Advanced Visualization of Spark jobs

1. Advanced visualization of
Spark jobs
Zoltán Zvara
zoltan.zvara@sztaki.hu
Márton Balassi
marton.balassi@sztaki.hu

2. Agenda
• Introduction
• Motivation and the challenges
• The Spark execution visualization
• An elegant way to tackle data skew
• The visualizer extended to Flink
• Future plans

3. Introduction
• Hungarian Academy of Sciences (MTA SZTAKI)
• Research institute with strong industry ties
• Big Data projects using Spark, Flink, Cassandra, Hadoop etc.
• Multiple telco use cases lately, with challenging data volume and
distribution

4. Motivation
• We have developped an application aggreagting telco data that tested
well on toy data
• When deploying it against the real dataset the application seemed
healthy
• However it could become suprisingly slow or even crash
• What did go wrong?

5. Data skew
• When some data-points are very frequent, and we need to process
them on a single machine – aka the Bieber-effect
• Power-law distributions are very common: from social-media to
telecommunication, even in our beloved WordCount
• Default hashing puts these into „random” buckets
• We can do better than that

6. Aim
1. Most of the telco data-processing workloads suffer from inefficient
communication patterns, introduced by skewed data
2. Help developers to write better applications by detecting issues of
logical and physical execution plans easier
3. Help newcomers to understand a distributed data-processing
system faster
4. Demonstrate and guide the testing of adaptive (re)partitioning
techniques

7. Spark’s execution model
• Extended MapReduce model, where a previous stage has to complete
all its tasks before the next stage can be scheduled (global
synchronization)
• Data skew can introduce slow tasks
• In most cases, the data distribution
is not known in advance
• „Concept drifts” are common
stage
boundary
& shuffle
even partitioning skewed data
slow task
slow task

8. Physical execution plan
• Understanding the physical plan
• can lead to designing a better application
• can uncover bottlenecks that are hard to identify otherwise
• can help to pick the right tool for the job at hand
• can be quite difficult for newcomers
• can give insights to new, adaptive, on-the-fly partitioning & optimization
strategies

9. Current visualizations in Spark
• Current visualization tools
• are missing the fine-grained input & output metrics – also not available in
most sytems, like Spark
• does not capture the data charateristics – hard to accomplish a lightweight
and efficient solution
• does not visualize the physical plan
DAG visualization in Spark

10. Collecting information during execution
• Data-transfer between tasks are accomplished in
the shuffle phase, in the form of shuffle block fetches
• We distinguish local & remote block fetches
• Data characteristics collected on shuffle write & read

11. A scalable sampling
• Key-distributions are approximated with a strategy, that
• is not sensitive to early or late concept drifts,
• lightweight and efficient,
• scalable by using a backoff strategy
keys
frequency
counters
keys
frequencysampling rate of 𝑠𝑖 sampling rate of 𝑠𝑖/𝑏
keys
frequency
sampling rate of 𝑠𝑖+𝑗/𝑏
truncateincrease by 𝑠𝑖
𝑇𝐵

12. Availability through the Spark REST API
• Enhanced Spark REST API to provide block fetch & data characteristics
information of tasks
• New queries to the REST API, for example: „what happend in the last
3 seconds?”
• /api/v1/applications/{appId}/{jobId}/tasks/{timestamp}

13. THE VISUALIZATION

14. Our way of tackling data skew
• Goal: use the collected information to handle data skew dynamically,
on-the-fly on any workload (batch or streaming)
• Currently paper in progress
heavy keys create
heavy partitions (slow tasks)
Bieber-key is alone,
size of the biggest partition
is minimized

15. Making visualization pluggable
• Let us also support Flink
• The execution model is quite different from Spark
• Minimal code change on the data reader module of the visualization
• Expose the necessary metrics through the Flink JM Rest API
• Map Flink’s model to Spark’s

16. Computational models
Batch data
Kafka, RabbitMQ ...
HDFS, JDBC ...
Stream Data
Spark RDDs break down the computation into stagesFlink computation is fully pipelined by default

17. The currently available Flink visualizer

18. Granular visualization of Flink jobs

19. Future plans
• Open-sourcing of the visualization framework
• Dynamic repartitioning paper is in progress
• Task-classification to make the visualization more compact
• Visual improvements, more features
• Opening PRs against Spark & Flink with the suggested metrics
• A rewrite is coming for the Flink metrics, track it at FLINK-1504

20. Conclusions
• The devil is in the details
• Visualizations can aid developers to better understand issues and
bottlenecks of certain workloads
• Specially data skew can hinder the completion time of the whole
stage in a batch engine
• Adaptive repartitioning strategies can be suprisingly efficient

21. Thank you for your attention
Q&A
stage
boundary
& shuffle
even partitioning skewed data
slow task
slow task