Parquet is an open source columnar storage format for Hadoop data. It was developed as a collaboration between Twitter and Cloudera to optimize IO and storage for analytics workloads on large datasets. Parquet supports efficient compression and encoding techniques that reduce storage size and enable faster scans by only loading the columns needed. It can be used with existing Hadoop tools and was implemented in Java, C++, and other languages to integrate with frameworks like Hive and Impala. Initial results at Twitter showed a 28% reduction in storage size and up to a 50% improvement in scan speeds compared to the previous Thrift format.

1. Parquet
Open columnar storage for Hadoop
Julien Le Dem @J_ Processing tools lead, analytics infrastructure at Twitter
http://parquet.io
http://parquet.io
Julien
1

2. Twitter Context
•
Twitter’s data
•
•
100TB+ a day of compressed data
•
•
200M+ monthly active users generating and consuming 500M+ tweets a day.
Scale is huge: Instrumentation, User graph, Derived data, ...
Analytics infrastructure:
•
•
Log collection pipeline
•
•
Several 1K+ nodes Hadoop clusters
Processing tools
Role of Twitter’s analytics infrastructure team
•
Platform for the whole company.
•
Manages the data and enables analysis.
•
Optimizes the cluster’s workload as a whole.
The Parquet Planers
Gustave Caillebotte
2
Julien
- Hundreds of TB
- Hadoop: a distributed file system and a parallel execution engine focused on data locality
- clusters: ad hoc, prod, ...
- Log collection pipeline (Scribe, Kafka, ...)
- Processing tools (Pig...)
- Every group at Twitter uses the infrastructure: spam, ads, mobile, search, recommendations, ...
- optimize space/io/cpu usage.
- the clusters are constantly growing and never big enough.
- We ingest hundreds of TeraBytes, IO bound processes:
Improved compression saves space. take advantage of column storage for better scans.

3. Twitter’s use case
•
Logs available on HDFS
•
Thrift to store logs
•
example: one schema has 87 columns, up to 7 levels of nesting.
struct LogEvent {
1: optional logbase.LogBase log_base
2: optional i64 event_value
3: optional string context
4: optional string referring_event
...
18: optional EventNamespace event_namespace
19: optional list<Item> items
20: optional map<AssociationType,Association> associations
21: optional MobileDetails mobile_details
22: optional WidgetDetails widget_details
23: optional map<ExternalService,string> external_ids
}
struct LogBase {
1: string transaction_id,
2: string ip_address,
...
15: optional string country,
16: optional string pid,
}
http://parquet.io
Julien
- Logs stored using Thrift with heavily nested data structure.
- Example: aggregate event counts per IP address for a given event_type.
We just need to access the ip_address and event_name columns
3

4. Goal
To have a state of the art columnar storage available
across the Hadoop platform
Hadoop is very reliable for big long running queries but also IO
heavy.
•
Incrementally take advantage of column based storage in existing
framework.
•
•
Not tied to any framework in particular
http://parquet.io
4

5. Columnar storage
•
Limits the IO to only the data that is needed.
•
Saves space: columnar layout compresses better
•
Enables better scans: load only the columns that
need to be accessed
•
Enables vectorized execution engines.
@EmrgencyKittens
http://parquet.io
5

6. Collaboration between Twitter and Cloudera:
•
Common file format definition:
•
•
•
Language independent
Formally specified.
Implementation in Java for Map/Reduce:
•
•
https://github.com/Parquet/parquet-mr
C++ and code generation in Cloudera Impala:
•
https://github.com/cloudera/impala
http://parquet.io
6

7. Twitter: Initial results
Data converted: similar to access logs. 30 columns.
Original format: Thrift binary in block compressed files
Scan time
Space
120.0%
100.0%
80.0%
60.0%
40.0%
20.0%
0%
100.00%
75.00%
50.00%
25.00%
0%
Thrift
Space
Parquet
1
Thrift
Parquet
30
columns
Space saving: 28% using the same compression algorithm
Scan + assembly time compared to original:
One column: 10%
All columns: 114%
http://parquet.io
Julien
7

8. Additional gains with dictionary
13 out of the 30 columns are suitable for dictionary encoding:
they represent 27% of raw data but only 3% of compressed data
Scan time
to thrift
Space
100.00%
60.0%
50.0%
75.00%
40.0%
50.00%
30.0%
20.0%
25.00%
0%
10.0%
Space
Parquet compressed (LZO)
Parquet Dictionary uncompressed
Parquet Dictionary compressed (LZO)
0%
1
13
Parquet compressed
Parquet dictionary compressed
Columns
Space saving: another 52% using the same compression algorithm (on
top of the original columnar storage gains)
Scan + assembly time compared to plain Parquet:
All 13 columns: 48% (of the already faster columnar scan)
http://parquet.io
Julien
8

9. Format
Row group
Column a
Page 0
Column b
Page 0
Column c
Page 0
Page 1
Page 1
Page 2
•
Row group: A group of rows in columnar format.
•
•
roughly: 50MB < row group < 1 GB
Page 4
Page 2
Page 3
Row group
Column chunk: The data for one column in a row group.
•
•
Page 2
Page 3
One (or more) per split while reading.
•
•
Max size buffered in memory while writing.
Page 1
Column chunks can be read independently for efficient scans.
Page: Unit of access in a column chunk.
•
Should be big enough for compression to be efficient.
•
Minimum size to read to access a single record (when index pages are available).
•
roughly: 8KB < page < 1MB
http://parquet.io
9

10. Format
Layout:
Row groups in columnar
format. A footer contains
column chunks offset and
schema.
Language independent:
Well defined format.
Hadoop and Cloudera
Impala support.
http://parquet.io
10

11. Nested record shredding/assembly
• Algorithm borrowed from Google Dremel's column IO
• Each cell is encoded as a triplet: repetition level, definition level, value.
• Level values are bound by the depth of the schema: stored in a compact form.
Schema:
Document
message Document {
required int64 DocId;
optional group Links {
repeated int64 Backward;
repeated int64 Forward;
}
}
Max rep.
level
Max def. level
DocId
Links
Backward
Forward
0
0
Links.Backward
DocId
Record:
DocId: 20
Links
Backward: 10
Backward: 30
Forward: 80
Columns
1
2
Links.Forward
1
2
Document
Column
value
R
D
DocId
Backward
10
30
Forward
80
0
Links.Backward
Links
20
0
10
0
2
Links.Backward
DocId
20
30
1
2
Links.Forward
80
0
2
http://parquet.io
Julien
- We convert from nested data structure to flat columnar so we need to store extra information to maintain the record structure.
- We can picture the schema as a tree. The columns are the leaves identified by the path in the schema.
- Additionally to the values we store two integers called the repetition level and definition level.
- When the value is NULL, the definition level captures up to which level in the path the value was defined.
- the repetition level is the last level in the path where there was a repetition. 0 means we have a new record.
- both levels are bound by the depth of the tree. They can be encoded efficiently using bit packing and run length encoding.
- a flat schema with all fields required does not have any overhead as R and D are always 0 in that case.
Reference: http://research.google.com/pubs/pub36632.html
11

12. Repetition level
record 1: [[a, b, c], [d, e, f, g]]
record 2: [[h], [i, j]]
Level: 0,2,2,1,2,2,2,0,1,2
Data: a,b,c,d,e,f,g,h,i,j
http://parquet.io
12

13. Differences of Parquet and ORC
Parquet:
•
Repetition/Definition levels capture the structure.
Document
DocId
=> one column per Leaf in the schema.
Links
Backward
•
Array<int> is one column.
•
Nullity/repetition of an inner node is stored in each of its children
•
Forward
=> One column independently of nesting with some redundancy.
ORC:
•
An extra column for each Map or List to record their size.
=> one column per Node in the schema.
•
Array<int> is two columns: array size and content.
•
=> An extra column per nesting level.
http://parquet.io
13

14. Iteration on fully assembled records
•
To integrate with existing row based engines (Hive, Pig, M/R).
a1
a2
•
b1
a2
Aware of dictionary encoding: enable optimizations.
a1
b2
a3
b3
a3
b1
b2
b3
•
Assembles projection for any subset of the columns:
- only those are loaded from disc.
Document
DocId
Document
Document
Links
Document
Links
Links
20
Backward
10
30
Backward
10
30
Forward
80
Forward
80
http://parquet.io
Julien
Explain in the context of example.
- columns: for implementing column based engines
- record based: for integrating with existing row based engines (Hive, Pig, M/R).
14

15. Iteration on columns
Row:
R
D
V
0
0
1
A
1
0
1
B
1
1
C
2
0
0
3
0
1
•
To implement column based execution engine
•
D<1 => Null
Iteration on triplets: repetition level, definition level, value.
•
R=1 => same row
Repetition level = 0 indicates a new record.
D
Encoded or decoded values: computing aggregations on
integers is faster than on strings.
•
http://parquet.io
Julien
15

16. APIs
•
Schema definition and record materialization:
Hadoop does not have a notion of schema, however Impala, Pig, Hive, Thrift,
Avro, ProtocolBuffers do.
•
•
Event-based SAX-style record materialization layer. No double conversion.
Integration with existing type systems and processing
frameworks:
•
•
Impala
•
Pig
•
Thrift and Scrooge for M/R, Cascading and Scalding
•
Cascading tuples
•
Avro
•
Hive
http://parquet.io
Julien
all define a schema in a different way.
16

17. Encodings
•
Bit packing:
1
3
2
0
0
2
2
0
01|11|10|00 00|10|10|00
•
•
•
Small integers encoded in the minimum bits required
Useful for repetition level, definition levels and dictionary keys
Run Length Encoding:
1
1
1
1
1
1
1
1
•
•
Cheap compression
•
Works well for definition level of sparse columns.
Dictionary encoding:
•
•
1
Used in combination with bit packing,
•
8
Useful for columns with few ( < 50,000 ) distinct values
Extensible:
•
Defining new encodings is supported by the format
http://parquet.io
17

18. Contributors
Main contributors:
•
Julien Le Dem (Twitter): Format, Core, Pig, Thrift integration, Encodings
•
Nong Li, Marcel Kornacker, Todd Lipcon (Cloudera): Format, Impala
Jonathan Coveney, Alex Levenson, Aniket Mokashi, Tianshuo Deng
(Twitter): Encodings, projection push down
•
•
Mickaël Lacour, Rémy Pecqueur (Criteo): Hive integration
Dmitriy Ryaboy (Twitter): Format, Thrift and Scrooge Cascading
integration
•
•
Tom White (Cloudera): Avro integration
•
Avi Bryant, Colin Marc (Stripe): Cascading tuples integration
•
Matt Massie (Berkeley AMP lab): predicate and projection push down
•
David Chen (Linkedin): Avro integration improvements
http://parquet.io
18

19. Parquet 2.0
- More encodings: more compact storage without heavyweight compression
•
Delta encodings: for integers, strings and dictionary
•
Improved encoding for strings and boolean
- Statistics: to be used by query planners and
- New page format: faster predicate push down,
http://parquet.io
19

20. How to contribute
Questions? Ideas?
Contribute at: github.com/Parquet
Look for tickets marked as “pick me up!”
http://parquet.io
20