Building a high scale machine learning pipeline with ...biconsulting.hu/letoltes/2016budapestdata/bedodaniel_spark_final.pdf · Building a high scale machine learning pipeline with

Building a high scale machine learning pipeline

with Apache Spark and Kafka

https://www.flickr.com/photos/sanjayaprime/5013115478Bedő Dániel

• biggest community-driven question & answer website in Germany

• 20 million questions, 70 million answers

• similar to Quora, Yahoo Answers

Google Update Impact

Ordering of answers

supervised machine learning

determine the type of the training data

gather a training set

find a representation of the data

pick a learning algorithm

run the training algorithm

evaluate the accuracy

Regression Prototype

Identify the problems

• Model not complex enough

• Similar inputs, different outputs?

• Not enough training data

The old ETL pipeline

ETL v2Spark

Spark ecosystem

API

Scala Python Java R

Spark Streaming

Spark SQL MLLib GraphX

K

Kafka

Producer

Producer

Consumer

Consumer

Consumer

Topic 1 Partition 0

Broker 1

Topic 1 Partition 1

Broker 2

Topic 1 Partition 2

Broker 3

Kafka Cluster

Kafka topic

• scale

• parallelism

0 1 2 3 4 5 6 7 8 9

0 1 2 3 4 5 6 7

0 1 2 3 4 5 6 7 8

Producer

partition 0

partition 1

partition 2

Parquet

id cc votes

1 DE 2

2 DE 3

3 AT 1

4 DE 2

id cc votes

1 DE 2

2 DE 3

3 AT 1

4 DE 2

id cc votes

1 DE 2

2 DE 3

3 AT 1

4 DE 2

SELECT votes FROM logs WHERE cc = ‘AT’

push-down filters

Kafka

Rabbit MQServices

Tracking

Spark Cluster

KafkaKafka

HDFS(Tracking)

Stre

amin

g Worker

MySQL Read Slave

MySQL Master

Redis Master

Redis Read Slave

ElasticSearch

ETL v2

Project Moria

Clean training data?

Project Angmar

• tried lots of different supervised learning methods

• feature engineering - most crucial part

• analyse the domain, chart everything

FeaturesContent

length

syntactic complexity

number of links

probability of deletion

Social

votes

most helpful answer

number of comments

answered by expert

Author

gained votes

credibility score

role

deleted answer ratio

number of answers

number of comments

reported answer ratio

The structure of the network

21

3

1

0,2

0,4

0,1

2 0,8

Answer vector

AV normalized

0,9

0,6

0,2

0,1

Output

The Result

Calculate features for all answers

Batch Layer (Spark Batch)

Insert features in Redis

Calculate Score and store in MySQL

Speed Layer (Spark Streaming)

Listen for events

Insert or update Redis

Calculate Score and store in MySQL

Serving Layer

Lambda Architecture

Back pressure

• Bulk jobs insert too fast

• MySQL: sendQueue size, threads connected

• ElasticSearch: load on the instance creating the new index

Debugging the network

+1Change individual features

real world test (deleted vs non-deleted)

deletednon-deleted

Switching models

Amount of questions for a

score range

Old Score

New score

Insights

MenWomen

Learnings• If your use case is complex, you need a complex

model

• If you have a complex model, you need lots of data

• If you have lots of data, you need an ETL pipeline that can process huge amounts of data fast

• Think about your use case first, then design the pipeline

Questions?You can ask them on gutefrage too :)