hadoop at ayasdi

18
Hadoop at Ayasdi Mohit Jaggi and Huang Xia, Zhen Li, Ajith Warrier

Upload: mohit-jaggi

Post on 19-Aug-2015

190 views

Category:

Technology


0 download

TRANSCRIPT

Hadoop at AyasdiMohit Jaggi

andHuang Xia, Zhen Li, Ajith Warrier

Overview

- HDFS for storage- YARN for integration into Hadoop data lake- Parquet as the file format- bigdf based on Spark for feature

engineering, data wrangling

Hadoop

Data

Lake

bigdf

Resource Broker

Apache Spark

YARN

Algorithms

UI/API handler

Architecture

!! Audience Poll !!

1. How many data scientists?2. How many backend engineers?3. UI/frontend engineers?4. Using hadoop in production?5. Using Spark in production?6. Personally worked on data bigger than 100GB? 1TB? 10TB?

1PB?

HDFS

HDFS - Motivation

- installed base, large community- ecosystem to connect to most other data

sources- commodity cluster- experiments with distributed NAS didn't

show enough promise to justify the additional cost and complexity

HDFS - Usage

- used as distributed storage- jobs dispatched to least loaded node- run Spark jobs

FishFiFishing for insights...

big data

YARN - Motivation

- Ayasdi scheduler- maximize throughput for batch jobs- minimize latency for interactive “tasklets”

- wanted to deploy in existing Hadoop data lakes

- integrated inhouse scheduler with YARN- “tasklets” get a long running container- batch jobs get a container on demand

YARN - Challenges

- increased latency observable for small batch jobs

- early adopter pains- sparse documentation- not the best API design

big data store

compressed data

Parquet - Motivation

- legacy: data stored in both row and column major

- requires expensive transpose on ingestion- were designing a “tiled file format” when

discovered parquet

Parquet - Challenges

- early adopter challenges- sparse documentation- needed to access package private APIs

big data

bigdf - Motivation

- born out of experience using spark for feature engineering

- creating classes for RDDs not reusable across projects

- SQL not expressive enough

bigdf - details- open source since Sep 2014- precedes Spark DataFrame, so built on spark-core

engine- experimenting with Catalyst using Spark DataFrame

APIs, looks promising- python and scala APIs- feature engineering library [not open source :-( ]- fast CSV reader(and other features) contributed to

spark-csv

bigdf - future

- wrapper around Spark DF - to protect from API changes- to add features e.g. “sparse column set” as “round-

trip time” for pull requests into large open source projects is high

Thanks!

www.ayasdi.comhttp://engineering.ayasdi.com

https://github.com/AyasdiOpenSource/bigdfhttp://www.ayasdi.com/company/careers/