Spark RDDs are almost identical to Scala collection, just in a distributed manner, all of the transformations and actions are derived from the Scala collections API.
As Martin Odersky mentioned, “Spark - The Ultimate Scala Collections” is the right way to look at RDDs. But with that great distributed power comes a great many data problems: at first you’ll start tackling the concept of partitioning, then the actual data becomes the next thing to worry about.
In the talk we’ll go through an overview on Spark's architecture, and see how similar RDDs are to the Scala collections API. We'll then shift to the world of problems that you’ll be facing when using Spark for processing a vast volume of time-series data with multiple data stores (S3, MongoDB, Apache Cassandra, MySQL).
When you start tackling many scale and performance problems, many questions arise:
> How to handle missing data?
> Should the system handle both serving and backend processes, or should we separate them out?
> Which solution is cheaper?
> How do we get the best performance for money spent?
In the talk we will tell the tale of all of the transformations we’ve made to our data and review the multiple data persistency layers... and I’ll try my best NOT to answer the question “which persistency layer is the best?” but I do promise to share our pains and lessons learned!
2. About Me
Demi Ben-Ari, Co-Founder & CTO @ Panorays
● BS’c Computer Science – Academic College Tel-Aviv Yaffo
● Co-Founder “Big Things” Big Data Community
In the Past:
● Sr. Data Engineer - Windward
● Team Leader & Sr. Java Software Engineer,
Missile defense and Alert System - “Ofek” – IAF
Interested in almost every kind of technology – A True Geek
3. Agenda
● Scala and Spark analogies
● Data flow and Environment
● What’s our time series data like?
● Where we started from - where we got to
○ Problems and our decisions
● Conclusions
8. What kind of DSL is Apache Spark
● Centered around Collections
● Immutable data sets equipped with functional transformations
● These are exactly the Scala collection operations
map
flatMap
filter
...
reduce
fold
aggregate
...
union
intersection
...
9. Spark vs. Scala Collections
● So, Spark is exactly Scala Collections, but running in a Cluster?
● Not quite. There are Two main differences:
○ Spark is Lazy, Scala collections are strict
○ Spark has added functionality, eg. PairRDDs.
■ Gives us the power doing lots of operations in the NoSQL distributed
world
17. Structure of the Data
● Geo Locations + Metadata
● Arriving over time
● Different types of messages being reported by sattelites
● Encoded
● Might arrive later than acttually transmitted
18. Data Flow Diagram
Externa
l Data
Source
Analytics
Layers
Data Pipeline
Parsed
Raw
Entity Resolution
Process
Building insights
on top of the entities
Data
Output
Layer
Anomaly
Detection
Trends
20. Basic Terms
● Idempotence
is the property of certain operations in mathematics and computer
science, that can be applied multiple times without changing the
result beyond the initial application.
● Function: Same input => Same output
21. Basic Terms
● Missing Parts in Time Series Data
◦ Data arriving from the satellites
⚫ Might be causing delays because of bad transmission
◦ Data vendors delaying the data stream
◦ Calculation in Layers may cause Holes in the Data
● Calculating the Data layers by time slices
22. Basic Terms
● Partitions == Parallelizm
◦ Physical / Logical partitioning
● Resilient Distributed Datasets (RDDs) == Collections
◦ fault-tolerant collection of elements that can be operated on in
parallel.
◦ Applying immutable transformations and actions over RDDs
24. The Problem - Receiving
DATA
Beginning state, no data, and the timeline
begins
T = 0
Level 3
Entity
Level 2
Entity
Level 1
Entity
25. The Problem - Receiving
DATA
T = 10
Level 3
Entity
Level 2
Entity
Level 1
Entity
Computation sliding window size
Level 1 entities data
arrives and gets stored
26. The Problem - Receiving
DATA
T = 10
Level 3
Entity
Level 2
Entity
Level 1
Entity
Computation sliding window size
Level 3 entities are created
on top of Level 2’s Data
(Decreased amount of data)
Level 2 entities are
created on top of Level 1’s
Data
(Decreased amount of
data)
27. The Problem - Receiving
DATA
T = 20
Level 3
Entity
Level 2
Entity
Level 1
Entity
Computation sliding window size
Because of the sliding window’s
back size, level 2 and 3 entities
would not be created properly
and there would be “Holes” in the
Data
Level 1 entity's
data arriving late
28. Solution to the Problem
● Creating Dependent Micro services forming a data pipeline
◦ Mainly Apache Spark applications
◦ Services are only dependent on the Data - not the previous
service’s run
● Forming a structure and scheduling of “Back Sliding Window”
◦ Know your data and it’s relevance through time
◦ Don’t try to foresee the future – it might Bias the results
30. How we started?
● Spark Standalone – via ec2 scripts
◦ Around 5 nodes (r3.xlarge instances)
◦ Didn’t want to keep a persistent HDFS – Costs a lot
◦ 100 GB (per day) => ~150 TB for 4 years
◦ Cost for server per year (r3.xlarge):
● On demand: ~2900$
● Reserved: ~1750$
● Know your costs: http://www.ec2instances.info/
31. Decision
● Working with S3 as the persistence layer
◦ Pay extra for
● Put (0.005 per 1000 requests)
● Get (0.004 per 10,000 requests)
◦ 150TB => ~210$ for 4 years of Data
● Same format as HDFS (CSV files)
◦ s3n://some-bucket/entity1/201412010000/part-00000
◦ s3n://some-bucket/entity1/201412010000/part-00001
◦ ……
36. The Problem
● Batch jobs
◦ Should run for 5-10 minutes in total
◦ Actual - runs for ~40 minutes
● Why?
◦ ~20 minutes to write with the Java mongo driver – Async
(Unacknowledged)
◦ ~20 minutes to sync the journal
◦ Total: ~ 40 Minutes of the DB being unavailable
◦ No batch process response and no UI serving
37. Alternative Solutions
● Sharded MongoDB (With replica sets)
◦ Pros:
● Increases Throughput by the amount of shards
● Increases the availability of the DB
◦ Cons:
● Very hard to manage DevOps wise (for a small team of
developers)
● High cost of servers – because each shared need 3 replicas
39. Our DevOps – After that solution
We had no
DevOps guy at
that time at all
☹
40. Alternative Solutions
● Apache Cassandra
◦ Pros:
● Very large developer community
● Linearly scalable Database
● No single master architecture
● Proven working with distributed engines like Apache Spark
◦ Cons:
● We had no experience at all with the Database
● No Geo Spatial Index – Needed to implement by ourselves
41. The Solution
● Migration to Apache Cassandra
● Create easily a Cassandra cluster using DataStax Community AMI
on AWS
◦ First easy step – Using the spark-cassandra-connector
(Easy bootstrap move to Spark ⬄ Cassandra)
◦ Creating a monitoring dashboard to Cassandra
43. Result
● Performance improvement
◦ Batch write parts of the job run in 3 minutes instead of ~ 40
minutes in MongoDB
● Took 2 weeks to go from “Zero to Hero”, and to ramp up a running
solution that work without glitches
45. Transferring the Heaviest Process
● Micro service that runs every 10 minutes
● Writes to Cassandra 30GB per iteration
◦ (Replication factor 3 => 90GB)
● At first took us 18 minutes to do all of the writes
◦ Not Acceptable in a 10 minute process
47. Transferring the Heaviest Process
● Solutions
◦ We chose the i2.xlarge
◦ Optimization of the Cluster
◦ Changing the JDK to Java-8
● Changing the GC algorithm to G1
◦ Tuning the Operation system
● Ulimit, removing the swap
◦ Write time went down to ~5 minutes (For 30GB RF=3)
Sounds good right? I don’t think so
49. The Solution
● Taking the same Data Model that we held in Cassandra (All of the
Raw data per 10 minutes) and put it on S3
◦ Write time went down from ~5 minutes to 1.5 minutes
● Added another process, not dependent on the main one, happens
every 15 minutes
◦ Reads from S3, downscales the amount and Writes them to
Cassandra for serving
50. How it looks after all?
Parsed
Raw
Static /
Aggregated
Data
Spark Analytics Layers
UI Serving
Downscale
d Data
Heavy
Fusion
Process
51. Conclusion
● Always give an estimate to your data
◦ Frequency
◦ Volume
◦ Arrangement of the previous phase
● There is no “Best” persistence layer
◦ There is the right one for the job
◦ Don’t overload an existing solution
52. Conclusion
● Spark is a great framework for distributed collections
◦ Fully functional API
◦ Can perform imperative actions
● “With great power,
comes lots of partitioning”
◦ Control your work and
data distribution via partitions
● https://www.pinterest.com/pin/155514993354583499/ (Thanks)