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
由Spanner來看Google資料庫的前世今生
Search
Szu-Kai Hsu (brucehsu)
November 07, 2012
Technology
4
280
由Spanner來看Google資料庫的前世今生
2012年秋,網際網路資料庫 @ 國立中正大學資工所
Szu-Kai Hsu (brucehsu)
November 07, 2012
Tweet
Share
More Decks by Szu-Kai Hsu (brucehsu)
See All by Szu-Kai Hsu (brucehsu)
Running Life Lean
brucehsu
0
170
Core Unleashed Part II: Introduction to GobiesVM (and STM) @ RubyKaigi 2014
brucehsu
0
2.1k
[RubyConf.tw 2014] Cores unleashed - Exploiting Parallelism in Ruby with STM
brucehsu
0
2.2k
用 Go 打造程式語言執行環境:實例剖析 [OSDC.tw 2014]
brucehsu
3
2.3k
pickbox @ OSDC.tw 2013 Lightning Talk
brucehsu
0
57
Building Web 2.0 APIs
brucehsu
1
150
Rapid Web Development by Example
brucehsu
3
3.1k
TechWed@CCU #0
brucehsu
2
510
Chromium OS
brucehsu
2
200
Other Decks in Technology
See All in Technology
FOSS4G 2025 KANSAI QGISで点群データをいろいろしてみた
kou_kita
0
410
関数型プログラミングで 「脳がバグる」を乗り越える
manabeai
2
210
CDKTFについてざっくり理解する!!~CloudFormationからCDKTFへ変換するツールも作ってみた~
masakiokuda
1
180
Lakebaseを使ったAIエージェントを実装してみる
kameitomohiro
0
160
cdk initで生成されるあのファイル達は何なのか/cdk-init-generated-files
tomoki10
1
240
american airlines®️ USA Contact Numbers: Complete 2025 Support Guide
supportflight
1
110
american aa airlines®️ USA Contact Numbers: Complete 2025 Support Guide
aaguide
0
400
【LT会登壇資料】TROCCO新コネクタ「スマレジ」を活用した直営店データの分析
kazari0425
1
110
How Do I Contact HP Printer Support? [Full 2025 Guide for U.S. Businesses]
harrry1211
0
130
NewSQLや分散データベースを支えるRaftの仕組み - 仕組みを理解して知る得意不得意
hacomono
PRO
3
190
アクセスピークを制するオートスケール再設計: 障害を乗り越えKEDAで実現したリソース管理の最適化
myamashii
1
180
SRE不在の開発チームが障害対応と 向き合った100日間 / 100 days dealing with issues without SREs
shin1988
1
480
Featured
See All Featured
Practical Orchestrator
shlominoach
189
11k
We Have a Design System, Now What?
morganepeng
53
7.7k
Rebuilding a faster, lazier Slack
samanthasiow
83
9.1k
Why Our Code Smells
bkeepers
PRO
336
57k
Into the Great Unknown - MozCon
thekraken
40
1.9k
Visualizing Your Data: Incorporating Mongo into Loggly Infrastructure
mongodb
46
9.6k
Agile that works and the tools we love
rasmusluckow
329
21k
Scaling GitHub
holman
460
140k
A Modern Web Designer's Workflow
chriscoyier
695
190k
Keith and Marios Guide to Fast Websites
keithpitt
411
22k
Building a Scalable Design System with Sketch
lauravandoore
462
33k
Raft: Consensus for Rubyists
vanstee
140
7k
Transcript
由 Spanner來看 Google資料庫 的 前世今⽣生 Szu-Kai Hsu (brucehsu)
Spanner is a scalable multi-version globally-distributed synchronously-replicated database
BigTable
Handling
Handling really
Handling really BIG DATA
key-value
key-value { “CCU”: “123”, “NCTU”: “113”, “NTU”: “112” }; key
key-value { “CCU”: “123”, “NCTU”: “113”, “NTU”: “112” }; value
distributed
Lack of transaction, think of our first project.
CAP
C A P
Consistency A P
Consistency Availability P
Consistency Availability Partition tolerance
Consistency Availability Partition tolerance Consistency
Megastore
NoSQL datastores are highly scalable, but their limited API and
loose consistency models complicate application development. “ “
In Megastore, data model is declared in a strong-typed schema
strong-typed schema CREATE TABLE User { required int64 user_id; required string name; } PRIMARY KEY(user_id), ENTITY GROUP ROOT;
Based on BigTable BigTable
PRIMARY user_id PRIMARY user_id, nyan_id
Local and Global Indexes are introduced: Local Index Find corresponding
data in entity group Global Index Find corresponding data in external groups Local Index Global Index
(user_id, born,nyan_id) For local index CREATE LOCAL INDEX NyanByBorn ON
Nyan(user_id, born); CREATE LOCAL INDEX NyanByBorn ON Nyan(user_id, born);
Consistency achieved via Paxos algorithm Paxos 2 Replicas 1 Witness
At least
Replica consists of Replication server and Coordinator Replication server Coordinator
write oversee
Witness’ Replication server only writes logs logs
Average Latency: 100-400ms Poor write throughput 100-400ms
Spanner ,finally.
We believe it is better to have application programmers deal
with performance problems due to overuse of transactions as bottlenecks arise, rather than always coding around the lack of transactions. “ “
Data model is almost identical to Megastore almost identical Basic
unit defined as Directory Directory
Data model is almost identical to Megastore almost identical Basic
unit defined as Directory Directory Same prefix key, therefore adjacent
Data model is almost identical to Megastore almost identical Basic
unit defined as Directory Directory Same prefix key, therefore adjacent Fine-grained mapping
Data model is almost identical to Megastore almost identical Basic
unit defined as Directory Directory Same prefix key, therefore adjacent Fine-grained mapping Interleaved rows gain performance
Two-phase commit for distributed transactions Two-phase commit 1Vote Coordinator Participants
Two-phase commit for distributed transactions Two-phase commit 2Commit Coordinator Participants
Locking remains a big issue Locking Especially when someone went
down, causing deadlock, literally.
Paxos is here to rescue, again Paxos will make sure
ALL logs are copied to every replicas. ALL logs
Real Innovation lies in time TrueTime API utilizes atomic clock
& GPS to determine the order of each transactions atomic clock GPS
NewSQL is the new NoSQL and Spanner is the best
example so far.