Skip to main content

Mutable Ideas

Spark Summit 2014 - Day 2 (Afternoon)

Following the Spark Summit 2014 - Day 2

# BI-style analytics on Spark (without Shark) using SparkSQL & SchemaRDD

Justin Langseth, Farzad Aref (Zoomdata)


  • Moving from Storm to Spark Stream

## Why they are using Spark?

  • flexible

  • distributed and fast!

  • rich math library (MLlib, graphX, Bagel)

  • Holding small DS

  • Holding aggregation datasets

  • data fusion across disparate sources

  • complex math

## Challenges

  • Sharing Spark contexts
  • Sharing RDDs across contexts
  • Not sure about Tachyon

# A Deeper Understanding of Spark Internals

Aaron Davidson (Databricks)

## Major core components for performance

  • exec model
  • shuffle
  • caching

## Create exec plan

Pipeline as much as possible Split into stages, on need to reorganize data.

  • Single KV must fit in memory!

## Common issues checklist


## Tuning the number of partitions


## Memory problems


# Spark on YARN: a Deep Dive

Sandy Ryza (Cloudera)

YARN: Execution/Scheduling (decides who/what/WHERE gets to run) ![img] (/images/yarn.jpg)

## Why to run on YARN?

  • manage workloads (allocate shares)
  • security (kerberos cluster)

YARN Spark 1.0 + CDH 5.1: Easier app submission spark-submit. Stable since CDH 5.0

## Yarn Client


## Yarn Cluster


## Problem with data locality

When running Spark on Yarn, solution is to include on the SparkContext definition where of files location, so yarn can select better containers.

# Productionizing a 24/7 Spark Streaming service on YARN

Issac Buenrostro, Arup Malakar (Ooyala)


# Going Live – Things to Address Before Your First Live Deployment

Gary Malouf (MediaCrossing Inc.)

  • Spark Standalone would be better if only Spark were running.
  • Using MESOS, Chronos for job scheduling
  • Cassandra (Long Term Data)

“If you’re starting on 2014, try to go with Spark”

HDFS for small data -> KV data prefer uses Cassandra (rollups, reports, etc)

# A Web application for interactive data analysis with Spark

Romain Rigaux (Cloudera)

Submitting spark jobs directly

Hue –> Spark Job Server –> Spark

  • Leverage on Spark Job Server Convert your Job to Spark Job Server, using trait SparkJob

Get started with Spark: deploy Spark Server and compute Pi from your Web Browser

## That’s all!