This document discusses optimizing data ingestion into Hadoop for high volume event streaming. It describes how Hadoop was not designed for small, high volume event data and outlines several techniques to improve ingest performance: buffering events to reduce mapper wakeups, compressing events into larger blocks for network transfer, and analyzing event data during ingestion to store metadata that improves later data access and analysis. The key is developing "mechanical sympathy" and balancing resource usage to fully utilize available hardware and remove bottlenecks.

1. © Hortonworks Inc. 2012
High Speed Smart Data Ingest
into Hadoop
Oleg Zhurakousky
Principal Architect @ Hortonworks
Twitter: z_oleg
Tom McCuch
Solution Engineer @ Hortonworks
Twitter: tmccuch

2. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Agenda
• Domain Problem
–Event Sourcing
–Streaming Sources
–Why Hadoop?
• Architecture
–Compute Mechanics
–Why things are slow – understand the problem
–Mechanics of the ingest
– Write >= Read | Else ;(
–Achieving Mechanical Sympathy –
– Suitcase pattern – optimize Storage, I/O and CPU

3. © Hortonworks Inc. 2012
Domain Problem

4. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Domain Events
• Domain Events represent the state of entities at a
given time when an important event occurred
• Every state change of interest to the business is
materialized in a Domain Event
Page 4
Retail & Web
Point of Sale
Web Logs
Telco
Network Data
Device Interactions
Financial
Services
Trades
Market Data
Energy
Metering
Diagnostics
Healthcare
Clinical Trials
Health Records

5. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Event Sourcing
• All Events are sent to an Event Processor
• The Event Processor stores all events in an Event Log
• Many different Event Listeners can be added to the
Event Processor (or listen directly on the Event Log)
Event
Processor
Web
Messaging
Search
Batch
Analytics
Page 5
Apps

6. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Streaming Sources
• When dealing with Event data (log messages, call
detail records, other event / message type data) there
are some important observations:
–Source data is streamed (not a static copy), thus always in motion
–More then one streaming source
–Streaming (by definition) never stops
–Events are rather small in size (up to several hundred bytes), but
there is a lot of them
Hadoop was not design with that in mind
E E E
Geos
Telco Use Case
– Tailing the syslog of a network device
– 200K events per second per streaming source
– Bursts that can double or triple event production

7. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Why Hadoop for Event Sourcing?
• Event data is immutable, an event log is append-only
• Event Sourcing is not just event storage, it is event
processing (HDFS + MapReduce)
• While event data is generally small in size, event
volumes are HUGE across the enterprise
• For an Event Sourcing Solution to deliver its benefits
to the enterprise, it must be able to retain all history
• An Event Sourcing solution is ecosystem dependent –
it must be easily integrated into the Enterprise
technology portfolio (management, data, reporting)
Page 7

8. © Hortonworks Inc. 2012
Ingest Mechanics

9. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Mechanics of Data Ingest & Why and which
things are slow and/or inefficient ?
• Demo
Page 9

10. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Mechanics of a compute
Atos, Portos, Aramis and the D’Artanian – of the compute

11. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Why things are slow?
• I/O | Network bound
• Inefficient use of storage
• CPU is not helping
• Memory is not helping
Imbalanced resource usage leads to limited growth potentials
and misleading assumptions about the capability of yoru
hardware
Page 11

12. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Mechanics of the ingest
• Multiple sources of data (readers)
• Write (Ingest) >= Read (Source) | Else
• “Else” is never good
–Slows the everything down or
–Fills accumulation buffer to the capacity which slows everything
down
–Renders parts or whole system unavailable which slows or brings
everything down
–Etc…
“Else” is just not good!
Page 12

13. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Strive for Mechanical Sympathy
• Mechanical Sympathy – ability to write software with
deep understanding of its impact or lack of impact on
the hardware its running on
• http://mechanical-sympathy.blogspot.com/
• Ensure that all available resources (hardware and
software) work in balance to help achieve the end
goal.
Page 13

14. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Achieving mechanical sympathy during the
ingest?
• Demo
Page 14

15. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
The Suitcase Pattern
• Storing event data in uncompressed form not feasible
–Event Data just keeps on coming, and storage is finite
–The more history we can keep in Hadoop the better our trend
analytics will be
–Not just a Storage problem, but also Processing
– I/O is #1 impact on both Ingest & MapReduce performance
–Consider event data “traveling” over the network to Hadoop
– we want to compress it before it makes the trip, but in a way that
facilitates how we intend to process it once it arrives
• Suitcase Pattern
–Before we travel, we take our clothes off the
rack and pack them (easier to store)
–We then unpack them when we arrive and put
them back on the rack (easier to process)

16. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
What’s next?
• How much data is available or could be available to us
during the ingest is lost after the ingest?
• How valuable is the data available to us during the
ingest after the ingest completed?
• How expensive would it be to retain (not lose) the data
available during the ingest?
Page 16

17. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Data available during the ingest?
• Record count
• Highest/Lowest record length
• Average record length
• Compression ratio
But with a little more work. . .
• Column parsing
–Unique values
–Unique values per column
–Access to values of each column independently from the record
–Relatively fast column-based searches, without indexing
–Value encoding
–Etc…
Imagine the implications on later data access. . .
Page 17

18. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
How expensive is to do that extra work?
• Demo
Page 18

19. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Conclusion - 1
• Hadoop wasn’t optimized to process <1 kB records
–Mappers are event-driven consumers: they do not have a running
thread until a callback thread delivers a message
–Message delivery is an event that triggers the Mapper into action
• “Don’t wake me up unless it is important”
–In Hadoop, generally
speaking, several thousand bytes to
several hundred thousand bytes
is deemed important
–Buffering records during collection allows us to collect more data
before waking up the elephant
–Buffering records during collection also allows us to compress the
whole block of records as a single record to be sent over the
network to Hadoop – resulting in lower network and file I/O
–Buffering records during collection also helps us handle bursts

20. © Hortonworks Inc. 2012© Hortonworks Inc. 2012
Conclusion - 2
• Understand your SLAs, but think ahead
–While you may have accomplished your SLA for the ingest your
overall system may still be in the unbalanced state
–Don’t settle on “throw more hardware” until you can validate that
you have approached the limits of your existing hardware
–Strive for mechanical sympathy. Profile your software to
understand its impact on the hardware its running on.
–Don’t think about the ingest in isolation. Think how the ingested
data is going to be used.
–Analyze, the information that is available to you during the ingest
and devise a convenient mechanism for storing it. Your data
analysts will thank you for that.
Page 20

21. © Hortonworks Inc. 2012
Thank You!
Questions & Answers
Follow: @z_oleg, @tmccuch, @hortonworks
Page 21