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
Kubernetes (EKS) で構築する スケーラブルなジョブ実行基盤
Search
Kenta Suzuki
August 28, 2019
Technology
7
6.4k
Kubernetes (EKS) で構築する スケーラブルなジョブ実行基盤
Kenta Suzuki
August 28, 2019
Tweet
Share
More Decks by Kenta Suzuki
See All by Kenta Suzuki
trocco®での インシデント重大度表(SEVレベル表)運用
kekekenta
0
3.4k
Ruby はじめての型 Gotanda.rb 2023/04/20
kekekenta
0
290
trocco® の品質を守る、とても普通な取り組み
kekekenta
0
5.4k
Embulkを利用したデータ統合SaaSの構築と運用
kekekenta
2
5k
Other Decks in Technology
See All in Technology
[OCI Skill Mapping] AWSユーザーのためのOCI – IaaS編(Compute/Storage/Networking) (2025年10月8日開催)
oracle4engineer
PRO
1
130
Kubernetes self-healing of your workload
hwchiu
0
310
20251007: What happens when multi-agent systems become larger? (CyberAgent, Inc)
ornew
1
510
SCONE - 動画配信の帯域を最適化する新プロトコル
kazuho
1
280
だいたい分かった気になる 『SREの知識地図』 / introduction-to-sre-knowledge-map-book
katsuhisa91
PRO
0
530
組織改革から開発効率向上まで! - 成功事例から見えたAI活用のポイント - / 20251016 Tetsuharu Kokaki
shift_evolve
PRO
2
220
現場データから見える、開発生産性の変化コード生成AI導入・運用のリアル〜 / Changes in Development Productivity and Operational Challenges Following the Introduction of Code Generation AI
nttcom
1
410
サイバーエージェント流クラウドコスト削減施策「みんなで金塊堀太郎」
kurochan
4
2.2k
OpenTelemetry が拡げる Gemini CLI の可観測性
phaya72
2
970
Introduction to Bill One Development Engineer
sansan33
PRO
0
300
会社を支える Pythonという言語戦略 ~なぜPythonを主要言語にしているのか?~
curekoshimizu
1
230
Introduction to Sansan, inc / Sansan Global Development Center, Inc.
sansan33
PRO
0
2.8k
Featured
See All Featured
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
34
2.5k
Building Better People: How to give real-time feedback that sticks.
wjessup
369
20k
Done Done
chrislema
185
16k
The Web Performance Landscape in 2024 [PerfNow 2024]
tammyeverts
10
880
Embracing the Ebb and Flow
colly
88
4.9k
Context Engineering - Making Every Token Count
addyosmani
7
280
Into the Great Unknown - MozCon
thekraken
40
2.1k
[RailsConf 2023 Opening Keynote] The Magic of Rails
eileencodes
31
9.7k
Fantastic passwords and where to find them - at NoRuKo
philnash
52
3.5k
The Invisible Side of Design
smashingmag
302
51k
Music & Morning Musume
bryan
46
6.9k
XXLCSS - How to scale CSS and keep your sanity
sugarenia
249
1.3M
Transcript
Kubernetes (EKS) Ͱߏங͢Δ εέʔϥϒϧͳδϣϒ࣮ߦج൫ Kenta Suzuki (twitter: @kekekenta) CTO, primeNumber
inc 2019/8/28 Kubernetes Meetup Tokyo #22
ࣗݾհ • Kenta Suzuki • twitter: @kekekenta • CTO, primeNumber
inc. εϥΠυͷ࠷ޙʹձࣾհࡌͤ·ͨ͠ͷͰɺ͝ڵຯ͋Γ·ͨ͠Βޚཡ͍ͩ͘͞ •ੳج൫ߏஙɾӡ༻ࢧԉSaaSʮtrocco®ʯ
.Z42- 4 4BMFTGPSDF #JH2VFSZ ੳج൫ߏஙɾӡ༻ σʔλసૹ αϚϦʔςʔϒϧ࡞ σʔλࢹ ֎෦࿈ܞ
.Z42- 4 4BMFTGPSDF #JH2VFSZ σʔλసૹ αϚϦʔςʔϒϧ࡞ σʔλࢹ ֎෦࿈ܞ ੳج൫ͷߏஙɾӡ༻Λ·ͱΊͯΧόʔ
troccoͷσʔλసૹδϣϒ • σʔλιʔε • MySQLɺPostgreSQLɺSQLServerͳͲ • εέδϡʔϧ࣮ߦ • ࠷1ִؒ ҙͷσʔλιʔεͷςʔϒϧͷ༰ΛDWHʹసૹ
troccoͷσʔλసૹ෦OSSͷembulkΛଟ༻͍ͯ͠·͢ɻ ͜ͷΛआΓͯɺίϛολʔͷํʑɺembulkίϛϡχςΟͷํʑʹײँਃ্͛͠·͢ɻ ࠓޙembulkͷpluginެ։Ͱߩݙ͍ͯ͘͠༧ఆͰ͢ʂ
troccoͷసૹδϣϒ • δϣϒ࣮ߦ • ʙ10,000/ • δϣϒ͝ͱͷ࣮ߦ࣌ؒ • min: 5s
ʙ max: 5h • ࣮ߦ࣌ؒଳ • ࣌ؒʹ1ͷεύΠΫ
Fan out on ECS چΞʔΩςΫνϟ
Fan out on ECSαʔϏε SQS εέδϡʔϥʔ 8PSLFS &$4UBTL 8PSLFS &$4UBTL
8PSLFS &$4UBTL ECS ECS NODE δϣϒ polling
λεΫεέʔϧΞτͷλΠϛϯά͍͠ • SQSͷΩϡʔͷαΠζΛࢹ • CloudwatchϝτϦΫε͕࠷1͝ͱ • εέʔϧ·Ͱɺ1ͨͳ͍ͱߦ͚ͳ͍
ϗετεέʔϧΞτͷλΠϛϯά͍͠ • λεΫ͕͏ՃͰ্ཱ͕ͪΔϦιʔε͕ ͳ͘ͳͬͨΒεέʔϧ • ىಈͪͷλεΫ͕ग़དྷͨλΠϛϯάͰϗ ετεέʔϧͯ͠΄͍͠ɻɻ
λεΫεέʔϧΠϯ͍͠ • ECSλεΫΛ҆શʹऴྃͤ͞Δඞཁ͕͋Δ • సૹδϣϒ5hʹٴͿ߹͋Δ • TERMγάφϧͷτϥοϓͯ͠graceful shutdownΛ ࣮ •
ͨͩ͠ɺECSͷίϯιʔϧ্ͷ࣮ߦதλεΫҰཡ͔ΒɺTERMΛૹͬͯ͠Β͘͢Δ ͱফ͑ͯ͠·͏
ϗετͷεέʔϧΠϯ͍͠ • Auto Scaling Group ͷ Life cycle hook ΛͬͯɺΠϯ
ελϯεఀࢭલʹDRAININGΛ࣮ࢪ • ىಈதͷλεΫ͕ͳΔ·Ͱࢹ͠ɺͳ͘ͳͬͨΒΠϯε λϯεఀࢭΛ͢͢ΊΔ • લड़ͷ௨ΓɺىಈதͷλεΫҰཡ͔Βফ͑ΔͨΊ͝ʹΐ ͝ʹΐͬͯͨ
Ϧιʔεׂ • సૹߴԽΦϓγϣϯͱͯ͠ɺCPU/ϝϞϦ্ݶΛઃఆ Ͱ͖ΔΑ͏ʹ͍ͨ͠ • ECSͰλεΫ͝ͱʹϦιʔε্ݶΛઃఆͰ͖Δ͕ɺى ಈ࣌ͷΈ • Ϧιʔε্ݶ͝ͱʹαʔϏεΛ͚Δඞཁ͕͋Γ
͍͠
ͦ͜Ͱɺk8s
k8sͷpodεέδϡʔϥʔʹͭ ͍ͯ
k8sͷpodεέδϡʔϥʔʹͭ ͍ͯ
k8sͷpodεέδϡʔϥʔʹͭ ͍ͯ statusͷભҠ "Pending" -> "Running" -> "Succeeded"
k8sͷpodεέδϡʔϥʔ • nodeϦιʔεͷۭ͖ঢ়گʹԠͯ͡ɺεέδϡʔϥʔ͕ॱ࣍podΛ nodeʹׂΓͯͯ͘ΕΔ • podΛ࡞͓͚ͯ͠ɺ͋ͱ࣮ߦ͞ΕΔͷΛ͚ͭͩ ,VCFSOFUFT QPE /0%& QPE
QPE pod্ཱͪ͛ ࣮ߦͪpod create pod
ͱΓ͋͑ͣࡶʹpodΛcreate͢Ε ɺ͋ͱk8s͕࣮ߦ·Ͱ໘ݟ ͯ͘ΕΔ
ϊʔυͷεέʔϧΞτ/Πϯ • EKSͷ߹ɺcluster autoscalerΛ͏ͷ͕σϑΝΫτʢͱ ࢥΘΕΔʣ https://github.com/kubernetes/autoscaler/tree/master/ cluster-autoscaler/cloudprovider/aws • podεέδϡʔϦϯάͷঢ়گʹԠͯ͡ɺEC2ϊʔυͷASG ܦ༝ͰεέʔϧΠϯɾΞτΛͬͯ͘ΕΔ
ϊʔυͷεέʔϧΞτ QPE &$/0%& QPE QPE 1FOEJOH ࠓͷnodeϦιʔεͰɺ podΛ্ཱͪ͛Δ͜ͱ͕ग़དྷͳ͍
ϊʔυͷεέʔϧΞτ QPE &$/0%& QPE &$/0%& QPE ClusterAutoscaler͕ASGܦ༝ͰnodeΛεέʔϧΞτ
ϊʔυεέʔϧΠϯ QPE &$/0%& QPE &$/0%& QPE
ϊʔυεέʔϧΠϯ QPE &$/0%& QPE &$/0%& Ϧιʔεʹ༨༟͕͋Δ
ϊʔυεέʔϧΠϯ QPE &$4/0%& QPE εέʔϧΠϯ
1. create pod ࣮ͯ͠ߦΛͭ 2. ඞཁʹԠͯ͡ɺnode͕εέʔϧΞ τ͢ΔͷΛͭ εέʔϥϒϧδϣϒج൫ര
k8sʹΑΔδϣϒج൫
ݱࡏͷΞʔΩςΫνϟ SQS εέδϡʔϥʔ 8PSLFS 1PE +PC %JTQBUDIFS Kubernetes δϣϒ create
pod schuedule pod &,4/0%& 8PSLFS 1PE $MVTUFS "VUPTDBMFS Scale in/out
Job Dispatcher SQS εέδϡʔϥʔ 8PSLFS 1PE Kubernetes δϣϒ create pod
schuedule pod &,4/0%& 8PSLFS 1PE $MVTUFS "VUPTDBMFS Scale in/out +PC %JTQBUDIFS
Job Dispatcher ˞ίʔυΠϝʔδͰ͢ • SQS Λ polling • k8s API
ʹϦΫΤετΛ͛ɺ pod Λ࡞
Job Dispatcher ˞ίʔυΠϝʔδͰ͢ LT"1*ʹϦΫΤετ • SQS Λ polling • k8s
API ʹϦΫΤετΛ͛ɺ pod Λ࡞
Job Dispatcher ˞ίʔυΠϝʔδͰ͢ EPDLFSΠϝʔ δͷλάࢦఆ • SQS Λ polling •
k8s API ʹϦΫΤετΛ͛ɺ pod Λ࡞
Job Dispatcher ˞ίʔυΠϝʔδͰ͢ QPEͰ࣮ߦ͢Δ ίϚϯυΛࢦఆ • SQS Λ polling •
k8s API ʹϦΫΤετΛ͛ɺ pod Λ࡞
Job Dispatcher • SQS Λ polling • k8s API ʹϦΫΤετΛ͛ɺ
pod Λ࡞ ˞ίʔυΠϝʔδͰ͢ $16ϝϞϦࢦఆ
Worker pod SQS εέδϡʔϥʔ +PC %JTQBUDIFS Kubernetes δϣϒ create pod
schuedule pod &,4/0%& $MVTUFS "VUPTDBMFS Scale in/out 8PSLFS 1PE 8PSLFS 1PE
Worker pod • ࣮ࡍͷసૹΛ࣮ߦ • సૹ։࢝Ͱ্ཱ͕ͪΓɺసૹ͕ऴྃͨ͠Βpodऴྃ ˞ίʔυΠϝʔδͰ͢
Cluster autoscaler SQS εέδϡʔϥʔ +PC %JTQBUDIFS Kubernetes δϣϒ create pod
schuedule pod &,4/0%& Scale in/out 8PSLFS 1PE 8PSLFS 1PE $MVTUFS "VUPTDBMFS
Cluster autoscaler ˞ίʔυΠϝʔδͰ͢
k8sδϣϒج൫ಋೖͷՌ • ѹతίετݮʹޭ • εέʔϧͷλΠϛϯά͕͍͕͍͍͠ ײ͡ʹղܾͰ͖ͨ • cluster autoscaler༷༷ •
։ൃڥͱຊ൪ڥͷಉҰԽ
ࢹʹ͍ͭͯ • create podͰ͖ͳ͍ • pod͕pendingͷ··্ཱ͕ͪΒͳ͍ • pod͕runningதʹಥવࢮ ຊ൪ӡ༻࣮͕ແ͍ͷͰɺຊʹظ௨Γʹ ಈ͔͘ͱ͍͏ҙຯͰࢹΛ࣮ࢪ
troccoͷδϣϒεςʔλεભҠ • DB্ʹδϣϒ͝ͱͷεςʔλεΛอ࣋ • dispatcherpodͳͲ͕ͦΕͧΕߋ৽͢Δ RVFVFE TFUUJOH@VQ FYFDVUJOH TVDDFFEFE SQSʹΩϡʔΠϯά
create pod ྃ pod্ཱ͕ͪΓ·Γ pod͕࣮ߦத సૹྃ
redashͰͷࢹ RVFVFE TFUUJOH@VQ FYFDVUJOH TVDDFFEFE ֤εςʔλεͰཹ͍ͯͨ͠ΒΞϥʔτ
·ͱΊ • γϯϓϧ • k8sͷݞʹΕɺεέʔϥϒϧͳδϣϒج൫ߏஙγϯϓ ϧʹͳΔ • ࢹ • ӡ༻࣮͕ແ͍ͷͰɺͱΓ͋͑ͣऔΕΔσʔλͱ͓ͬͯ͘
• k8sͷҠߦΛৼΓฦͬͯ • όον͔Βk8sಋೖ͢ΔͷׂͱΓ͍͢ͷͰ͓͢͢Ί • কདྷతʹ • কདྷతʹɺϗετͷཧͱ͔Βͳͯ͘ྑ͘ͳΔͱخ͍͠
͝ਗ਼ௌ ͋Γ͕ͱ͏͍͟͝·ͨ͠ :bow Α͚Εձࣾհޚཡ͍ͩ͘͞
גࣜձࣾ primeNumber • 201511݄ ઃཱ • ࣾһ9໊ʢ͏ͪΤϯδχΞ8໊ʣ ύʔτφʔΤϯδχΞ30໊Ҏ্ • 201811݄
trocco®ϦϦʔε • 20197݄ γϦʔζAࢿۚௐୡʢ1.3ԯԁʣ࣮ࢪ https://primenumber.co.jp
ಇ͘ڥ • @தࠇͷԊ͍ • ίΞλΠϜ੍ • 11ʙ15࣌ • िதɺ2·ͰϦϞʔτՄ ※ࢼ༻த
• δϜશֹิॿ
We are hiring • ιϑτΣΞΤϯδχΞ • ΠϯϑϥΤϯδχΞ • ͳͲͳͲɺଟํ໘ઈࢍืूத