SLIDE 1
Summary Related work Our work- Wren Introduction
WR WREN: A A Fa Fast and Scalable Transactional Causally Co - - PowerPoint PPT Presentation
WR WREN: A A Fa Fast and Scalable Transactional Causally Co - - PowerPoint PPT Presentation
WR WREN: A A Fa Fast and Scalable Transactional Causally Co Consistent Geo-Re Replicated Ke Key-Va Value Store Diego Didona, Kristina Spirovska, Willy Zwaenepoel RainbowFS workshop Paris, May 3rd Outline Introduction Our work- Wren
SLIDE 2
SLIDE 3
3
Geo-replicated partitioned key-value data store
A-D E-H J-M
SLIDE 4
Consistency
Causal Consistency Eventual Consistency Linearizability (Strong Consistency)
Performance Consistency Guarantees
4
SLIDE 5
5
Causal Consistency
Alice
7 minutes ago ·
I’ll miss the train ;(
Like · Comment · Share 0 people like this.
Write a comment … J
>
AliceYes! Caught it J
Like · Reply · 2 minutes ago
Bob Happy for you!
Like · Reply · 1 minute ago
Alice
7 minutes ago ·
I’ll miss the train ;(
Like · Comment · Share 0 people like this.
Write a comment … J
>
Bob Happy for you!
Like · Reply · 1 minute ago
Eve
SLIDE 6
Transactional Causal Consistency
Causal Consistency Read-write Transactions Transactional Causal Consistency
- Transactions read from a causally consistent snapshot
- Transactions updates are visible atomicallyin a DC
6
SLIDE 7
Summary Related work Our work- Wren Introduction
Outline
SLIDE 8
8
Limitations of State-of-the-art Systems
- Metadata scalability
Dependency tracking
- Clock skew induced latency
Update timestamping
SLIDE 9
9
Main Contributions
WREN
- 2. Low Latency
Loose synchronization of physical clocks without suffering from clock skew A novel dependency tracking and stabilization protocol
Hybrid Stable Time (HST)
- 1. Constant Metadata
Hybrid Logical/ Physical Clocks
Addresses metadata scalability Addresses clock skew induced latency
SLIDE 10
10
Main Contributions
WREN
- 2. Low Latency
Loose synchronization of physical clocks without suffering from clock skew A novel dependency tracking and stabilization protocol
Hybrid Logical/ Physical Clocks Hybrid Stable Time (HST)
- 1. Constant Metadata
SLIDE 11
- Only two scalar timestamps
11
Hybrid Stable Time (HST)
1 2
Tracks the dependencies on local items Summarizes dependencies on remote items
Local Dependency Time Remote Stable Time
SLIDE 12
Existing dependency tracking methods
12
Dependency vectors
Fresh, non-blocking snapshot
Single timestamp Cure[ICDCS’16] GentleRain[SOCC’14]
Metadata size O(#DCs)
+ ⎯
Metadata size O(1) Inter DC sync to install snapshot
+ ⎯
SLIDE 13
13
Single Scalar: GentleRain
DC 0 DC 1 DC 2 Lower Bound: Client
Write Read BLOCK untill: lower bound >= 5 7 15 15 Client dependency time 15 15
5
O(1)
SLIDE 14
14
Dependency vectors: Cure
DC 0 DC 1 DC 2 Lower Bound:
[ 5, 7, 15]
Client
Write 5 7 15 15 Client dependency time
O(n)
SLIDE 15
15
Dependency vectors: Cure
DC 0 DC 1 DC 2 Lower Bound:
[ 5, 7, 15]
Client
Write Read NO BLOCKING 5 7 15 15 Client dependency time [5,7,15]
O(n)
SLIDE 16
16
Local and Remote Time: Wren
DC 0 DC 1 DC 2 Lower Bound:
[ 5, 15]
Client
Write 5 7 15 15 Client dependency time
O(1)
SLIDE 17
17
Local and Remote Time: Wren
DC 0 DC 1 DC 2 Lower Bound:
[ 5, 15]
Client
Write Read 5 7 15 15 Client dependency time [5,15]
O(1)
NO BLOCKING
SLIDE 18
- Only two scalar timestamps
18
Hybrid Stable Time (HST)
Tracks the dependencies on local items Summarizes dependencies on remote items
Local Dependency Time Remote Stable Time
SLIDE 19
- Computed periodically
- Lower bound on updates from remote DCs
- No additional inter DCs sync
19
Remote Stable Time (RST)
SLIDE 20
HST Benefits & Trade-off
- Hits sweet spot in the meta-data size vs
performance spectrum
.
20
SLIDE 21
21
Main Contributions
WREN
- 2. Low Latency
Loose synchronization of physical clocks without suffering from clock skew A novel dependency tracking and stabilization protocol
Hybrid Stable Time (HST)
- 1. Constant Metadata
Hybrid Logical/ Physical Clocks
SLIDE 22
22
Existing update timestamping methods
Physical Clocks PC
(Loose) clock synchr. enables efficient dependency tracking
Logical Clocks LC
Clock skew introduces uncertainty (latencies)
+ ⎯
No syhncronization makes dependency tracking more costly Capture dependency among events easily
+ ⎯
SLIDE 23
23
Hybrid Logical/Physical clocks (HCL)
- Logical Physical Clocks, OPODIS’14
- Be
Best of
- f bot
- th wo
worlds
- Ca
Captures the causality y re relationship
- In
Inher erits loose e sy synchronization of
- f PC
PC
- No
No clock skew as LC LC
SLIDE 24
24
Hybrid Logical/Physical clocks (HCL)
p :l
Physical component Logical component
SLIDE 25
25
Clock skew problem
Client
Write(x2)
B L O C K
p
Wait until <
t physical clock
t
10
t client dependency time
10
x2
5 11
In Invariant: timestamps must reflect causality
Cl Client’s dep. time < x2’s update time
SLIDE 26
26
Clock skew solution
Client
Write(x2)
p
hybrid clock client dependency time
10 :0 5 :0
p:l p:l
t physical clock
5
Update hybrid clock
In Invariant: timestamps must reflect causality
Cl Client’s dep. time < x2’s update time
SLIDE 27
27
Clock skew solution
Client
Write(x2)
p
hybrid clock client dependency time
x2
10 :0
p:l p:l
t physical clock
5 10 :1
SLIDE 28
28
Availability
HS HST always tracks stable remote dependencies The system will continue to work, everything is there locally Do not block in case of network partitions
SLIDE 29
Summary Related work Our work- Wren Introduction
Outline
SLIDE 30
30
Related Systems
Low Latency Metadata Availability (Transact.) Snapshot freshness Cure [ICDCS’16] Clock skew #DCs Yes Higher GentleRain [SOCC’14] Clock skew
Inter-DC syn
O(1) No Lower Occult [NSDI’17] Inter DC sync O(#DCs) No Highest Wren Yes O(1) Yes Lower
SLIDE 31
The first transactional causally consistent geo-replicated system that at the same time has:
31
Our work: Wren
Constant metadata Low latency Always-available
SLIDE 32
- Wren: The first transactional causally consistent
geo-replicated system that at the same time has:
- Constant metadata
- Hybrid Stable Time
- Low latency
- Hybrid Logical/Physical Clocks
- Always-available
32