External Consistency and Spanner CS425/ECE428 SPRING 2020 NIKITA - - PowerPoint PPT Presentation

External Consistency and Spanner

CS425/ECE428 — SPRING 2020 NIKITA BORISOV, UIUC

Transactions so far

Objects distributed / partitioned among different servers

• For load balancing (sharding)
• For separation of concerns / administration

Isolation enforced using two-phase locking (2PL)

• Each server maintains locks on own objects
• Deadlocks detected using e.g., edge-chasing

Atomic commit using 2PC

• Prepare to commit ensures durability
• Recover from coordinator and participant crashes

Dealing with Failures

Node failure

• Objects unavailable until recovery
• 2PC “stuck” after coordinator failure

But! Node failure is common Drive failures => no recovery!

Replication

Objects distributed among 1000’s cluster nodes for load-balancing (sharding) Objects replicated among a handful of nodes for availability / durability

• Replication across data centers, too

Two-level operation:

• Use transactions, coordinators, 2PC per object
• Use Paxos / Raft among object replicas

Note: can be expensive!

• Coordinator sends Prepare message to leaders of each replica group
• Each leader uses Paxos / Raft to commit the Prepare to the group logs
• Once commit succeeds, reply to coordinator
• Coordinator uses Paxos / Raft to commit decision to its group log

Example transaction

read A -> acquire read lock on A read B -> acquire read lock on B write A -> promote A’s lock to write lock commit -> perform 2PC

• Coordinator -> A, B: prepare
• A, B -> OK
• Coordinator -> A, B: commit

Read transactions

Read transactions often access many data items

• E.g., Facebook ”news feed”
• E.g., Amazon front page
• E.g., balances across all accounts

Read transactions still need (read) locks (Why?) Acquiring locks requires consensus (Why?) Locks prevent write transactions from moving forward

Linearizability

Serial equivalence:

• Total effect on system is equivalent to a run that is serial and consistent with each client’s order

Linearizability

• Total effect on system is equivalent to a run that is serial and consistent with actual order of events

E.g., buying a movie

• Client makes RPC to bank transfers $3.99 to Amazon account
• Client requests video from Amazon
• Amazon makes RPC to bank, does not see transfer, rejects request!

Spanner: Google’s Globally-Distributed Database

Wilson Hsieh representing a host of authors OSDI 2012

What is Spanner?

• Distributed multiversion database
• General-purpose transactions (ACID)
• SQL query language
• Schematized tables
• Semi-relational data model
• Running in production
• Storage for Google’s ad data
• Replaced a sharded MySQL database

OSDI 2012 9

Example: Social Network

OSDI 2012 User posts Friend lists User posts Friend lists User posts Friend lists User posts Friend lists US Brazil Russia Spain San Francisco Seattle Arizona Sao Paulo Santiago Buenos Aires Moscow Berlin Krakow London Paris Berlin Madrid Lisbon User posts Friend lists 10 x1000 x1000 x1000 x1000

Overview

• Feature: Lock-free distributed read transactions
• Property: External consistency of distributed

transactions

– First system at global scale

• Implementation: Integration of concurrency

control, replication, and 2PC

– Correctness and performance

• Enabling technology: TrueTime

– Interval-based global time

OSDI 2012 11

Read Transactions

• Generate a page of friends’ recent posts

– Consistent view of friend list and their posts

OSDI 2012

Why consistency matters

• 1. Remove untrustworthy person X as friend
• 2. Post P: “My government is repressive…”

12

User posts Friend lists User posts Friend lists

Single Machine

Friend2 post Generate my page Friend1 post Friend1000 post Friend999 post Block writes OSDI 2012 … 13

User posts Friend lists User posts Friend lists

Multiple Machines

User posts Friend lists Generate my page Friend2 post Friend1 post Friend1000 post Friend999 post User posts Friend lists Block writes OSDI 2012 … 14

User posts Friend lists User posts Friend lists User posts Friend lists

Multiple Datacenters

User posts Friend lists Generate my page Friend2 post Friend1 post Friend1000 post Friend999 post OSDI 2012 … US Spain Russia Brazil 15 x1000 x1000 x1000 x1000

Version Management

• Transactions that write use strict 2PL

– Each transaction T is assigned a timestamp s – Data written by T is timestamped with s

OSDI 2012 16 Time 8 <8 [X] [me] 15 [P] My friends My posts X’s friends [] []

Synchronizing Snapshots

== External Consistency: Commit order respects global wall-time order

OSDI 2012 17

== Timestamp order respects global wall-time order given timestamp order == commit order

Global wall-clock time

Timestamps, Global Clock

• Strict two-phase locking for write transactions
• Assign timestamp while locks are held

T Pick s = now() Acquired locks Release locks OSDI 2012 18

Timestamp Invariants

OSDI 2012 19

• Timestamp order == commit order
• Timestamp order respects global wall-time order

T2 T3 T4 T1

TrueTime

• “Global wall-clock time” with bounded

uncertainty

time earliest latest TT.now() 2*ε OSDI 2012 20

Timestamps and TrueTime

T Pick s = TT.now().latest Acquired locks Release locks Wait until TT.now().earliest > s s OSDI 2012 average ε Commit wait average ε 21

Commit Wait and Replication

OSDI 2012 T Acquired locks Release locks Start consensus Notify slaves Commit wait done Pick s 22 Achieve consensus

Commit Wait and 2-Phase Commit

OSDI 2012 TC Acquired locks Release locks TP1 Acquired locks Release locks TP2 Acquired locks Release locks Notify participants of s Commit wait done Compute s for each 23 Start logging Done logging Prepared Compute overall s Committed Send s

Example

OSDI 2012 24 TP Remove X from my friend list Remove myself from X’s friend list sC=6 sP=8 s=8 s=15 Risky post P s=8 Time <8 [X] [me] 15 TC T2 [P] My friends My posts X’s friends 8 [] []

What Have We Covered?

• Lock-free read transactions across datacenters
• External consistency
• Timestamp assignment
• TrueTime

– Uncertainty in time can be waited out

OSDI 2012 25

What Haven’t We Covered?

• How to read at the present time
• Atomic schema changes

– Mostly non-blocking – Commit in the future

• Non-blocking reads in the past

– At any sufficiently up-to-date replica

OSDI 2012 26

TrueTime Architecture

Datacenter 1 Datacenter n … Datacenter 2 GPS timemaster GPS timemaster GPS timemaster Atomic-clock timemaster GPS timemaster Client OSDI 2012 27 GPS timemaster

Compute reference [earliest, latest] = now ± ε

TrueTime implementation

time ε 0sec 30sec 60sec 90sec +6ms

now = reference now + local-clock offset ε = reference ε + worst-case local-clock drift

reference uncertainty OSDI 2012 28 200 μs/sec

What If a Clock Goes Rogue?

• Timestamp assignment would violate external consistency
• Empirically unlikely based on 1 year of data

– Bad CPUs 6 times more likely than bad clocks

OSDI 2012 29

Network-Induced Uncertainty

OSDI 2012 Mar 29 Mar 30 Mar 31 Apr 1

Date

2 4 6 8 10

Epsilon (ms)

99.9 99 90 6AM 8AM 10AM 12PM

Date (April 13)

1 2 3 4 5 6 30

Conclusions

• Reify clock uncertainty in time APIs

– Known unknowns are better than unknown unknowns – Rethink algorithms to make use of uncertainty

• Stronger semantics are achievable

– Greater scale != weaker semantics

OSDI 2012 33