http://flink-forward.org/kb_sessions/flink-and-beam-current-state-roadmap/ It is no secret that the Dataflow model, which evolved from Google’s MapReduce, Flume, and MillWheel, has been a major influence to Apache Flink’s streaming API. The essentials of this model are captured in Apache Beam. Beam provides the Dataflow API with the option to deploy to various backends (e.g. Flink, Spark). In this talk we will examine the current state of the Flink Runner. Beam’s Runners manage the translation of the Beam API into the backend API. The Beam project itself has made an effort to summarize the capabilities of each Runner to provide an overview of the supported API concepts. From all open sources backends, Flink is currently the Runner which supports the most features. We will look at the supported Beam features and their counterpart in Flink. Further, we will look at potential improvements and upcoming features of the Flink Runner.

1. Flink and Beam:
Current State & Roadmap
Maximilian Michels
PMC Apache Flink and Apache Beam
FlinkForward 2016
Apache Flink® Apache Beam
(incubating)

2. 2
Google Cloud Dataflow On Top of Apache Flink®
Last year’s talk:
This year’s talk:
Apache Flink® and Apache Beam: Current State &
Roadmap
What happened?

3. Two projects, one idea
• Dec 2014 Google releases the Cloud Dataflow Java SDK
• Feb 2016 Cloud Dataflow Java SDK becomes Apache Beam
3
MapReduce
BigTable DremelColossus
FlumeMegastoreSpanner
PubSub
Millwheel
Apache
Beam
Google Cloud
Dataflow
Taken from the official Beam material

4. Apache Beam
(incubating)
• Jan 2016 Google proposes project to the Apache incubator
• Feb 2016 Project enters incubation
• Jun 2016 Apache Beam 0.1.0-incubating released
• Jul 2016 Apache Beam 0.2.0-incubating released
4
Dataflow Java 1.x
Apache Beam Java 0.x
Apache Beam Java 2.x
Bug Fix
Feature
Breaking Change

5. The Dataflow/Beam Model
• 2004 MapReduce
• 2010 Flume Java (Beam’s API)
• 2013 MillWheel (Watermarks, exactly-once,
Windows)
• Sep 2015 Dataflow Model published at VLDB
• Nov 2015 Apache Flink 0.10.0 DataStream API
features Event Time in line with the Dataflow model
5

6. Beam Runners
• A Runner ports Beam code to a backend
• A Runner’s concern is
• a) translation
• b) runtime execution
• Runners available:
• Google Cloud Dataflow
• Apache Flink
• Apache Spark
• In Development: Gearpump, Apache Apex
6
Beam Model: Fn Runners
Apache
Flink
Apache
Spark
Beam Model: Pipeline Construction
Other
LanguagesBeam Java
Beam
Python
Execution Execution
Cloud
Dataflow
Execution
Taken from the official Beam material

7. Flink Runner - Status Quo
• Stable: 0.2.0-incubating
• Powered by Flink 1.0.3
• Development: 0.3.0-incubating-SNAPSHOT
• Powered by Flink 1.1.2
• Changes coming in regularly, use the snapshot
version to get the latest features
7

8. Evolution of the Flink Runner
2014
Dec
• Google Cloud Dataflow SDK released
2015
Jan
• Initial commit and drafting
Mar
• Batch support without Windows
Dec
• Streaming support with Watermarks
and Windows
8
2016
Mar
• Contribution to Apache Beam
May
• Parallel and checkpointed sources in
streaming
• Batch support with Windows and Side
Inputs
• RunnableOnService for batch
Aug
• Side Inputs for streaming
• RunnableOnService for streaming

9. Batched vs Streamed
Execution
• Batched Execution
• Built on top of the DataSet API
• Supports only bounded sources
• Managed Memory
• Streamed Execution
• Built on top of the DataStream API
• Supports bounded and unbounded sources
• Watermark support which leverages full Dataflow Model
• Checkpointing
• When to pick which? If undecided, just let the Runner decide
9

10. The Flink Runner’s Strength
• Backed by a runtime which is streaming and batch aware
• Embeds and reuses Beam logic whenever possible
• Window and Triggering
• Source API
• State Internals / Checkpointing
• Whenever possible, translate primitive transforms
• Integration with Beam’s RunnableOnService tests (batch/
streaming)
10

11. Capability Matrix
• Compares Runners with the reference model
• What results are being calculated?
• Where in event time?
• When in processing time?
• How do refinements of results relate?
11
For more information see
http://beam.incubator.apache.org/learn/runners/capability-matrix/

12. What
12

13. Where
13

14. When
14

15. How
15

16. So we’re good?
• Flink Runner currently the most advanced Runner
which is backed by an open-source engine
• Does that mean the Flink Runner is perfect?
16

17. Roadmap Flink Runner
• Side Input streaming: size restrictions
• Integrate with new PipelineResult
• Dynamic Scaling
• Incremental Checkpointing
• Connectors (Kafka 8/9, Cassandra, Elasticsearch, RabbitMQ, Redis, NiFi,
Kinesis)
• Libraries (Gelly, CEP, Storm Compatibility, ML)
• Performance testing
• Bug fixes
17

18. Apache Flink Features
powering the Runner
• Streaming
• Checkpointing / Savepoints
• State backends
• Batch
• Pipelined execution
• Managed Memory
• High Availability
• Flink’s native sources / sinks
18

19. Demo
• Develop a Beam program with the Flink Runner
• Monitor applications using the web interface
• Handle failures by restoring from checkpointed
state
• Create a Savepoint, stop the Beam program,
resume execution from it
19

20. Thank you for your
attention!