1 of 26

Taking Omid to the Clouds:�Fast, Scalable Transactions for Real-Time Cloud Analytics

1

Ohad Shacham

Edward Bortnikov

Eshcar Hillel

Idit Keidar

Aran Bergman

Yonatan Gottesman

2 of 26

Agenda

Introduction to Omid (FAST’17)

​

Omid low latency optimizations (VLDB’18)

​

Apache Phoenix integration

2

3 of 26

Agenda

Introduction to Omid

​

Omid low latency optimizations

​

Apache Phoenix integration

3

4 of 26

Omid (Hope in Persian)

Transactional API over NoSQL key value

​

Client Library + Runtime Service

​

Snapshot Isolation consistency

​

Highly available

​

Open source Apache incubator

​

4

5 of 26

Transactions and Snapshot Isolation

Aborts only write-write conflict

5

Begin

Commit

read(x)

Write(x)

Begin

Commit

Write(x)

Read(x)

6 of 26

Omid Architecture - Write

6

Transaction Manager

Client

Begin

Data store

Data store

Data store

Commit table

Timestamp - tsr

commit

Put(k1,v1,tsr)

API

tsr

tsc

Commit(k1,k2…)

tsr

Conflict Detection

Timestamp – tsc

K1

tsr

V1

tsc

7 of 26

Omid Architecture – Read

7

Transaction Manager

Client

Begin

Data store

Data store

Data store

Commit table

Timestamp – tsr’

Get tsc

API

Get(k1,t<=tsr’)

tsr’

K1

tsr

V1

tsr

tsc

8 of 26

Omid Bottleneck

  • Omid preferred throughput over latency – batch
  • High latency

8

Transaction Manager

Commit table

Bottleneck

9 of 26

Taking Omid to the Cloud

Introduction to Omid

​

Omid low latency optimizations

​

Apache Phoenix integration

9

10 of 26

Omid Low Latency

Distribute commit table writes

    • Remove Omid bottleneck

​

Fast path transactions

    • New API for single key transactions

​

10

11 of 26

Distribute Commit Table Updates

11

Transaction Manager

Client

Begin

Data store

Data store

Data store

Commit table

Timestamp - tsr

Put(k1,v1,tsr)

API

tsr

tsc

Commit(k1,k2…)

tsr

Conflict Detection

Timestamp – tsc

K1

tsr

V1

Not so trivial!

12 of 26

SI Violation Example

12

Transaction Manager

Client 1

Data store

Data store

Data store

Commit table

API

tsr

tsc

K1

tsr

V1

K2

tsr

V2

Client 2

API

Put

Get(k1)

Begin

Timestamp - tsr

Commit(k1,k2…)

tsr

Timestamp – tsc

Get(k2)

13 of 26

SI Violation Solution - Invalidation

13

Transaction Manager

Client 1

Data store

Data store

Data store

Commit table

API

K1

tsr

tsr

INVALID

V1

K2

tsr

V2

Client 2

API

Get(k1)

Timestamp - ts

Commit(k1,k2…)

tsr

Invalidate tsr

14 of 26

Fast Path Transactions

Many workloads have singe key transactions

​

Wasteful access to TM for timestamps

​

New API – Only access data table without TM

    • brc(key)
    • bwc(key,val)
    • br(key) + wc(key,val)

​

​

​

14

15 of 26

Fast Path Transactions

15

Transaction Manager

Client

Data store

Commit table

API

Local Clock

Global Clock

Data store

Local Clock

bwc(k1,v1)

​

K1

tsr

V1

tsc

16 of 26

SI Requires Local Validation

16

Transaction Manager

Client 1

Commit table

API

Global Clock

Client 2

API

Begin

Timestamp - tsr

bwc(k1,v1)

​

Put

Conflict Detection

Data store

Local Clock

Data store

Local Clock

K1

tsr

V1

tsc

17 of 26

Evaluation

HBase cluster

​

YCSB

​

Transaction sizes 1-10

17

18 of 26

Throughput Latency

18

Transaction Size = 1

Transaction Size = 10

7X

2.5X

19 of 26

Latency Breakdown

19

Single key transactions

Begin Time

​

Data Time

​

Commit Time

20 of 26

Agenda

Introduction to Omid

​

Omid low latency optimizations

​

Apache Phoenix integration

20

21 of 26

Apache Phoenix

SQL interface over HBase

Transforms SQL queries into native HBase API calls

​

Requires transaction manager for

  • SQL transactions
  • Consistent secondary index

21

22 of 26

Omid Integration

Support Phoenix coprocessors

    • Integrate Omid’s code within HBase coprocessors

​

Add functionality

    • Construction of secondary indexes
    • Snapshot isolation exclude current - SIX

22

23 of 26

Secondary Index Creation

Phoenix on the fly index creation

    • Secondary index from non empty table
    • During creation table get updated

​

New fence API

    • Abort transaction that overlap fence

​

23

24 of 26

Snapshot Isolation Exclude Current - SIX

BEGIN;

​

INSERT INTO T

SELECT ID+10 FROM T;

​

COMMIT;

​

​

​

24

read

write

25 of 26

Checkpoints

BEGIN;

​

INSERT INTO T

SELECT ID+10 FROM T;

​

COMMIT;

​

​

​

25

checkpoint

read

write

26 of 26

Summary

Omid is a mature transactional layer over HBase

​

Omid low latency improves throughput latency and scalability

​

Integrated into Phoenix with new features

​

Available in Omid release 1.0.1 and Phoenix releases 4.15 and 5.1

26