < 10000; ++i)
    
    
    {
    
    
    /* ... */
    
    
    }
    
    
    return 0;
    
    
    }
    

    View Slide

14. OpenMP
    14
    0...10000
    int main(int argc, char *argv[])
    
    
    {
    
    
    int i;
    
    
    #pragma omp parallel for
    
    
    for (i = 0; i < 10000; ++i)
    
    
    {
    
    
    /* ... */
    
    
    }
    
    
    return 0;
    
    
    }
    

    View Slide

15. OpenMP
    15
    0...2000
    2000...4000
    4000...6000
    6000...8000
    8000...10000
    

    View Slide

16. Agenda
    Concurrency is everywhere
    
    
    Introduction to Ractor
    
    
    case study
    16
    

    View Slide

17. Ractor
    parallel
    
    
    avoid race condition
    
    
    isolation
    本当のスレッドと保護されたメモリー（オブジェクト）で
    17
    

    View Slide

18. Thread to Ractor
    単純なケースではThreadのように使えるよ
    18
    >> t = Thread.new(*arg) {|*arg| ... }
    
    
    ...
    
    
    >> t.value
    >> r = Ractor.new(*arg) {|*arg| ... }
    
    
    ...
    
    
    >> r.take
    

    View Slide

19. Thread to Ractor
    Ractorはtakeで値を取り出すのがちがうよ。
    19
    >> t = Thread.new(25) {|n| (1..n).inject(1, :*)}
    
    
    ...
    
    
    >> t.value
    
    
    => 15511210043330985984000000
    >> r = Ractor.new(25) {|n| (1..n).inject(1, :*)}
    
    
    ...
    
    
    >> r.take
    
    
    => 15511210043330985984000000
    

    View Slide

20. avoid race conditions
    isolation
    
    
    block scope
    
    
    global variables
    
    
    shareable object
    競合状態を嫌ってRactor間をまたぐオブジェクトに強い制限がある
    20
    

    View Slide

21. isolation
    block scope
    Procだけど、外側の変数にはアクセスできないよ
    21
    >> hash = {'foo'=>'bar'}
    
    
    >> r = Ractor.new {hash['foo']}
    
    
    :271:in `new': can not isolate a Proc
    because it accesses outer variables (i). (ArgumentError)
    

    View Slide

22. isolation
    global variable
    グローバル変数にはアクセスできないぞ
    22
    >> $foo = 1
    
    
    >> r = Ractor.new {$foo}
    
    
    # terminated with exception
    (report_on_exception is true):
    
    
    (irb):11:in `block in ': can not access
    global variables $foo from non-main Ractors
    (Ractor::IsolationError)
    

    View Slide

23. pass by value
    deep copy
    引数経由して渡せるよ。でもRactorをまたぐときに複製になる。dRubyみたいだ
    23
    >> hash = {'foo'=>'bar'}
    
    
    >> r = Ractor.new(hash) {|it| it['foo'] = 'baz'}
    
    
    ...
    
    
    >> r.take
    
    
    => "baz"
    
    
    >> hash
    
    
    => {"foo"=>"bar"}
    
    
    >> r = Ractor.new(hash) {|it| it.object_id}
    
    
    ...
    
    
    >> [hash.object_id, r.take]
    
    
    => [2070360, 2006160]
    

    View Slide

24. shareable
    Ractor.make_shareable
    make_shareableするとRactor間で共有されるぞ！でもfreezeされてしまうのだ
    DRbUndumpedみたい
    24
    >> hash = {'foo'=>'bar'}
    
    
    >> Ractor.make_shareable(hash)
    
    
    >> Ractor.new(hash) {|it| it.object_id}.take
    
    
    => 221520
    
    
    >> hash.object_id
    
    
    => 221520
    
    
    >> hash['foo'] = 'baz'
    
    
    (irb):16:in `': can't modify frozen Hash:
    {"foo"=>"bar"} (FrozenError)
    

    View Slide

25. can't make shareable
    Queue, Monitor, TupleSpace
    同期プリミティブは渡せない
    25
    >> Ractor.make_shareable(Queue.new)
    
    
    :820:in `make_shareable': can not make
    shareable object for #
    (Ractor::Error)
    
    
    >> Ractor.make_shareable(Monitor.new)
    
    
    :820:in `make_shareable': can not make
    shareable object for #
    (Ractor::Error)
    
    
    >> Ractor.make_shareable(Rinda::TupleSpace.new)
    
    
    :820:in `make_shareable': can not make
    shareable object for #
    (Ractor::Error)
    

    View Slide

26. shareable object
    immutable
    
    
    class/module
    
    
    special object
    
    
    Ractor
    
    
    ENV
    26
    

    View Slide

27. Ractor has two queues.
    takeしか使わなかった...
    27
    incoming queue
    
    
    Ractor#send / Ractor.receive
    
    
    outgoing queue
    
    
    Ractor.yield / Ractor#take
    

    View Slide

28. avoid race conditions
    Using Ractor is easy, but writing for Ractor is hard.
    
    
    straitjacket
    拘束衣。オブジェクト指向プログラミングのように書くのはたいへんだと思う
    28
    

    View Slide

29. PG
    Thought it wouldn't work, but it didn't
    動かないと思うけど動かないもの
    29
    >> conn = PG.connect("postgres:///masaki")
    
    
    >> Ractor.make_shareable(conn)
    
    
    :820:in `make_shareable': can not make shareable
    object for # (Ractor::Error)
    
    
    >> Ractor.new { PG.connect("postgres:///masaki").exec("select 1") }
    
    
    # terminated with exception
    (report_on_exception is true):
    
    
    /opt/homebrew/lib/ruby/gems/3.2.0/gems/pg-1.4.6/lib/pg/
    connection.rb:70:in `conninfo_parse': ractor unsafe method called
    from not main ractor (Ractor::UnsafeError)
    
    
    

    View Slide

30. pp
    Thought it would work, but it didn't
    動きそうで動かなかったもの。私の使い方が悪いのかもしれないです。
    30
    >> Ractor.new {pp 1}
    
    
    # terminated with exception
    (report_on_exception is true):
    
    
    core_ext/kernel_require.rb>:40:in `require': can not access
    non-shareable objects in constant
    Kernel::RUBYGEMS_ACTIVATION_MONITOR by non-main ractor.
    (Ractor::IsolationError)
    

    View Slide

31. a situation suitable for Ractor
    Accelerate loops with parallelism
    
    
    numeric calculation
    
    
    in-memory database
    さっき説明したOpenMP見たい領域が得意そう。
    31
    

    View Slide

32. There is an in-memory search engine
    用意しておきました！
    32
    

    View Slide

33. Agenda
    Concurrency is everywhere
    
    
    Introduction to Ractor
    
    
    case study
    33
    

    View Slide

34. RubyKaigi 2022
    Create my own search engine.
    [email protected]
    ࣗ෼༻ͷݕࡧγεςϜΛ࡞Δ࿩
    - There is an in-memory search engine
    dRubyについて話せなかったが、ポケモンカードの検索エンジンの話
    

    View Slide

35. Pokémon
    Trainer Energy
    Pokémon TCG
    Build a deck with 60 cards
    

    View Slide

36. Pokémon TCG similar deck search
    Heroku → Oracle Cloud Infrastructure
    
    
    VM : Arm x 4 core
    
    
    do not use containers
    OCIに移行したので、コンテナを使うのもやめた
    36
    2023
    

    View Slide

37. Weekly deck trend analysis
    新機能！
    37
    case 1
    

    View Slide

38. Hierarchical clustering
    
    
    Calculate distances for all combinations
    
    
    Merge closest node into cluster
    38
    

    View Slide

39. Hierarchical clustering
    
    
    Calculate distances for all combinations
    
    
    Merge closest node into cluster
    39
    

    View Slide

40. Hierarchical clustering
    
    
    Calculate distances for all combinations
    
    
    Merge closest node into cluster
    m
    C2
    40
    

    View Slide

41. Hierarchical clustering
    
    
    Calculate distances for all combinations → mCn
    m
    C2
    41
    irb(main):001:0>
    Array.new(6).combination(2).size
    
    
    => 15
    
    
    irb(main):002:0>
    Array.new(900).combination(2).size
    
    
    => 404550
    

    View Slide

42. Hierarchical clustering
    
    
    Calculate distances for all combinations
    
    
    Merge closest node into cluster
    42
    

    View Slide

43. Hierarchical clustering
    
    
    Calculate distances for all combinations
    
    
    Merge closest node into cluster
    43
    

    View Slide

44. Hierarchical clustering
    
    
    Calculate distances for all combinations
    
    
    Merge closest node into cluster
    44
    

    View Slide

45. Hierarchical clustering
    
    
    Calculate distances for all combinations
    
    
    Merge closest node into cluster
    45
    

    View Slide

46. Hierarchical clustering
    0.0
    0.1
    0.2
    0.3
    0.4
    0.5
    0.6
    0.7
    0.8
    ↑ Height
    vVkVkf-CfQ2Me-5wF5k5
    pM2ySp-NqdVVy-SEXyyy
    Ffvkv5-urPxNL-VkFdvk
    x888Y4-ZEG7rl-8a8c88
    pySESM-DfgV3e-EpSpEy
    QggnnQ-rTu8JY-iingnn
    cY88a8-41UAjT-88Gc4D
    888Y8K-M5mgp6-Y4cxcD
    pyySSU-fpvTh7-Upy2XM
    cY48J8-P6fNIl-Kxac8J
    HHnQnL-P97pwl-PHgnng
    MSpSpy-vAWjQl-MySMpS
    fkFkFF-zaC6UD-v1kkbw
    pSypUM-wkyESS-p2yyM3
    8YDcxc-feOc08-4GY884
    kFk1Fv-0TdM4n-dv5kkV
    Yx8ax4-JDBmvc-cYG8c8
    SSyXy2-N8v9C7-y2p3yp
    pXy3yy-9MrduS-yUpSy3
    Fdv1k5-WkoeFa-FkvVkk
    84888c-xCG2yN-Y8cGD8
    wkVvFV-nyF8eL-FVkkfk
    xa8a88-dom7r2-x8Y4cx
    gnngiP-TQ1apH-nPnnQn
    nNL9gn-eoFUGm-nNnngQ
    fVwwVw-ubXExr-FFkkv5
    QQn6ng-IKDkHa-nNiNnn
    Y8ccYa-GItA31-cY488x
    NHnnPg-jQhceA-nQnnLL
    yEUMRy-2mG9g3-pSSMyM
    Fkk1Vw-d4Wozk-FFvvFf
    6QQg9n-AbEJHX-giPngn
    gN6nQn-ckH59n-9inQnL
    D88xY8-hpaIDr-8cxcYc
    LQggQn-hv20uI-HnnnQg
    YcK48x-cDdQ0X-xxGD8D
    nHL6Hg-knDcUp-6nQngQ
    SSMUpp-INYSgw-ppyySy
    nng96n-lBix3f-LQQLgg
    SMRSyy-GHEWTP-MyXpyE
    Vbkv1k-tOY61g-F5kvVk
    PnnnLn-PEaw1c-QLnngg
    pypSpy-c9cBp2-SSyypp
    XyyyyX-TrT2ek-SpyMpy
    LHHgHQ-mLxVMM-nngnn9
    kVFkFb-AN5NWI-kvvfFb
    kv1v1f-vaYx04-kFfkkV
    x88J88-jBN0Jj-8Yc8Kx
    ySM3yy-gnhe17-XyUSyp
    8xYD8D-npQXH8-KYYcxY
    Y4Y84x-MrJeFq-cDD488
    nnnNgQ-VNO3JO-igNnNg
    5FFk1k-hWutyR-vwVk5F
    Q6nLNN-H90bPb-nLginn
    pyySS2-H7DN4A-pUSyRy
    SEpEyy-hbxATA-SyyppM
    YKc8xY-R3feUE-8cc8a8
    g69ngL-gWWKti-QinnnH
    nLgNQn-JVUT8W-HPHQLn
    D8Yx88-lrJIuu-JD84c4
    kv55F5-3erdMp-VkFFwk
    v5VFf1-UKYbZn-VbVkkk
    Y48Y8a-RJ2LDC-8c8xcD
    pS2y2p-ZSaQvC-MSUyyy
    Sypyy2-K6PSF4-MppXXy
    K8c8YK-4jx3NW-8Dac48
    XyRyyU-qrp4V9-pMRySM
    J888cc-lxoCeB-4DYJ8c
    giPgni-OJ09Yi-gngnnQ
    Y84cxx-3rJwVM-cYY84c
    ngiQgN-CnQ4gv-6nQngn
    kkkFVb-zoKBee-vFkFvk
    6QgHnn-2VFveI-QnnngH
    3yRSyy-D8i0T9-yy2ppM
    gnLnQQ-Cyt2rM-L6nnLN
    MUpSyX-xcRVC3-pyyRpS
    dkvkfk-77tnCm-FFVfFV
    F1kkFf-MVW9eT-vkvFkd
    kkkFkk-xYY2UD-kvw15V
    nngHng-OBmOQL-NQHLQL
    8xKD84-DzKFZJ-4xK8cJ
    Q9ngLn-bjLzmZ-gQnQgQ
    NPnLgn-58T69t-nQnniH
    8GD84D-DuK5jM-xYc84x
    Vk5vkk-bcQSA0-FkvFkv
    LLQnQn-cTWB0Y-gNQNnn
    nLnQnn-mtbSBC-ngQLHL
    6ngniN-9Dj0L4-NninnQ
    iinngg-njYbYw-ngnNHQ
    ypUyy3-3rxk2j-p3pSy2
    SpyyRM-Cq2rZQ-UyyMEy
    55kvkF-p1hUSP-kFkFkb
    FkFVdF-Clq0WL-kvv1kd
    gnLngg-3NcLDa-Q9nPPN
    1FVkvk-2dvvTK-kkdkVF
    Mypyyp-g0z6Cl-EMpSSy
    nnngQQ-QwK7at-9P9gnn
    cxY8G8-59CrtL-8c488x
    9LNQg9-kuPro1-ggnQnn
    D888Dx-mAJVcL-KaYx8c
    LLLgnn-VdBYdx-gLnNnQ
    gHniQg-bmCra9-Lnggni
    Eyp23X-6lKVUo-MpppyS
    nPLgPN-2Vu2fH-gLgnQL
    x8Y48c-XDRyTC-KYaccc
    x8Y8Y8-LANoAw-88Jc8D
    acY88c-uS50Np-8xYxYc
    x88DYD-yXtIEy-8cJ88x
    8Y4ccx-kwQSDk-48x8DG
    gngHng-dMAxgW-n6nQQ6
    LPNnLn-G0fVYk-iQNnng
    NQLnnN-vDiFBq-nP9gNn
    Fvbkff-MmzaQA-kvVwVk
    c88Yc4-YeMfSt-8c48ca
    kkvdVV-OgMIQX-kFkFkk
    48YKYc-HRXkZA-Ya8c88
    nnQng6-DRu6oG-gnigin
    kFvkkv-iKWU1R-kbVvfk
    Fv5FFk-HQjlQO-kwkvkv
    LngLgN-Sdu3m1-nNLnQn
    SyMyy2-9owvQk-SypXMp
    kkvVfk-Q24Ze6-kkfFvF
    gHing9-jvXlft-QgLQni
    8YaxDD-5uPjMv-c84xKc
    gLnQQn-oProA3-gLQgn9
    gngQgg-bLqQCo-g6nQn6
    8xca8Y-YGIUkZ-8x8Ycx
    nnn6NQ-oUJKe0-HQLngg
    D8GcY8-rW9jGo-Yxx8cK
    SypypS-g77hSn-MypUyp
    E2My2p-GEKcv4-pUySyR
    gQQNgL-oBkSWV-LQnPnn
    FkdVVv-O2oHwv-bbvkd5
    88xxGa-Og85PD-c8Ya8D
    EMSyp2-2fQqfN-yyESUp
    YY84c8-RR4n4d-8c8xJx
    gnLnQg-euTM2D-gnnHQL
    xcc4YJ-8JSn76-Y8KDc8
    Y8888c-GraJIz-x4Kcc4
    c88x8Y-RdPWMA-88cD8Y
    k5FV1v-975h5I-FVFkkv
    LPNgQn-3q0i7b-6ngHNn
    nggnn9-4PJZ5K-gLQHng
    8YJ88K-iGHqSA-xcxDc8
    g9QngQ-trb93k-NnnHng
    pS3ypy-0GDmyP-yXMSUp
    Y8GJx8-c3qvQb-KY4Kcx
    bFkdkF-QDqqYK-vFkbbk
    QQgPnn-vVzh0f-gnNnLg
    ySyyR2-K40qf2-pXyMpp
    kkVFkF-2fCbID-vkdFFV
    x8Y8K8-t24oZW-acYx8c
    8Y8Jc8-lxt7oi-x84cJY
    LLinnn-WGGLjk-LQnQgn
    EXpyMM-kVHZFw-MXyS2y
    88xYDc-Utex6g-8cxx8c
    DKY884-G2Uobg-8aJcxx
    Eyp2yp-X8YjPR-3pSySU
    F5vVVv-acRcO1-FkFkdF
    cx8xJx-I5OlcH-888caY
    XRXSMy-00j6oU-pyppyX
    LnQnnH-mXhDzP-NgHngQ
    5vvFkk-lOTpux-Vvkffk
    xKJcY8-k6OH37-8D8c8c
    VkvV5F-fZPXeQ-FkvvFV
    FvFwvw-PO6yND-kfbVFk
    vkb1kk-JjQzd4-VfFFfb
    p3SXEp-zAnOdf-yMSyXp
    6nQgLg-796wEi-9innHL
    ngnNNP-YJpVX9-n9QnHN
    SXXppM-Aio17g-3MyMyE
    gQLgng-qJnyrq-nninQP
    6nnQnn-ZbOkdT-LPLggN
    FkwvVk-0JXBtY-Vk1kv5
    9n9nnQ-LG4gBy-L9LnQg
    dvkkFk-nu07S3-FFF1df
    L6Ln6i-ecLiWp-NgQQng
    p2pyMy-YUHLGz-pMpU2S
    8cY48D-Kqr0g6-8cJ88D
    xD8c8K-1V2oBq-Y8aac8
    1v1VkF-lwRkxD-FkVvkk
    cY8D8c-zdx6Xc-Y8cx8a
    a8KcKc-hzlU7N-8Dcx8Y
    y2pyMX-4GZa8x-MMyS2M
    yyppMX-i88khM-pS3SSU
    kkFk5k-pQFm92-kvvkFv
    bFFvfV-LdQ9FU-vk5kwF
    gnnHHP-3TnKB4-QngLQg
    ngnQnQ-Kr6ml6-LLgNnH
    x84YJc-BKoP5V-c8cx8D
    YcD84c-CKpkih-8K8DY4
    8YDcx8-kAMY8T-D448cY
    kvwkFk-OCov4S-kwvVVf
    gnnnNg-lpf6hT-gLnQQg
    88xYxc-3H7q7f-J8YaYY
    Xyypyp-qFmtto-My2ySE
    nQngPP-57dCZZ-NnnngL
    NLnnQg-WXipiX-LNgQin
    w5kbVk-XttGD8-vkkFkF
    4cGx8G-S0wNBC-8Ycc4x
    8GxcYx-OH2UUH-8cc48a
    SyXyyp-CI48IF-XMyMpy
    6nN6nL-KBuiUY-nggQnL
    ax8cY8-6q8I5j-aKY4Yx
    F1vFvd-BqFbRu-FkfkFk
    XMpySy-Pdzy6z-yMEM2p
    Fwkbkk-uMrdbS-vVfFkk
    2MypyM-gQ7P5G-XySEMy
    QLgQLn-1nKVTu-gQQQnH
    Spy2yM-4Nvzt6-URyyXU
    ypUpXM-ztrfJL-RyyypS
    8ccJcY-kkMIDl-8Kc8a8
    gLLggn-yvCK39-PnHQgn
    8aac8D-zIwhY9-88GxYx
    xaca88-4E9ceZ-Yx8c48
    ngnQQg-oxqaRX-nLH9L6
    vkvkVF-RxL7MG-kVVv5v
    8Y8D88-fJIb8F-xYx8c8
    8x8x8K-82daju-xcD8Yx
    c8Jc8Y-GU8JRy-8a88cY
    D848cK-0jaJxg-J88YYc
    Vv5FkF-bOeyK8-Fkkf5f
    gQNnHn-WGS9K0-gN6Pnn
    bvkkwF-NhDuQr-kFvkVF
    kk1Vfk-4aWYnb-Fvbk1f
    kVFbFF-erSOuJ-wkk1kv
    HPHQ6n-9lHaAh-LPng6n
    88YcaY-0qIYTk-xDY88J
    YY88Yx-AgFk80-8xDDDc
    Yc8c8K-nldoLK-888YKc
    UpypES-Ol9Ujw-yyyMyy
    YcacxD-KbQrek-8c88Yc
    D8J8Y8-SJKX5n-84YGc8
    Fvk5dk-K6auXt-5VVV5F
    MyySXy-LHYR2U-p3pRMp
    vvkkwk-7TNI6g-vFFkfF
    RSMy3p-JMw8qI-yyySpS
    RyySSy-60b8OH-pXSSMy
    8Y8888-AVb4Pb-DacxcD
    MyXpSE-PSA5X7-ypySSy
    gQQnni-IzZz4R-nPnnLn
    bkkkvV-j2TTzD-1Fkfvk
    gHnQLg-2LcmM7-nHnigP
    yUXMyU-RA8N0r-SMSpy3
    LnnLgQ-CWkC2Y-niHnQH
    888cc8-HWX2ML-8YcGcY
    k5kvFf-dxaYgv-VFvk5V
    xxc8Yc-q57E1T-JY88cD
    kVddkF-6pt4pn-VFfv1f
    gPgnnL-VbQ3lG-Q9nnQg
    FFvfkk-wDPW9d-FkfdFb
    S2yypX-QgNONa-MyRpyy
    Sp22Ep-mRnyTJ-yMy2Ep
    nnHgQQ-Pcwdh7-nHgigL
    FFkfdv-EJeIDW-5kFVkV
    2ySpUM-fGliRn-pyMSSy
    iLiggn-fTlEhz-nnLQHP
    3Spyyp-brGSW8-pUyMS2
    3Spyyp-brGSW8-pUyMS2
    nQnnnL-E6HSwW-gPnng6
    ax8D88-0tZR3G-DcGYcc
    FFbkfk-Ahqclr-kkFvFV
    pypU3R-RHKaOf-3yMySy
    S2MMyp-tfWxGF-Upyyyp
    yyySyM-nM9cMj-Mp2EyX
    V1FdkF-A4ny3T-kkbVvd
    nNQLgQ-0Hvaqt-ngnngi
    fVvFFk-NUHTax-Fkkkkv
    kFkFf1-eRmTqv-vkv1kF
    yXMSyM-Qh65tM-yyUXSp
    yppMSX-yGCakE-2pyM2y
    fwVFVk-183UMQ-vkFFkv
    pSy2yS-UKDq4o-yMXyyS
    88G4Yx-DwlHjt-c8cx8Y
    ypyyXy-pb3u0i-pySpyM
    Fkfk15-4XDphg-VbvVvF
    gNL6nN-f26zyi-ngQnQQ
    gQLggn-BTZkL7-QNHQin
    2yyyMX-5YjiUW-XXppSX
    Fk1kvw-B62poY-FkvvFF
    MMXyyy-NSCi2H-pXpSER
    ggnLHn-LXeSkS-gPQQnQ
    8ax8Yc-B1iwrH-K4D8cc
    kFbfF5-74ahY9-wkVVVv
    88cY88-eFXdZr-xxYxcc
    K8xcxY-o5yMOo-c84YYD
    yyXSpM-Kcb8JW-MyppMy
    kvdVkF-5kFlwN-FFwfkF
    888acc-8WUCbS-DxcYaD
    ngnQiH-QhNNz0-iPngng
    yM2pU2-QDc6cT-yyp2SU
    kkv5kV-7Pc53X-VfkFFk
    nLngQH-yer1jS-HggQLn
    UypMp2-3gaNFr-yMXp2S
    SpyyUM-4jH8rz-M2XMp2
    kVkbwF-ajPl3H-vvkVdb
    VVkVFv-44HtbN-Fkkbdv
    kbkk5k-izx11l-FFFFvV
    UMSSSp-AJ7hpH-MREyyM
    G8caD8-jp4j8f-K4xYcx
    SyySyy-NuaovJ-MXXSpp
    yyEXpE-urWr81-EyyS2p
    gnLQLg-kDTow4-LLnngn
    NNQnHn-IDa7pf-QggNLn
    fVkVFF-tjWVdW-vVFkkd
    pyMypR-2T25Tq-RSyMpS
    pUXMyy-ja9d9w-SyMEUp
    gnQgnH-Ri9g32-LgHgNL
    8DD8c8-OxZqL1-cYJacK
    kfkvVk-rPQHjL-b5kFb1
    nNQnng-82kbpU-Hng9nN
    cYa888-p2GCPQ-88Y8cD
    xYx88c-E7tRMc-8cDaa8
    pySyS2-qgNbnp-yMppME
    bk5kkk-o0af4U-kv5dFk
    Vk5Fkv-w4UgEO-FvvkdF
    kVkwFb-SxtreL-5Vkkvd
    Nnngnn-6o7P2u-ggPQ9Q
    cc8Kac-0EDf1N-Y8DxY8
    VFFvFk-2NUr78-kwv51V
    SyEppM-Q7oNnr-pypUyy
    ngnQ6H-GtyUpN-nnnHnn
    LnninL-Gfjigu-PiLggQ
    nnLPQg-EHCVvo-NQgnnN
    nQLnQn-6Uj2Lk-gPnNQH
    nggQig-eA0RQN-HQnnnH
    8Dc88Y-cw0ufi-cKc88D
    XyMyyp-UYPEdR-SX3SUp
    8Ycca8-XkkqQg-aGY88K
    yUypSS-4WrxnC-XMppSy
    4K8cx8-YqtVwW-cY88DY
    pM2pyR-kKwzxW-ppUSMy
    LnnnQi-hRaaWB-LnnP9g
    88cc84-oeLGiw-KDcY8a
    FFdbkk-1iiZ6G-F5kv5k
    kfFvFv-8pxrcr-Fkkkkk
    ngHNgQ-LFjcFH-ngLLnQ
    QgLHgn-PhNwl4-gLQnNQ
    p2MyMM-A6l922-S22Spy
    6ggnnL-3n7MKb-NnnnHQ
    8cK8Yx-43UM4i-cY88c8
    4YD8c8-udiY9H-K8c88c
    9HPnQn-T5VVou-NgHnnn
    c8axx8-qVDXly-cxYcYD
    kkV5dF-vHHHJQ-Vfkvkk
    5FkwdF-tf6KQ5-VkvFv5
    ySpyyy-tNjCkC-ppp3My
    n6gQLn-DJ4wkP-gngnng
    nLnLgN-L3ErQj-Q9Hgng
    66n6Ln-irXTVx-LngnQg
    a8YD88-7wjh7o-8ccca8
    nQPgHH-N1XnV8-LLQLnn
    HnnLLN-kEtUgV-nHLngQ
    fkvkk1-uIqKay-k1kFvV
    SXXyM2-Zp7pPp-pRpp3y
    nggngg-6QpEdq-nQiHLN
    VF1fkk-gCRSjm-vkFVvk
    8Y8aD8-PxTvBD-c88xJ8
    kFkV1v-88YvWJ-kkkkFf
    c88YJ8-LzcBD4-xcc8ca
    pUppyp-w2d575-ypSyyp
    8GYYcD-5rRgXA-8Gccx8
    ypSMpM-ywy4eI-yy2pyS
    8Yccc8-tYneHr-xDx888
    9Qgn6N-riIe1q-nnngig
    vkVFkV-g7pgZ9-kkFwFv
    QnLLnn-Jl04DF-nnQngL
    kwdFFv-WWH1i6-wF5kkk
    8DGcx8-lqWmWH-acDcY4
    SpyEXy-rDPjYw-UyRppy
    2EyMpS-6pzr09-ypMyyp
    dkkvkk-hViWbW-FwvkFw
    F5V5kV-F54puu-VvkdVv
    XyXyp3-xkCbUh-yyMSpM
    4c8DY8-Ot68y0-ccYxGx
    nNQnQn-aCXGqU-9QgQQn
    VkFkvk-Hl7mGl-fkfk5k
    2XySyp-yxhDQO-3SppUy
    MRypyS-2i4Url-pyypSS
    X2SyRU-LyWtXP-ySXpyy
    fVfkdv-RZJBcC-vFkVFb
    NnnnLQ-cFkMCg-nLgQin
    FFFkFv-RwtTEA-fkkFkf
    dkfkkk-R8qMjG-vFbfkk
    kvFFdF-UpHnMV-5k5kkk
    Fkkk5V-4oBPWI-VkkkvV
    ggHLnn-M0Pj8x-iHgn6Q
    yEpRp2-AvXSox-yySyyp
    gnnN9i-E039Jw-nLgggQ
    6ngg6L-43MyEK-nggQnL
    RXXyp2-0pDXRa-yypSXM
    yX2yyp-VUSzza-pUypSM
    kvfFvk-GXYMh2-FVvwFF
    Ff5vkk-1HlTBb-F5kkkF
    y2XXSS-a7bZvT-ypEypM
    kVvF5V-GmUVQq-kdvvFk
    gnggn6-a2Ihzl-nQgiNQ
    LnnLQ9-xPq7uk-gLnNnN
    MXpyMy-AnJ1Pl-MSSppy
    QnnHnN-7Yhh3J-iLHnng
    HgLgHg-4OjrcS-nQnnnL
    yypS3p-rK8Xap-yyMRMS
    nNPQng-Fbu6qi-LgnNPL
    ngHnnn-qDHVrz-nngNQL
    x8x888-BdMNIF-D8Yacc
    NQniPn-0CeUUw-HgQngn
    Nnngnn-y9TBzW-Qin9Qn
    VkkkkV-RXSWMY-VkfFbv
    SpypEp-1S1DBJ-y3M2yS
    nnHLnL-oJk1fq-QiLgLg
    giQn9n-X9l3YG-ngNLQH
    gLinHL-lgIWYS-LPnNQn
    kkvwF1-cuoDmI-vkVkvk
    xY8488-iitdyJ-Kc8xac
    pXXyyS-QO24lS-pySUXy
    fVFFkk-k8iczo-1d5fkv
    NginNL-KTIcT4-nnnHQi
    cG888Y-7l8Nje-8DDK8c
    inNgng-LZEk1P-PgnQLn
    pySSRS-AQlvCZ-Upyyyy
    dkkF55-UPt79F-f5kv5k
    kFvfk5-v3CpqY-1bVFVk
    6LQnng-M8eQmc-ngg9gn
    88YcGY-iSyEBR-xcDDYc
    pyEEyy-rjtInL-ppSpMS
    nnnLLQ-xFzJFn-Lnngnn
    XypS2E-Pf3SLV-EpXyyy
    MyUyyp-qVn3L4-SSySpE
    c88c8c-tVBpGY-x8YYKG
    Fvkkkd-3fIdSd-kkk5Vd
    XyMMyp-1uGxEG-pySXEy
    k5fkvF-JVNG1E-kkkVbV
    84Y8xJ-lx9wZe-Y8cxY8
    yySyyy-JQPu4V-yRyUSp
    Y8Jcc8-qpCJbt-8D88YK
    fvv1Vk-UFvCIT-kFwkkf
    HNnnnQ-C0A9Nm-gHNQLn
    SypySM-VQTvjR-My2yyp
    6PHngg-60tYV7-6QnnnL
    vkFkkF-iiI1X7-FFVkdv
    wkwdFv-cZuRig-VdFkkk
    8YG8cY-z6ci46-8cD8Kc
    P9QnHg-kN8ED4-inLgng
    8YY8Jc-o9uHiH-88xYDx
    kkdkF5-CIGccH-fF5kvV
    SUy3yM-OaMBcv-yypUpy
    ypMy2p-FFzmW2-ySpySy
    ypyXRS-L16SGq-SpyUpy
    Y8x8a8-8MNGCw-DccYx4
    Ffv1vk-NxUbDV-fkVFFw
    FkfFkk-6vHnSh-VbvdVf
    F5k5kV-CciOJd-kvkVVF
    c8c4Dc-Teghco-c88YaD
    NQgnnn-3ibViR-L6ngnL
    5Vkkkk-Kc0t7m-V5vFkv
    f5kkvk-35IYvc-kF5VdF
    pESSyp-gn0CgU-pyMRyS
    8c888J-5hivVv-4acxYK
    yypMyX-MNe9fy-MpSUS3
    FVVvvk-DJQmUN-kFkwVd
    yppyXp-3u6POe-SS2yyy
    8KJaDa-Kj9bZ1-cDcY88
    8Y4x84-bwtANd-axcx8J
    gQNQnn-xuRV4o-LnnngN
    VvwkFv-iy7fiU-wFkVbf
    ngiiQg-lOpYlt-ngLLnL
    ac8G88-JAl5JM-8DYx8c
    QPnnNL-9edpKq-HggnnL
    k1kkk1-n3MtXH-Fvv1kF
    fVvb1k-FvDpbQ-Fkk5fF
    c88cYG-pC3MhR-D8DYxa
    c88Gc8-hwd3Pl-8DY8a4
    VdkkFv-ZYQbn4-fkVwkk
    88YDx8-BGE0rQ-cc8484
    2yppRS-BDzUOM-MMSEpy
    nQLQg6-rXiLHS-LLHggn
    vvVVFF-VATfWX-1Ffkkv
    cxac8c-PMljcj-88YccG
    MyXSyX-1GwVBR-S32ypp
    nnQgHH-OvNqkj-gNQngL
    cDc4YY-0JmWlW-Y88Ycx
    SUpSpy-vai0ka-XSyUyp
    cY8cxJ-xUWnJw-88DY48
    ngQLg6-ZkMy0I-9nnnin
    NgLnNn-Lwmd5n-HngniQ
    xYx8ac-a5TU6t-Dc8c88
    gNnnPn-ECzi8c-LnQngN
    SyXMM2-fSJmmA-SpMpMy
    kF5555-iF0iVE-vVkfVF
    FdkVvk-ij3jSg-fk51fF
    LgQgQn-9Jqh11-iQggQn
    FvkkvV-Ng42hI-FFFfkV
    bkfFVb-u7BzOm-fkVVfv
    ngQQnn-g3wfZb-igNPLP
    QgnNnN-qpdvJb-nHgnnH
    XMXypy-8mJmMd-SyRSXp
    kFVk5v-FTN4dQ-kVkkkf
    6nQNLQ-qvq8TJ-LnnHgn
    XMSXy2-0O7Yzy-pyySpy
    nLgNgL-jK0iJz-QQnnng
    

    View Slide

47. Hierarchical clustering
    0.0
    0.1
    0.2
    0.3
    0.4
    vVkVkf-CfQ2Me-5wF5k5
    pM2ySp-NqdVVy-SEXyyy
    Ffvkv5-urPxNL-VkFdvk
    x888Y4-ZEG7rl-8a8c88
    pySESM-DfgV3e-EpSpEy
    QggnnQ-rTu8JY-iingnn
    cY88a8-41UAjT-88Gc4D
    888Y8K-M5mgp6-Y4cxcD
    pyySSU-fpvTh7-Upy2XM
    cY48J8-P6fNIl-Kxac8J
    HHnQnL-P97pwl-PHgnng
    MSpSpy-vAWjQl-MySMpS
    fkFkFF-zaC6UD-v1kkbw
    pSypUM-wkyESS-p2yyM3
    8YDcxc-feOc08-4GY884
    kFk1Fv-0TdM4n-dv5kkV
    Yx8ax4-JDBmvc-cYG8c8
    SSyXy2-N8v9C7-y2p3yp
    pXy3yy-9MrduS-yUpSy3
    Fdv1k5-WkoeFa-FkvVkk
    84888c-xCG2yN-Y8cGD8
    wkVvFV-nyF8eL-FVkkfk
    xa8a88-dom7r2-x8Y4cx
    gnngiP-TQ1apH-nPnnQn
    nNL9gn-eoFUGm-nNnngQ
    fVwwVw-ubXExr-FFkkv5
    QQn6ng-IKDkHa-nNiNnn
    Y8ccYa-GItA31-cY488x
    NHnnPg-jQhceA-nQnnLL
    yEUMRy-2mG9g3-pSSMyM
    Fkk1Vw-d4Wozk-FFvvFf
    6QQg9n-AbEJHX-giPngn
    gN6nQn-ckH59n-9inQnL
    D88xY8-hpaIDr-8cxcYc
    LQggQn-hv20uI-HnnnQg
    YcK48x-cDdQ0X-xxGD8D
    nHL6Hg-knDcUp-6nQngQ
    SSMUpp-INYSgw-ppyySy
    nng96n-lBix3f-LQQLgg
    SMRSyy-GHEWTP-MyXpyE
    Vbkv1k-tOY61g-F5kvVk
    PnnnLn-PEaw1c-QLnngg
    pypSpy-c9cBp2-SSyypp
    XyyyyX-TrT2ek-SpyMpy
    LHHgHQ-mLxVMM-nngnn9
    kVFkFb-AN5NWI-kvvfFb
    kv1v1f-vaYx04-kFfkkV
    x88J88-jBN0Jj-8Yc8Kx
    ySM3yy-gnhe17-XyUSyp
    8xYD8D-npQXH8-KYYcxY
    Y4Y84x-MrJeFq-cDD488
    nnnNgQ-VNO3JO-igNnNg
    5FFk1k-hWutyR-vwVk5F
    Q6nLNN-H90bPb-nLginn
    pyySS2-H7DN4A-pUSyRy
    SEpEyy-hbxATA-SyyppM
    YKc8xY-R3feUE-8cc8a8
    g69ngL-gWWKti-QinnnH
    nLgNQn-JVUT8W-HPHQLn
    D8Yx88-lrJIuu-JD84c4
    kv55F5-3erdMp-VkFFwk
    v5VFf1-UKYbZn-VbVkkk
    Y48Y8a-RJ2LDC-8c8xcD
    pS2y2p-ZSaQvC-MSUyyy
    Sypyy2-K6PSF4-MppXXy
    K8c8YK-4jx3NW-8Dac48
    XyRyyU-qrp4V9-pMRySM
    J888cc-lxoCeB-4DYJ8c
    giPgni-OJ09Yi-gngnnQ
    Y84cxx-3rJwVM-cYY84c
    ngiQgN-CnQ4gv-6nQngn
    kkkFVb-zoKBee-vFkFvk
    6QgHnn-2VFveI-QnnngH
    3yRSyy-D8i0T9-yy2ppM
    gnLnQQ-Cyt2rM-L6nnLN
    MUpSyX-xcRVC3-pyyRpS
    dkvkfk-77tnCm-FFVfFV
    F1kkFf-MVW9eT-vkvFkd
    kkkFkk-xYY2UD-kvw15V
    nngHng-OBmOQL-NQHLQL
    8xKD84-DzKFZJ-4xK8cJ
    Q9ngLn-bjLzmZ-gQnQgQ
    NPnLgn-58T69t-nQnniH
    8GD84D-DuK5jM-xYc84x
    Vk5vkk-bcQSA0-FkvFkv
    LLQnQn-cTWB0Y-gNQNnn
    nLnQnn-mtbSBC-ngQLHL
    ypUyy3-3rxk2j-p3pSy2
    SpyyRM-Cq2rZQ-UyyMEy
    55kvkF-p1hUSP-kFkFkb
    FkFVdF-Clq0WL-kvv1kd
    gnLngg-3NcLDa-Q9nPPN
    nnngQQ-QwK7at-9P9gnn
    cxY8G8-59CrtL-8c488x
    LLLgnn-VdBYdx-gLnNnQ
    gHniQg-bmCra9-Lnggni
    Eyp23X-6lKVUo-MpppyS
    x88DYD-yXtIEy-8cJ88x
    8Y4ccx-kwQSDk-48x8DG
    LPNnLn-G0fVYk-iQNnng
    NQLnnN-vDiFBq-nP9gNn
    Fvbkff-MmzaQA-kvVwVk
    c88Yc4-YeMfSt-8c48ca
    kkvdVV-OgMIQX-kFkFkk
    48YKYc-HRXkZA-Ya8c88
    nnQng6-DRu6oG-gnigin
    8YaxDD-5uPjMv-c84xKc
    gLnQQn-oProA3-gLQgn9
    gngQgg-bLqQCo-g6nQn6
    8xca8Y-YGIUkZ-8x8Ycx
    nnn6NQ-oUJKe0-HQLngg
    88xxGa-Og85PD-c8Ya8D
    EMSyp2-2fQqfN-yyESUp
    YY84c8-RR4n4d-8c8xJx
    k5FV1v-975h5I-FVFkkv
    LPNgQn-3q0i7b-6ngHNn
    nggnn9-4PJZ5K-gLQHng
    QQgPnn-vVzh0f-gnNnLg
    ySyyR2-K40qf2-pXyMpp
    kkVFkF-2fCbID-vkdFFV
    x8Y8K8-t24oZW-acYx8c
    p3SXEp-zAnOdf-yMSyXp
    6nQgLg-796wEi-9innHL
    yyppMX-i88khM-pS3SSU
    kkFk5k-pQFm92-kvvkFv
    F1vFvd-BqFbRu-FkfkFk
    XMpySy-Pdzy6z-yMEM2p
    8x8x8K-82daju-xcD8Yx
    c8Jc8Y-GU8JRy-8a88cY
    D8J8Y8-SJKX5n-84YGc8
    Fvk5dk-K6auXt-5VVV5F
    gPgnnL-VbQ3lG-Q9nnQg
    FFvfkk-wDPW9d-FkfdFb
    ax8D88-0tZR3G-DcGYcc
    FFbkfk-Ahqclr-kkFvFV
    pypU3R-RHKaOf-3yMySy
    88cY88-eFXdZr-xxYxcc
    K8xcxY-o5yMOo-c84YYD
    yM2pU2-QDc6cT-yyp2SU
    gnLQLg-kDTow4-LLnngn
    8DD8c8-OxZqL1-cYJacK
    nnLPQg-EHCVvo-NQgnnN
    4K8cx8-YqtVwW-cY88DY
    FFdbkk-1iiZ6G-F5kv5k
    kfFvFv-8pxrcr-Fkkkkk
    6ggnnL-3n7MKb-NnnnHQ
    4YD8c8-udiY9H-K8c88c
    9HPnQn-T5VVou-NgHnnn
    c8axx8-qVDXly-cxYcYD
    nLnLgN-L3ErQj-Q9Hgng
    66n6Ln-irXTVx-LngnQg
    a8YD88-7wjh7o-8ccca8
    nggngg-6QpEdq-nQiHLN
    8Y8aD8-PxTvBD-c88xJ8
    ypSMpM-ywy4eI-yy2pyS
    9Qgn6N-riIe1q-nnngig
    8DGcx8-lqWmWH-acDcY4
    F5V5kV-F54puu-VvkdVv
    VkFkvk-Hl7mGl-fkfk5k
    fVfkdv-RZJBcC-vFkVFb
    kvFFdF-UpHnMV-5k5kkk
    gnnN9i-E039Jw-nLgggQ
    Ff5vkk-1HlTBb-F5kkkF
    HgLgHg-4OjrcS-nQnnnL
    SpypEp-1S1DBJ-y3M2yS
    nnHLnL-oJk1fq-QiLgLg
    

    View Slide

48. Hierarchical clustering
    0.6
    0.7
    0.8
    ↑ Height
    

    View Slide

49. Weekly deck trend analysis
    適当な高さで切るとデッキアーキテクチャごとのクラスタになる！
    49
    

    View Slide

50. Pokémon TCG similar deck search 2023
    50
    card normalize map
    
    
    data/uniq_*.txt
    new deck
    metadata
    known deck
    
    
    sqlite3
    sqlite3
    My MacBook
    App
    deck
    
    similarity
    Search
    Initialize
    Crowler
    Web UI
    @deck
    
    
    @idf
    
    
    @norm
    @name
    
    
    @id_norm
    Make Vector
    Update Deck
    Build page
    periodic task
    検索エンジンを流用して
    

    View Slide

51. Weekly deck trend analysis
    \
    51
    card normalize map
    
    
    data/uniq_*.txt
    new decks
    known deck
    
    
    sqlite3
    sqlite3 Rakefile
    deck
    
    similarity
    Initialize
    Crowler
    Clustering
    @deck
    
    
    @idf
    
    
    @norm
    @name
    
    
    @id_norm
    Make Vector
    hierarchical
    
    
    clustering
    クラスタリング
    n weeks
    

    View Slide

52. Weekly deck trend analysis
    \
    52
    card normalize map
    
    
    data/uniq_*.txt
    new decks
    known deck
    
    
    sqlite3
    sqlite3 Rakefile
    deck
    
    similarity
    Initialize
    Crowler
    Clustering
    @deck
    
    
    @idf
    
    
    @norm
    @name
    
    
    @id_norm
    Make Vector
    hierarchical
    
    
    clustering
    クラスタリング
    n weeks
    

    View Slide

53. Ractor-ize
    Using Ractor is easy, but writing for Ractor is hard.
    
    
    Refactoring was very hard
    
    
    Move calculation part to mix-in
    めちゃめちゃリファクタリングした...
    53
    

    View Slide

54. n週間分のRactorをつくるよ
    54
    def weekly_analyze(world, deck_and_date)
    
    
    origin = Date.parse('2022-10-07') # ༵ۚ೔࢝·Γ
    
    
    weekly = deck_and_date.sort_by {|x| x[1]}.chunk {|x|
    
    
    (Date.parse(x[1]) - origin).to_i / 7
    
    
    }
    
    
    result = weekly.map {|_, decks|
    
    
    {
    
    
    'range' => Range.new(decks[0][1], decks[-1][1]),
    
    
    'deck_count' => decks.size,
    
    
    'cluster' => Cluster.new(world, decks.map {|x, d| x})
    
    
    }
    
    
    }
    
    
    result.each do |x|
    
    
    x['cluster'].join
    
    
    end
    
    
    result
    
    
    end
    Ractor.new
    take
    

    View Slide

55. 55
    2022-10-07 .. 2022-10-13
    2022-10-14 .. 2022-10-20
    2022-10-21 .. 2022-10-27
    2022-11-07 .. 2022-11-03
    2022-11-04 ... 2022-11-11
    

    View Slide

56. map { Ractor.new }.map {|r| r.take }
    Ractor起動！
    56
    >> result = weekly.map {|w| Ractor.new(w) {|it| do_cluster(it)}}
    [ ]
    

    View Slide

57. map { Ractor.new }.map {|r| r.take }
    終わったRactorもいるし、動いてるRactorもいるぞ
    57
    >> result = weekly.map {|w| Ractor.new(w) {|it| do_cluster(it)}}
    [ ]
    

    View Slide

58. map { Ractor.new }.map {|r| r.take }
    気にせずtakeで待ち合わせ！
    58
    >> result = weekly.map {|w| Ractor.new(w) {|it| do_cluster(it)}}
    [ ]
    >> result.map {|r| r.take}
    

    View Slide

59. map { Ractor.new }.map {|r| r.take }
    Linda本でもこういうイディオムあったような気がする
    59
    >> result = weekly.map {|w| Ractor.new(w) {|it| do_cluster(it)}}
    [ ]
    >> result.map {|r| r.take}
    [ ]
    

    View Slide

60. 28 Ractors, 8 cores (6 performance and 2 efficiency)
    40%くらい速くなった。デッキ数が週ごとにばらつくため、多い週に律速した？
    60
    w/o Ractor
    
    
    % /usr/bin/time -l rake
    
    
    51.03 real 50.28 user 0.55 sys
    
    
    Ractor
    
    
    % /usr/bin/time -l rake
    
    
    31.84 real 83.81 user 89.54 sys
    

    View Slide

61. Pokémon TCG similar deck search
    61
    card normalize map
    
    
    data/uniq_*.txt
    new deck
    metadata
    known deck
    
    
    sqlite3
    sqlite3
    My MacBook
    App
    deck
    
    similarity
    Search
    Initialize
    Crowler
    Web UI
    @deck
    
    
    @idf
    
    
    @norm
    @name
    
    
    @id_norm
    Make Vector
    Update Deck
    Build page
    periodic task
    検索の高速化
    case 2
    n decks
    

    View Slide

62. Pokémon TCG similar deck search
    62
    card normalize map
    
    
    data/uniq_*.txt
    new deck
    metadata
    known deck
    
    
    sqlite3
    sqlite3
    My MacBook
    App
    deck
    
    similarity
    Search
    Initialize
    Crowler
    Web UI
    @deck
    
    
    @idf
    
    
    @norm
    @name
    
    
    @id_norm
    Make Vector
    Update Deck
    Build page
    periodic task
    検索の高速化
    n decks
    

    View Slide

63. Ractor-ize
    Using Ractor is easy, but writing for Ractor is hard.
    
    
    Refactoring was very hard
    
    
    Add decks while searching
    
    
    I want an appendable shared object
    追記可能な共有可能なオブジェクトがほしいです！
    63
    

    View Slide

64. 変化がない領域だけRactorで検索する
    64
    if USING_RACTOR
    
    
    top = (_search_by_deck_core(@added_deck, v, n) + search_using_ractor(v, n)).max(n)
    
    
    else
    
    
    top = _search_by_deck_core(@deck, v, n)
    
    
    end
    
    
    search added decks
    
    
    (w/o Ractor)
    search static decks
    
    
    (Ractor)
    

    View Slide

65. 大きな配列を作るのを嫌ってイテレータを作った
    65
    def each_in(range)
    
    
    return enum_for(__method__, range) unless block_given?
    
    
    range.each {|x| yield(@deck[x])}
    
    
    end
    
    
    def _search_by_deck(v, n)
    
    
    (([email protected]).step(@deck.size / @nproc) + [@deck.size]).each_cons(2).map {|s, e|
    
    
    Ractor.new(self, each_in(s...e), v, n) {|world, sub_decks, v, n|
    
    
    world._search_by_deck_core(sub_decks, v, n)
    
    
    }
    
    
    }.map {|r| r.take}.sum([]).max(n)
    
    
    end
    
    
    enum of subrange
    subrange iterator
    
    
    

    View Slide

66. 10Ractors, 46400 decks
    66
    0...5811
    5911...11622
    11622...17433
    17433...23244
    23244...29055
    

    View Slide

67. map { Ractor.new }.map {|r| r.take }.sum.max
    max(n)を集めてmax(n)
    67
    .map {|r| r.take}
    e.map {Ractor.new(r) {|it| it.max(n)}}
    [ ]
    [ ]
    .sum([]).max(n)
    

    View Slide

68. 10Ractors, 46400 decks
    だいぶ速くなった！
    68
    M1 Pro MacBook
    
    
    w/o Ractor
    
    
    0.272264 sec
    
    
    Ractor
    
    
    0.070405 sec
    OCI ARM4 core
    
    
    w/o Ractor
    
    
    0.369918735 sec
    
    
    Ractor
    
    
    0.099914693 sec
    

    View Slide

69. Peak performance issue
    When tweeting, search slows down
    
    
    Multiple requests generate a lot of Ractor
    宣伝ツイートすると検索がめっちゃ遅くなる
    69
    

    View Slide

70. Agenda
    Concurrency is everywhere
    
    
    Introduction to Ractor
    
    
    case study
    70
    

    View Slide

71. Pokémon
    Trainer Energy
    Pokémon TCG
    Build a deck with 60 cards
    

    View Slide