2. About me
● Technical Leader of Data Platform in LivePerson
● Bird watcher and amateur bird photographer
Pharaoh Eagle-Owl / Bubo ascalaphus
This is what the people from previous slide were looking at…
Amir Silberman
3. Agenda
● Why we chose Kafka + Storm
● How implementation was done
● Measures of success
● Two examples of use
● Tips from our experience
4. Data in LivePerson
Visitor in Site
Chat Window
Agent console
LivePerson SaaS Server
LoginMonitor
Rules,
Intelligence,
Decision
Chat
Chat
Invite
DATA
DATA DATA
BIG
DATA
5. Legacy Data flow in LivePerson
BI DWH
(Oracle)
RealTime
servers
ETL
Sessionize
Modeling
Schema
View
Real-Time data
Historical data
6. Why Kafka + Storm?
● Need to scale out and plan for future scale
○ Limit for scale should not be technology
○ Let the limit be cost of (commodity) hardware
● What Data platforms can be implemented quickly?
○ Open source - fast evolving and community
○ Micro-services - do only what you ought to do!
● Are there risks in this choice?
○ Yes! technology is not mature enough
○ But, there is no other mature technology that can
address our needs!
8. Legacy Data flow in LivePerson
BI DWH
(Oracle)
RealTime
servers
Customers
ETL
Sessionize
Modeling
Schema
View
9. 1st phase - move to Hadoop
ETL
Sessionize
Modeling
Schema
View
RealTime
servers
BI DWH
(Vertica)HDFS
Hadoop
MR Job transfers
data to BI DWH
Customers
10. 2. move to Kafka
6
RealTime
servers
HDFS
BI DWH
(Vertica)
Hadoop
MR Job transfers
data to BI DWH
Kafka
Topic-1
Customers
11. 3. Integrate with new producers
6
RealTime
servers
HDFS
BI DWH
(Vertica)
Hadoop
MR Job transfers
data to BI DWH
Kafka
Topic-1 Topic-2
New
RealTime
servers
Customers
12. 4. Add Real-time BI
6
Customers
RealTime
servers
HDFS
BI DWH
(Vertica)
Hadoop
MR Job transfers
data to BI DWH
Kafka
Topic-1 Topic-2
New
RealTime
servers
Storm
Topology
Analytics
DB
16. 1st Strom Use Case: “Visitors List”
Use case:
● Show list of visitors in the “Agent Console”
● Collect data about visitor in real time
● Visitor stickiness in streaming process
18. Selected Analytics DB - Couchbase
1st Strom Use Case: “Visitors List”
● Document Store - for complex documents
● Searchable - possible to search by different
attributes.
● High throughput - Read & Write
20. Visitors List - Storm considerations
● Complex calculations before sending to DB
○ Ignore delayed events
○ Reorder events before storing
● Document cached in memory
● Fields Grouping to bolt that writes to CouchBase
● High parallelism in bolt that writes to CouchBase
25. Selected Analytics DB - Cassandra
2nd Storm Use Case: “Agent State”
● Wide Column Store DB
● Highly Available w/o Single point of failure
● High throughput
● Optimized for counters
26. First Storm Topology – Visitor Feed
Storm Topology
Kafka Spout Analyze relevant
events
Send events
emit emit
Kafka events stream
Add
“Agent Status” Topology:
Analytics DB: Cassandra - Document store
Parse Avro into
tuple
emit
Data
visualization
using Highcharts
27. Agent Status - Storm considerations
● Counters stored by topology
● Calculations done after reading from DB
● Delayed events should not be ignored
● Order of events does not matter
● Using Highcharts for data visualization
29. 3rd Storm Use Case: Data Auditing
Use case:
● Needs to be able to tell whether events arrived
○ Where there any missing events?
○ Where there any duplicated events?
○ How long did it take for events to arrive?
● Data not important - only count of events
30. 3rd Storm Use Case: Data Auditing
Realtime server
Kafka
Topics
Auditing
Topic
Storm Sync
topology
Audit-loader
topology
MySql
Hadoop
HDFS
audit
job
kafka
1
3
4
2
Auditor
31. First Storm Topology – Visitor Feed
Storm Topology
Kafka Spout Analyze relevant
events
Send events
emit emit
Kafka events stream
Add
“Sync Audit” Topology:
Sync messages between two topics
Parse Avro into
tuple
emit
Kafka Audit topic
33. “Load Audit” Topology:
● Stores statistics of events count
● SQL type DB
● Used for Auditing and other statistics
● Requires metadata in events header
34. Challenges:
● High network traffic
● Writing to Kafka is faster than reading
● All topologies read all events
● How to avoid resource starvation in Storm
Subalpine Warbler / Sylvia cantillans
Amir Silberman
35. Optimizations of Kafka
● Increase Kafka consuming rate by adding partitions
● Run on physical machines with RAID
● Set retention to the proper need
● Monitor data flow!
36. Optimizations of Storm
● #of Kafka-Spouts = number of total partitions
● Set “Isolation mode” for important topologies
● Validate Network cards can carry network traffic
● Set Storm cluster on high CPU machines
● Monitor servers CPU & Memory (Graphite)
● Assess min. #Cores that topology needs
○ Use “top” -> “load” to find server load
37. Demo
● Agent Console - https://z1.le.liveperson.net/
71394613 / rans@liveperson.com
● My Site - http://birds-of-israel.weebly.com/