Real-time Analytics with Spark Maciej Dabrowski, Chief Data Scientist, Altocloud Galway Data Meetup, 2015-02-03
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Real-time Analytics with Spark
Maciej Dabrowski, Chief Data Scientist, Altocloud !Galway Data Meetup, 2015-02-03
2
MEETS A SMALL STARTUP
source: https://media.licdn.com/mpr/mpr/p/1/005/0a0/167/2f98d60.jpg
‣ We built predictive communications software that uses analytics to make customer interactions and experience better
Altocloud
3
Monitoring live users
4
5
‣ Real-time for us is under 1-5s
‣ Q: How many customers are currently online?
‣ Q: How many chats/calls are taking place at the moment?
‣ Q: What is the utilisation of my customer support agents?
Use Case 1: Real-time analytics
7
‣ Q: How many calls were offered in the last week?
‣ Q: What is the acceptance rate of my chat offers?
Use Case 2: Reporting
8
‣ Q: Which customers currently on my site I should engage?
Use Case 3: Predictive Analytics
9
‣ Scalability
‣ Limited resources
‣ Various analytics use cases
Technical challenges
10
11
Real-time analytics with Hadoop
source: http://barbarashdwallpapers.com/funny-elephant-wallpapers/
APIs
QUERYING LAYER
STORAGE LAYER
PROCESSING LAYER
Altocloud Platform
12
MESSAGE QUEUES
FRONT-END APIs KAFKA
SPARK
RABBIT MQ
CASSANDRA
SPARK STREAMING
HDFS
BACK-END APIS
APPS
BACK-END APIs
MONGODB
DATA SOURCES
QUERYING LAYER
STORAGE LAYER
PROCESSING LAYER
Altocloud Data Platform
13
MESSAGE QUEUES
FRONT-END APIs KAFKA
MONGODB OPLOG
SPARK
RABBIT MQ
CASSANDRA
SPARK STREAMING
HDFS
FRONT-END APIS
APPS
MONGODB
‣ One code base for streaming and batch processing
‣ Rich API in Scala/Python/Java
‣ Fast for iterative algorithms (important for ML)
‣ Growing community
‣ The concept of a micro-batch
‣ Nicely integrates with Kafka and Cassandra
‣ Fairly easy setup
Why Spark
14
Spark components
15
‣ Hadoop
!
!
!
!
!
!
‣ Spark
Word count in Spark
16
‣ Example: user event aggregation stored in Cassandra
‣ Still much better than Hadoop!
What about something more useful?
17
‣ User activity is an input (e.g. page view)
‣ Users for multiple businesses online
‣ Scale 100s to 100 000s activities per second
‣ Response time under 5s
‣ A perfect use case for spark streaming
Counting users currently online
18
‣ Pub-sub message broker
‣ Fast: 100s MBs /s on a single broker
‣ Scalable: partitioned data streams
‣ Durable: messages persisted and replicated
‣ Distributed: Strong durability with and fault-tolerance
‣ Downside: requires ZooKeeper
!see https://kafka.apache.org
Data source: Kafka
19
!
!
!
!
!
!
!‣ Kafka with Spark: http://www.michael-noll.com/blog/2014/10/01/kafka-spark-streaming-integration-example-tutorial/
Spark and Kafka
20
‣ Simple count unique events
!
!
‣ Count visit events for unique users
Count users online
21
‣ Twitter Algebird to the rescue!
‣ HyperLogLog - a probabilistic data structure saving a lot of memory!
‣ https://github.com/twitter/algebird
Sets can take a lot of memory!
22
‣ Easy to setup
‣ High availability - no master
‣ Great performance
‣ CQL - SQL like querying
‣ Great support and bug-free drivers from Datastax
‣ Key: Design your schema around queries; !!
see https://cassandra.apache.org
Storing your results
23
‣ Datastax driver is very easy to use
!
!
‣ Save our results to Cassandra
Store data in Cassandra
24
25source: http://top1walls.com
‣ Spark streaming job performs two major tasks:
• data processing • data receiving
‣ Receiver always takes one core
‣ Technically, you need 2N cores to run N streaming jobs
‣ Not a big deal in production, what about testing?
Spark streaming
26
‣ Containerise your app including all its dependencies
‣ Distribute your app in this standard container
‣ Run it on any machine with docker
‣ Very lightweight
Docker
27
c3.xlarge: 4 cores
‣ AWS example
Spark
SPARK EXECUTOR
c3.large: 2 cores
SPARK DRIVER
SPARK EXECUTOR
CORE 1 CORE 2 CORE 3 CORE 4
c3.xlarge: 4 cores
‣ AWS example
Spark on Docker
c3.large: 2 cores
SPARK DRIVER
CORE 1 CORE 2 CORE 3 CORE 4
docker-1: 4 “cores”
SPARK EXECUTOR
C1 C2 C4C3
docker-2: 4 “cores”
SPARK EXECUTOR
C1 C2 C4C3
SPARK EXECUTOR
‣ Spark Streaming is fast to deploy but tuning is VERY important
‣ The lower the number of tasks, the better (in general)
‣ When reading from Kafka make sure that you configure blockingInterval
‣ optimize your jobs when possible - similar jobs can be sometimes merged
‣ persist your data from workers, NOT the driver
Spark Streaming
30
‣ OLAP-type queries using Spark SQL
‣ More advanced performance testing
‣ Detailed unit testing
‣ More batch jobs
Where do we go from here?
31
‣ Spark Documentation
‣ Reference application: http://github.com/killrweather/killrweather
‣ Productionalizing Spark Streaming
‣ Spark and Kafka
‣ Docker
‣ Free Hadoop Training from MapR
‣ Free edX course on Spark
Resources
32
Related Documents
