Scalable Low-Latency Indexes for a Key-Value Store Talks/ankita.pdf · Scalable Low-latency Indexes...

Scalable Low-Latency Indexes for a Key-Value Store

Ankita Kejriwal

With Arjun Gopalan, Ashish Gupta, Greg Hill, Zhihao Jia, Stephen Yang

and John Ousterhout

A key value store can support

strongly consistent secondary indexes

while operating at low latency and large scale.

SLIK Slide 2

Thesis

●  Scalable Low-latency Indexes for a Key-value Store: SLIK §  Enables multiple secondary keys for each object §  Allows lookups and range queries on these keys

●  Key design features: §  Scalability using independent partitioning §  Strong consistency using an ordered write approach

●  Implemented in RAMCloud

●  Performance: §  Linear throughput increase with increasing number of partitions

§  11-13 µs indexed reads §  29-37 µs durable writes/overwrites of objects with one indexed attribute §  Latency approx. 2x non-indexed reads and writes

Summary of Results

Slide 3

SLIK Slide 4

Talk Outline

● Motivation ● Design

●  Performance

● Related Work

●  Summary

SLIK Slide 5

Motivation

Traditional RDBMs

MySQL

SLIK Slide 6

Motivation

Traditional RDBMs

NoSQL Systems

+ scalability

- data models - consistency MySQL

SLIK Slide 7

Motivation

Traditional RDBMs

NoSQL Systems

+ scalability


+ consistency

+ data models + low latency

H-Base Espresso

PNUTS Tao

Spanner Megastore HyperDex MongoDB H-Store

RAMCloud

Mem- cached

Yesquel

SLIK Slide 8

Motivation

Traditional RDBMs

NoSQL Systems

+ scalability


+ consistency

+ data models + low latency

H-Base Espresso

PNUTS Tao

Spanner Megastore HyperDex MongoDB H-Store

RAMCloud

Mem- cached

Yesquel

SLIK Slide 9

Talk Outline

● Motivation

● Design ●  Performance

● Related Work

●  Summary

●  Data model

●  Scalability

●  Strong consistency

●  Storage

●  Durability

●  Availability

SLIK Slide 10

Design

●  Scalability


SLIK Slide 11

Design

●  Scalability


SLIK Slide 12

Design

●  Scalability §  Nearly constant low latency irrespective of the server span §  Linear increase in throughput with the server span


SLIK Slide 13

Design

Index Partitioning: Colocation

Slide 14

●  Colocate index entries and objects

●  One of the keys used to partition the objects and indexes

Indexlet

Tablet

1 q rose

2 a tulip

3 n violet

Tablet

4 e clover

5 v daily

6 g iris

v è 5 e è 4 g è 6

Indexlet

Tablet

7 m lily

8 b dahlia

b è 8 m è 7

Indexlet

Server 1 Server 2 Server 3

n è 3 q è 1 a è 2

SLIK

primary key

index key

value index key

primary key


Slide 15

●  Colocate index entries and objects

●  One of the keys used to partition the objects and indexes

●  No association between index partitions and index key ranges

Indexlet

Tablet

1 q rose

2 a tulip

3 n violet

Tablet

4 e clover

5 v daily

6 g iris

v è 5 e è 4 g è 6

Indexlet

Tablet

7 m lily

8 b dahlia

b è 8 m è 7

Indexlet


n è 3 q è 1 a è 2 primary

key

index key

value index key

primary key Metadata:

tablet & indexlet w/ pk 1 to 3: S 1 tablet & indexlet w/ pk 4 to 6: S 2 tablet & indexlet w/ pk >= 7: S 3

Slide 16


Indexlet

Tablet

1 q rose

2 a tulip

3 n violet

a è 2 n è 3 q è 1

Tablet

4 e clover

5 v daily

6 g iris

v è 5 e è 4 g è 6

Indexlet

Tablet

7 m lily

8 b dahlia

b è 8 m è 7

Indexlet


SLIK

Client query: objects with index key between m - q

Slide 17



Indexlet

Tablet

1 q rose

2 a tulip

3 n violet

a è 2 n è 3 q è 1

Tablet

4 e clover

5 v daily

6 g iris

v è 5 e è 4 g è 6

Indexlet

Tablet

7 m lily

8 b dahlia

b è 8 m è 7

Indexlet


SLIK

Slide 18


Indexlet

Tablet

1 q rose

2 a tulip

3 n violet

a è 2 n è 3 q è 1

Tablet

4 e clover

5 v daily

6 g iris

v è 5 e è 4 g è 6

Indexlet

Tablet

7 m lily

8 b dahlia

b è 8 m è 7

Indexlet



SLIK

Slide 19


Not Scalable!

Indexlet

Tablet

1 q rose

2 a tulip

3 n violet

a è 2 n è 3 q è 1

Tablet

4 e clover

5 v daily

6 g iris

v è 5 e è 4 g è 6

Indexlet

Tablet

7 m lily

8 b dahlia

b è 8 m è 7

Indexlet



Index Partitioning: Independent

Slide 20

●  Partition each index and table independently

●  Partition each index according to sort order for that index

SLIK

Tablet

1 q rose

2 a tulip

3 n violet

Tablet

4 e clover

5 v daily

6 g iris

Tablet

7 m lily

8 b dahlia


Indexlet

a è 2 b è 8 e è 4 g è 6

Server 4

Indexlet

n è 3 m è 7 v è 5 q è 1

Server 5

primary key

index key

value index key

primary key


Slide 21

●  Partition each index and table independently

●  Partition each index according to sort order for that index

SLIK

Tablet

1 q rose

2 a tulip

3 n violet

Tablet

4 e clover

5 v daily

6 g iris

Tablet

7 m lily

8 b dahlia


Indexlet

a è 2 b è 8 e è 4 g è 6

Server 4

Indexlet

n è 3 m è 7 v è 5 q è 1

Server 5

primary key

index key

value index key

primary key

Metadata: tablet w/ pk 1 to 3: S 1 tablet w/ pk 4 to 6: S 2 tablet w/ pk >= 7: S 3 indexlet w/ sk a to g: S 4 indexlet w/ sk >= h: S 5


Slide 22 SLIK


Tablet

1 q rose

2 a tulip

3 n violet

Tablet

4 e clover

5 v daily

6 g iris

Tablet

7 m lily

8 b dahlia


Indexlet

a è 2 b è 8 e è 4 g è 6

Server 4

Indexlet

n è 3 m è 7 v è 5 q è 1

Server 5


Slide 23


Indexlet

a è 2 b è 8 e è 4 g è 6

Server 4

Indexlet

n è 3 m è 7 v è 5 q è 1

Server 5

Tablet

1 q rose

2 a tulip

3 n violet

Tablet

4 e clover

5 v daily

6 g iris

Tablet

7 m lily

8 b dahlia


SLIK


Slide 24


Tablet

1 q rose

2 a tulip

3 n violet

Tablet

4 e clover

5 v daily

6 g iris

Tablet

7 m lily

8 b dahlia


Indexlet

a è 2 b è 8 e è 4 g è 6

Server 4

Indexlet

n è 3 m è 7 v è 5 q è 1

Server 5

SLIK


Slide 25


Tablet

1 q rose

2 a tulip

3 n violet

Tablet

4 e clover

5 v daily

6 g iris

Tablet

7 m lily

8 b dahlia


Indexlet

a è 2 b è 8 e è 4 g è 6

Server 4

Indexlet

n è 3 m è 7 v è 5 q è 1

Server 5

Scalable!

Index Partitioning: Lookup Latency

0

20

40

60

80

100

0 10 20 30 40 50 60 70 80

Look

up L

aten

cy (µ

s)

Number of Servers

Colocation size 1Colocation size 10

Independent size 1Independent size 10

8.3

26.7

87.3

16.2

89.7

15.2 16.712.7

22.328.8

Index Partitioning: Lookup Throughput

0500

1000150020002500300035004000

0 2 4 6 8 10

Thro

ughp

ut (1

03 loo

kups

/sec

)

Number of Indexlets

Independent PartitioningColocation

634940

12491558

18592184

24782782

3092

335 461 423 463 447 457 447 357 441 418435

●  Scalability §  Nearly constant low latency irrespective of the server span §  Linear increase in throughput with the server span


SLIK Slide 28

Design

●  Scalability §  Nearly constant low latency irrespective of the server span §  Linear increase in throughput with the server span §  Solution: Use independent partitioning §  But: indexed object writes: distributed operations §  Potential consistency issues between indexes and objects


SLIK Slide 29

Design



SLIK Slide 30

Design


●  Strong consistency §  With minimal performance overheads

SLIK Slide 31

Design

SLIK Slide 32

Consistency Properties

●  If an object contains a given secondary key, then an index lookup with that key will return the object

●  If an object is returned by index lookup, then this object contains a secondary key for that index within the specified range

SLIK Slide 33




Frank

Bob Alice

Trent

Carol Peggy

SLIK Slide 34




students with name

between a – d?

Alice

Carol

Frank

Bob Alice

Trent

Carol Peggy

?

SLIK Slide 35




SLIK Slide 36




Frank

Bob Alice

Trent

Carol Peggy

SLIK Slide 37




students with name

between a – d?

Alice

Bob

Carol

Peggy

Frank

Bob Alice

Trent

Carol Peggy ?

SLIK Slide 38




students with name

between a – d?

Alice

Bob

Carol

Frank

Bob Alice

Trent

Carol Peggy

Consistency

Slide 39

●  Consistency properties: §  If an object contains a given secondary key, then an index lookup with

that key will return the object §  If an object is returned by index lookup, then this object contains a

secondary key for that index within the specified range

●  Solution: §  Longer index lifespan (via ordered writes) §  Object data is ground truth and index entries serve as hints

Consistency

Slide 40





Foo Bob ... Object

Index Entry

Bob è Foo

Consistency

Slide 41





Foo Bob ... Object

Index Entry

Bob è Foo Sam è Foo 1.  Add new index entry

Consistency

Slide 42





Foo Sam ... Object

Index Entry

Bob è Foo Sam è Foo 1.  Add new index entry 2.  Modify object

Consistency

Slide 43





Foo Sam ... Object

Index Entry

Sam è Foo 1.  Add new index entry 2.  Modify object 3.  Remove old index entry

Consistency

Slide 44





time

Bob è Foo

Sam è Foo

Foo Bob ... Foo Sam ... Object

Index Entry

modify object remove object write object

Consistency

Slide 45





time

Bob è Foo

Sam è Foo


Index Entry


Consistency

Slide 46





time

Bob è Foo

Sam è Foo

commit point commit point commit point


Index Entry


x

Consistency

Slide 47





time

Bob è Foo

Sam è Foo

commit point commit point commit point


Index Entry



●  Strong consistency §  With minimal performance overheads §  Solution: Ordered write approach + treat indexes as hints

SLIK Slide 48

Design

SLIK Slide 49

Talk Outline

● Motivation

● Design

● Performance

● Related Work

●  Summary

●  Does SLIK provide low latency?

●  Does SLIK provide scalability?

●  How does the performance of indexing with SLIK compare to other state-of-the-art systems?

SLIK Slide 50

Performance: Questions

●  H-Store: §  Main memory database §  Data (and indexes) partitioned based on specified attribute §  Many parameters for tuning

●  Got assistance from developers to tune for each test ●  Examples: txn_incoming_delay, partitioning column

●  HyperDex: §  Spaces containing objects §  Data (and indexes) partitioned using hyperspace hashing §  Each index contains all object data §  Designed to use disk for storage

Slide 51

Performance: Systems for Comparison

SLIK

SLIK Slide 52

Hardware

Slide 53

Latency

Experiments:

1.  Lookups: table with single secondary index

2.  Overwrites: table with single secondary index

3.  Overwrites: varying number of secondary indexes

Configuration:

●  Single client

●  Single partition for table and (each) index

●  Object: 30 B pk, 30 B sk, 100 B value

●  SLIK: Three-way replication to durable backups

●  H-Store: No replication, durability disabled, single server

Slide 54

Lookup Latency

10

100

1000

100 101 102 103 104 105 106

(c) O

verw

rite

Late

ncy

(µs)

Size of Index (# objects)

H-Store SK Partitioned

179.53 160.16195.06 202.07 192.46 207.10 209.37

H-Store PK Partitioned

987.25 939.84 961.17 1019.82 1048.86 1010.92 968.72

HyperDex

648 667 671 667 660 657 661

SLIK tcp

124.4 135.8 126.5 125.9 124.3 123.8 129.7

SLIK

31.4 32.7 34.2 34.3 35.2 35.2 37.0

0

50

100

150

200

250

100 101 102 103 104 105 106

(a) L

ooku

p La

tenc

y (µ

s)

H-Store

142.10 138.18150.11 152.37

134.97147.42

132.00

SLIK TCP

45.2 44.9 42.7 48.5 49.7 45.654.5

SLIK

11.0 10.2 11.7 11.6 12.7 12.8 13.1

Slide 55

Lookup Latency

10

100

1000

100 101 102 103 104 105 106

(c) O

verw

rite

Late

ncy

(µs)


H-Store SK Partitioned

179.53 160.16195.06 202.07 192.46 207.10 209.37

H-Store PK Partitioned

987.25 939.84 961.17 1019.82 1048.86 1010.92 968.72

HyperDex

648 667 671 667 660 657 661

SLIK tcp

124.4 135.8 126.5 125.9 124.3 123.8 129.7

SLIK

31.4 32.7 34.2 34.3 35.2 35.2 37.0

0

50

100

150

200

250

100 101 102 103 104 105 106

(a) L

ooku

p La

tenc

y (µ

s)

H-Store

142.10 138.18150.11 152.37

134.97147.42

132.00

SLIK TCP

45.2 44.9 42.7 48.5 49.7 45.654.5

SLIK

11.0 10.2 11.7 11.6 12.7 12.8 13.1

Overwrite Latency

Slide 56

0

50

100

150

200

250

100 101 102 103 104 105 106

(b) O

verw

rite

Late

ncy

(µs)


H-Store

143.51 143.00151.39 153.28

137.74148.88

133.54

SLIK TCP

124.4135.8

126.5 125.9 124.3 123.8 129.7

SLIK

31.4 32.7 34.2 34.3 35.2 35.2 37.0

Multiple Secondary Indexes

Slide 57

10

100

1000

0 1 2 3 4 5 6 7 8 9 10

Over

writ

e La

tenc

y (µ

s)

Number of Indexes

H-Store via PK

152.97

265.91 273.41 285.99 287.72

H-Store via SK

181.98

1474.39 1682.16 1704.69 1756.99

SLIK TCP

138.5 139.1 156.2 165.2 164.3 175.3 175.7 179.1 182.0 184.6

SLIK

33.0 35.3 39.8 39.0 42.7 42.4 46.4 47.3 49.5 51.2


Slide 58

10

100

1000

0 1 2 3 4 5 6 7 8 9 10

Over

writ

e La

tenc

y (µ

s)

Number of Indexes

H-Store via PK

152.97

265.91 273.41 285.99 287.72

H-Store via SK

181.98

1474.39 1682.16 1704.69 1756.99

SLIK TCP

138.5 139.1 156.2 165.2 164.3 175.3 175.7 179.1 182.0 184.6

SLIK

33.0 35.3 39.8 39.0 42.7 42.4 46.4 47.3 49.5 51.2


Slide 59

10

100

1000

0 1 2 3 4 5 6 7 8 9 10

Over

writ

e La

tenc

y (µ

s)

Number of Indexes

H-Store via PK

152.97

265.91 273.41 285.99 287.72

H-Store via SK

181.98

1474.39 1682.16 1704.69 1756.99

SLIK TCP

138.5 139.1 156.2 165.2 164.3 175.3 175.7 179.1 182.0 184.6

SLIK

33.0 35.3 39.8 39.0 42.7 42.4 46.4 47.3 49.5 51.2

Slide 60

Scalability

Experiments:

1.  Lookup throughput with increasing number of partitions

2.  Lookup latency with increasing number of partitions

Configuration:

●  Single table with one secondary index

●  Table and index partitioned across servers

●  Object: 30 B pk, 30 B sk, 100 B value

●  Throughput experiment: Loaded system

●  Latency experiment: Unloaded system

Slide 61

Scalability: Lookup Throughput

0

1000

2000

3000

4000

5000

0 2 4 6 8 10 12 14 16 18 20

Thro

ughput

(10

3

lookups/s

ec)

Number of Servers

H-Store

SLIK TCP

SLIK

96.90312.35 347.32 396.27

430653

7941001

11991352 1445

16631807

220

580

1127

1619

2197

2655

3199

3629

4248

4629

5069

Slide 62

Scalability: Lookup Latency

0

50

100

150

200

250

300

0 1 2 3 4 5 6 7 8 9 10

Avera

ge L

ate

ncy p

er

Lookup (

µs)

Number of Indexlets

H-Store

119.4

152.6

179.6

199.2

215.6

243.4 240.3

267.4 269.8 267.2

SLIK TCP

49.955.6 58.2

69.181.6

90.6 91.0

112.7 114.4 113.3

SLIK

13.1 13.3 13.9 13.7 14.4 14.7 14.5 14.4 14.6 14.7

SLIK Slide 63

Talk Outline

● Motivation

● Design

●  Performance

● Related Work ●  Summary

SLIK Slide 64

Related Work

Data storage system §  Data model (spectrum from key-value to relational) §  Consistency (spectrum from eventual to strong) §  Performance: latency and/or throughput

Current Web Scale Datastores

Slide 65 SLIK

100s 10s 1s 100ms 10ms 1ms 100µs

Eventual

Strong

10µs

Causal, SI, “Define

your own”

Read / write latency (approx)

Con

sist

ency

Lev

el

Better

Bett

er


Slide 66 SLIK

100s 10s 1s 100ms 10ms 1ms 100µs

Eventual

Strong

10µs


your own”


Con

sist

ency

Lev

el

Better

Bett

er

CouchDB Tao

Espresso PNUTS


Slide 67 SLIK

100s 10s 1s 100ms 10ms 1ms 100µs

Eventual

Strong

10µs


your own”


Con

sist

ency

Lev

el

Better

Bett

er

CouchDB Tao

H-Store

HyperDex

MongoDB Spanner

Espresso

H-Base

PNUTS


Slide 68 SLIK

100s 10s 1s 100ms 10ms 1ms 100µs

Eventual

Strong

10µs


your own”


Con

sist

ency

Lev

el

Better

Bett

er

Tao

Cassandra

CouchDB

MegaStore

H-Store

HyperDex

MongoDB Spanner

Espresso

H-Base

PNUTS

SLIK

Cassandra

CouchDB Tao

MegaStore

H-Store

HyperDex

MongoDB Spanner

Espresso

H-Base

PNUTS


Slide 69 SLIK

100s 10s 1s 100ms 10ms 1ms 100µs

Eventual

Strong

10µs


your own”


Con

sist

ency

Lev

el

Better

Bett

er

SLIK Slide 70

Talk Outline

● Motivation

● Design

●  Performance

● Related Work

● Summary

A key value store can support

strongly consistent secondary indexes

while operating at low latency and large scale.

SLIK Slide 71

Summary

Lookups and range queries on secondary keys

By using approaches that have minimal overheads we get: 11-13 µs lookups and 30-37 µs (over)writes

By using ordered writes and treating indexes as hints

By using independent partitioning we get: linear throughput increase and minimal impact on latency as the scale increases

Thank you!

Code available free and open source: github.com/PlatformLab/RAMCloud My papers and other information at: http://stanford.edu/~ankitak

I can be reached at: [email protected]

Date post:	14-May-2020
Category:	Documents
Upload:	others
View:	18 times
Download:	0 times

Scalable Low-Latency Indexes for a Key-Value Store Talks/ankita.pdf · Scalable Low-latency Indexes...

Documents