RubiX: A caching framework for big data engines in the cloud. Helps provide data caching capabilities to engines like Presto, Spark, Hadoop, etc transparently without user intervention.

1. RubiX
A caching framework for big data engines in the
cloud
Strata + Hadoop World
March 2017
Shubham Tagra (stagra@qubole.com)

2. Agenda
● Intro
● Why Caching?
● Path to Rubix
● Rubix Architecture
● Future of Rubix
● QnA

3. Built for Anyone who Uses Data
Analysts l Data Scientists l Data Engineers l Data Admins
Optimize performance,
cost, and scale through
automation, control and
orchestration of big data
workloads.
A Single Platform for Any Use Case
ETL & Reporting l Ad Hoc Queries l Machine Learning l
Streaming l Vertical Apps
Open Source Engines, Optimized for the Cloud
Native Integration with multiple cloud providers

4. Qubole operates at Cloud Scale
500 PB
Data Processed in the
Cloud Monthly
6
PB
80
PB
150
PB
500
PB
500 Nodes
Largest Spark Cluster in
the Cloud
2000
Clusters Started per
month

5. Why Caching
● Popularity of Cloud Stores like S3
+ Near-infinite capacity
+ Inexpensive
+ Ease of use
- Network Latencies
- Back-offs

6. Rubix ancestors
● File cache

7. Rubix ancestors
● File cache
○ Benefits: as much as 10x performance improvement
○ Problems
■ Huge warm-ups
■ Cache size
■ Tied to Presto
■ Required Presto scheduler changes

8. ● Improve performance
● Abstracted from user
○ Easy of use
● Support Columnar formats
○ Improves speed
● Work well with autoscaling
○ Saves cost
● Ease of extension to clouds and engines
Requirements for new cache

9. Alternatives Considered: FUSE FileSystem
● Mount S3 paths on ec2
● OS for page caching, read ahead, etc
● Problems
○ Exclusive control over bucket
○ Data corruptions in external updates
○ Not production ready

10. Alternatives Considered: HTTP Caching

11. Alternatives Considered: HTTP Caching
● Worked fine with TXT data
● Problems
○ Columnar formats and Byte-Range based Varnish Keys
■ Poor hit ratio
■ Redundant copies

12. Tachyon/Alluxio
● More than just a caching system
● We required light weight system
● SQL first

13. Rubix
● Extendible to many engines
● Columnar format friendly
● Works well with autoscaling
● Share-able across engines/instances

14. Architecture
● Split ownership assignment system
● Data Caching System
● Plugins

15. Architecture
● Split ownership assignment
system
○ Used in master node during split computation
○ Calculates which node owns particular split of
file
○ Uses Consistent Hashing to work well with
Autoscaling

16. ● Data Caching System
○ Used in worker nodes when data is read
○ Read from disk or remote as per the
metadata
○ Metadata stored in units of block (1MB each)
○ BookKeeper provides metadata for the block
○ Metadata too Checkpointed to local disk
Architecture

17. ● Plugin
○ Provides two types of information
■ How to get the list of nodes in the
system
■ FileSystem for remote reads
○ E.g. presto plugin, hadoop1 plugin, hadoop2
plugin
Architecture

18. Plugins: Presto
● Presto provides tight control over scheduling local splits
● This ensured that splits will be always scheduled locally
● Worked well for our customers

19. Plugins: Hadoop
● Strict local scheduling was not possible with hadoop
● This meant lot of warm-ups and redundant copies of data
● Options:
○ Read directly from remote for non-local read
○ Figure out the correct owner and read from it
○ Implement Non-Local reads for Hadoop support
○ Learnings
■ 100% strict location based scheduling not possible in H2

20. Using Rubix with Presto
● Configure disk mount point
○ Assumes disks mounted on /media/ephemeral0, /media/ephemeral1, etc by default
● Start BookKeeper
● Place rubix jars in hive-hadoop2 plugin of Presto
● Configure Presto to use Rubix FileSystem for the cloud store

21. Using Rubix with Presto in Qubole

22. Using Rubix with Hadoop
● Configure disk mount point
○ Assumes disks mounted on /media/ephemeral0, /media/ephemeral1, etc by default
● Start BookKeeper
● Place rubix jars with hadoop libraries
● Configure Hadoop to use Rubix FileSystem for the cloud store

23. Extending to other Engines and Clouds

24. Performance gains

25. Future Work
● Extend to other clouds and engines
● Table aware objects in Rubix
● Caching policies for Hive Partitions
● Subquery caching

26. Questions?