Choosing an HDFS data storage format- Avro vs. Parquet and more - StampedeCon 2015

Choosing an HDFS data storage format: Avro vs.
Parquet and more
Stephen O’Sullivan | @steveos

© 2015 SILICON VALLEY DATA SCIENCE LLC. ALL RIGHTS RESERVED.
@SVDataScience

3 © 2015 SILICON VALLEY DATA SCIENCE LLC. ALL RIGHTS RESERVED.
@SVDataScience
• Prioritize for highest
business value when using
emerging technology
• Design with outcomes in
mind
• Be agile: deliver initial
results quickly, then adapt
and iterate
• Collaborate constantly with
our customers and partners
OUR PHILOSOPHY

4 © 2015 SILICON VALLEY DATA SCIENCE LLC. ALL RIGHTS RESERVED.
@SVDataScience
AGENDA
Introduction
Data formats
How to choose
Schema evolution
Summary
Questions

@SVDataScience
• Storage formats
• What they do
DATA FORMATS

@SVDataScience
Data format
• Storage Format
• Text
• Sequence File
• Avro
• Parquet
• Optimized Row Columnar (ORC)

@SVDataScience
Text
• More specifically text = csv, tsv, json records…
• Convenient format to use to exchange with other
applications or scripts that produce or read
delimited files
• Human readable and parsable
• Data stores is bulky and not as efficient to query
• Do not support block compression

@SVDataScience
Sequence File
• Provides a persistent data structure for binary key-
value pairs
• Row based
• Commonly used to transfer data between Map
Reduce jobs
• Can be used as an archive to pack small files in
Hadoop
• Support splitting even when the data is
compressed

@SVDataScience
Avro
• Widely used as a serialization platform
• Row-based, offers a compact and fast binary
format
• Schema is encoded on the file so the data can be
untagged
• Files support block compression and are splittable
• Supports schema evolution

@SVDataScience
Parquet
• Column-oriented binary file format
• Uses the record shredding and assembly algorithm
described in the Dremel paper
• Each data file contains the values for a set of rows
• Efficient in terms of disk I/O when specific columns
need to be queried

@SVDataScience
Optimized Row Columnar
• Considered the evolution of the RCFile
• Stores collections of rows and within the collection
the row data is stored in columnar format
• Introduces a lightweight indexing that enables
skipping of irrelevant blocks of rows
• Splittable: allows parallel processing of row
collections
• It comes with basic statistics on columns (min ,max,
sum, and count)

@SVDataScience
• ..for write
• ..for read
HOW TO CHOOSE

@SVDataScience
How to choose: For write
• Functional Requirements:
• What type of data do you have?
• Is the data format compatible with your
processing and querying tools?
• What are your file sizes?
• Do you have schemas that evolve over time?

@SVDataScience
• Speed Concerns
• Parquet and ORC usually needs some
additional parsing to format the data which
increases the overall read time
• Avro as a data serialization format: works well from
system to system, handles schema evolution (more
on that later)
• Text is bulky and inefficient but easily readable and
parsable

@SVDataScience
0
20
40
60
80
100
120
140
160
TimeinSeconds
Narrow – Hortonworks (Hive 0.14 )
0
500
1000
1500
2000
2500
TimeinSeconds
Wide – Hortonworks (Hive 0.14)
Narrow: 10 million rows, 10 columns
Wide: 4 million rows, 1000 columns

@SVDataScience
0
10
20
30
40
50
60
70
TimeinSeconds
Narrow - hive-1.1.0+cdh5.4.2
0
100
200
300
400
500
600
700
TimeinSeconds
Wide - hive-1.1.0+cdh5.4.2

@SVDataScience
0
10
20
30
40
50
60
70
Text Avro Parquet
TimeinSeconds
Narrow - Spark 1.3
0
200
400
600
800
1000
1200
Text Avro Parquet
TimeinSeconds
Wide - Spark 1.3

@SVDataScience
0
200
400
600
800
1000
1200
1400
Megabytes
File sizes for narrow dataset
0
2000
4000
6000
8000
10000
12000
Megabytes
File sizes for wide dataset

@SVDataScience
• Use case
• Avro – Event data that can change over time
• Sequence File – Datasets shared between MR
jobs
• Text – Adding large amounts of data to HDFS
quickly

@SVDataScience
How to choose: For read
• Types of queries:
• Column specific queries, or few groups of
columns -> Use columnar format like Parquet or
ORC
• Compression of the file regardless the format
increases query speed times
• Text is really slow to read
• Parquet and ORC optimize read performance at
the expense of write performance

@SVDataScience
• Set up:
• Narrow dataset:
• 10 million rows, 10 columns
• Wide dataset:
• 4 million rows, 1000 columns
• Compression:
• Snappy, except for Avro which is deflate

@SVDataScience
0
10
20
30
40
50
60
70
Query 2 (5 conditions) Query 3 (10 conditions)
TimeinSeconds
Narrow Dataset - Hortonworks Hive 0.14.0.2.2.4.2
Text
Avro
Parquet
Sequence
ORC

@SVDataScience
0
100
200
300
400
500
600
700
800
Query 2 (5 conditions) Query 3 (10 conditions) Query 4 (20 conditions)
TimeinSeconds
Wide Dataset - Hortonworks Hive 0.14.0.2.2.4.2
Text
Avro
Parquet
Sequence
ORC

@SVDataScience
0
10
20
30
40
50
60
70
TimeinSeconds
Narrow Dataset - CDH hive-1.1.0+cdh5.4.2
Text
Avro
Parquet
Sequence
ORC

@SVDataScience
0
50
100
150
200
250
Query 1 (no
conditions)
Query 2 (5
conditions)
Query 3 (10
conditions)
Query 4 (20
conditions)
TimeinSeconds
Wide Dataset - CDH hive-1.1.0+cdh5.4.2
Text
Avro
Parquet
Sequence
ORC

@SVDataScience
0
1
2
3
4
5
6
7
8
TimeinSeconds
Narrow Dataset - CDH Impala
Text
Avro
Parquet
Sequence

@SVDataScience
0
5
10
15
20
25
30
Query 1 (0 filters) Query 2 (5 filters) Query 3 (10 filters) Query 4 (20 filters)
TimeinSeconds
Wide Dataset - CDH Impala
Text
Avro
Parquet
Sequence

@SVDataScience
Ran 4 queries (using Impala)
over 4 Million rows (70GB raw),
and 1000 columns (wide table)
0.00
10.00
20.00
30.00
40.00
50.00
60.00
70.00
80.00
Query 1 (0
filters)
Query 2 (5
filters)
Query 3 (10
filters)
Query 4 (20
filters)
Seconds
Query times for different data formats
Avro uncompress
Avro Snappy
Avro Deflate
Parquet
Seq uncompressed
Seq Snappy
Text Snappy
Text uncompressed

@SVDataScience
• Use case
• Avro – Query datasets that have changed over
time
• Parquet – Query a few columns on a wide
table

@SVDataScience
• What is schema
evolution?
• Data formats that evolve
• Examples
• Use cases
SCHEMA
EVOLUTION

@SVDataScience
Schema evolution
• What is schema evolution?
• Adding columns
• Renaming columns
• Removing columns
• Why do we need it?

@SVDataScience
Schema evolution
• Data formats that can evolve
• Avro
• Parquet
• Can only add columns at the end
• ORC
• It’s coming (That’s what they tell me ;) )…

@SVDataScience
Schema evolution
• Avro Example
• The data – Dr Who episodes
• Original Dr Who & new Dr Who
• http://www.theguardian.com/news/datablog/2010/aug/20/doctor-who-time-travel-
information-is-beautiful
• Avro schema for the original Dr Who
{"namespace": "drwho.avro",
"type": "record",
"name": "drwho",
"fields": [
{"name": "doctor_who_season", "type": "string"}, {"name": "doctor_actor", "type": "string"},
{"name": "episode_no", "type": "string"}, {"name": "episode_title", "type": "string"},
{"name": "date_from", "type": "string"}, {"name": "date_to", "type": "string"},
{"name": "estimated", "type": "string"}, {"name": "planet", "type": "string"},
{"name": "sub_location", "type": "string"}, {"name": "main_location", "type": "string"}
]}

@SVDataScience
Schema evolution
• Avro Example
• Original Dr Who data
doctor_who_
season doctor_actor episode_no episode_title date_from date_to estimated planet sub_location main_location
3 Pertwee 51 Spearhead from Space 1970 1990 y Earth England London and other
3 Pertwee 55 Terror of the Autons 1971 1971 y Earth England Luigi Rossini's Circus
3 Pertwee 58 Colony in Space 1971 2472 planet Uxarieus
3 Pertwee 59 The Daemons 1971 1971 y Earth England Devil's End; Wiltshire
3 Pertwee 60 Day of the Daleks 1972 2100 Earth England
Auderly House and
environs
3 Pertwee 63 The Mutants 1972 2900 Solos
3 Pertwee 64 The Time Monster -2000 1972 Earth/ Atlantis
3 Pertwee 64 The Time Monster 1972 -2000 Earth/ Atlantis
3 Pertwee 66 Carnival of Monsters 1972 1928 n
Indian Ocean;
Planet Inter Minor Ocean; alien planet
3 Pertwee 67 Frontier in Space 1972 2540 n
Planet Draconia;
Orgon Planet alien planets
3 Pertwee 68 Planet of the Daleks 1972 2540 y Planet Spiridon Alien Planet
3 Pertwee 69 The Green Death 2540 1973 y Earth UK Llanfairfach; Wales
3 Pertwee 70 The Time Warrior 1973 1200 n Earth UK Wessex Castle
3 Pertwee 71
Invasion of The
Dinosaurs 1200 1974 y Earth UK London

@SVDataScience
Schema evolution
• Avro Example
• Lets add, rename, and delete some columns
• Avro schema for the new Dr Who
{"namespace": "drwho.avro",
"type": "record",
"name": "drwho",
"fields": [
{"name": "drwho_season", "type": ["null","string"], "aliases": ["doctor_who_season"]},
{"name": "drwho_actor", "type": ["null","string"], "aliases": ["doctor_actor"]},
{"name": "episode_no", "type": ["null","string"]}, {"name": "episode_title", "type": ["null","string"]},
{"name": "date_from", "type": ["null","string"]}, {"name": "date_to", "type": ["null","string"]},
{"name": "estimated", "type": "string"}, {"name": "planet", "type": ["null","string"]},
{"name": "sub_location", "type": ["null","string"]}, {"name": "main_location", "type": ["null","string"]},
{"name": "hd", "type": "string", "default": "no"}
]}

@SVDataScience
Schema evolution
• Avro Example
• Original & New Dr Who data
drwho_season drwho_actor episode_no episode_title date_from date_to planet sub_location main_location hd
10 Tennant 201 New Earth 2006 5000000023 New Earth New … New York yes
10 Tennant 202 Tooth and claw 2006 1879 Earth Scotland
Torchwood house;
Near Balmoral yes
10 Tennant 203 school Reunion 2007 2007 Earth England Deffry Vale yes
10 Tennant 204
the Girl in the
Fireplace 1727 1744 Earth France Paris yes
3 Pertwee 51
Spearhead from
Space 1970 1990 Earth England London and other no
3 Pertwee 55
Terror of the
Autons 1971 1971 Earth England Luigi Rossini's Circus no
3 Pertwee 58 Colony in Space 1971 2472
planet
Uxarieus no
3 Pertwee 59 The Daemons 1971 1971 Earth England Devil's End; Wiltshire no
3 Pertwee 60 Day of the Daleks 1972 2100 Earth England
Auderly House and
environs no

@SVDataScience
Schema evolution
• Use cases
• New data added to an event stream
• Need to see historic data with new data (and
the schema has changed a lot)
• Business has changed the field/column name

@SVDataScience
QUESTIONS?

44
Yes, we’re hiring!
info@svds.com
THANK YOU
Stephen O’Sullivan
stephen@svds.com
@steveos
Demo code is here:
github.com/silicon-valley-data-
science/stampedecon-2015

Choosing an HDFS data storage format- Avro vs. Parquet and more - StampedeCon 2015

More Related Content

What's hot (20)

Viewers also liked (18)

Similar to Choosing an HDFS data storage format- Avro vs. Parquet and more - StampedeCon 2015 (20)

More from StampedeCon (20)

Recently uploaded (20)

Choosing an HDFS data storage format- Avro vs. Parquet and more - StampedeCon 2015

Editor's Notes