Big Data Computing Architecture

Evolution of Big Data Architecture

Gang Tao

Computing Trend

Non Functional Requirements

• Latency

• Throughput

• Fault Tolerant

• Scalable

• Exactly Once Semantic

Hadoop

Hadoop Eco System

Map Reduce

Move computation to Data

Map Reduce

Hadoop the Limitation

• Map/Reduce is hard to use

• Latency is high

• inevitable data movement

Lambda

Lambda Architecture

query = function(all data)

Design Principle

human fault-tolerance – the system is unsusceptible to data loss or data corruption because at scale it could be irreparable.

data immutability – store data in it’s rawest form immutable and for perpetuity. (INSERT/ SELECT/DELETE but no UPDATE !)

recomputation – with the two principles above it is always possible to (re)-compute results by running a function on the raw data.

Lambda Architecture

Batch Processing

Batch Processing

Using only batch processing, leaves you always with a portion of non- processed data.

Realtime Stream Processing

query = function(all data) + function(new data)

Lambda Architecture

A Reference Case

Lambda the Bad?

What is Good?

What is Bad?

Inmutable Data , Reprocessing

Keeping code written in two different systems perfectly in sync

Over complicated Lambda

Batch VS. Stream

Stream & Realtime

Storm Logic View

Bolts

Spouts

Move data to Computation

Storm Deployment View

Code Compare

DAG• DAG : Directed Acyclic Graph

• Used in Spark, Storm, Flink etc.

Out of Control

Build Complexity with Simplicity

Stream Processing Model

One at a time Micro batch

Stream Processing Model

One at a time Micro Batch

Low Latency Y N

High Throughput N Y

at least once Y Y

excatly once Sometimes Y

simple programing model Y N

Stream Computing the Limitation• Queries must be written before data

• There should be another way to query past data

• Queries cannot be run twice

• All results will be lost when any error occurs All data have gone when bugs found

• Disorders of events break results

• Recorded time based queries? Or arrival time based queries?

Batch/Stream Unification

Stream with Spark

Apache Flink

Flink Distributed Snapshot

Fault Tolerance in Stream

• At Least Once : ensure all operators see all events

• Stream -> Replay on failure

• Exactly Once :

• Flink : distributed Snapshot

• Spark : Micro Batch

Jay Kreps

Kappa Architecture