My slides for Huguk #7

1. Using HFile outside
of HBase
Marc de Palol (marc@last.fm)
Huguk #7
19th November
1

2. Context: What’s Last.fm?
Last.fm is a:
– Music discovery website.
– powered by scrobbling.
– that provides personalized radio.
– And lots, lots of stats and informaLon about arLsts.
2

3. You can find numbers all around!
3

4. You can find numbers all around!
4

5. You can find numbers all around!
5

6. You can find numbers all around!
6

7. How does this work ?
7
Scrobble
Server
Chartserver
Chartserver
Memca
che
Web
Nodes
scrobble
Hadoop
Web / API
Users

8. How does this work ?
8
Scrobble
Server
Chartserver
Chartserver
Memca
che
Web
Nodes
scrobble
Hadoop
Web / API
Users

9. Closer look at this guy (chartserver)
• Java (used to be PHP).
• ThriZ API.
• Text file format + index.
• Disk I/O is the problem.
9
• Not only a Key‐Value store.
• It serves nearly all the data
we generate with Hadoop.

10. Closer look at this guy (chartserver)
• Java (used to be PHP).
• ThriZ API.
• Text file format + index.
• Disk I/O is the problem.
10
• Not only a Key‐Value store.
• It serves nearly all the data
we generate with Hadoop.

11. File Format
• Easy to grep / read, from the
command line.
• Server is easy to implement &
maintain.
• Very fast thanks to the index. Very
sparse though.
• Disk space not really and issue here.
We can always get rid of old indexes.
• Problem?
11
Key1 x Size
0
0
0
Key2 x Size
0
0
0
0
KeyN x Size
Key1 Value 1
Key1 Value 2
Key1 Value 3
Key1 Value 4
Key1 Value 5
Key2 Value 1
Key2 Value 2
Key2 Value 3
Key2 Value 4
Key2 Value 5
Key2 Value 6
...
...
...
...
...
...
...
...
...
KeyN Value 1
KeyN Value 2
KeyN Value 3
KeyN Value 4
KeyN Value 5
KeyN Value 6
KeyN Value 7
Index File Data File

12. File Format
• Easy to grep / read, from the
command line.
• Server is easy to implement &
maintain.
• Very fast thanks to the index. Very
sparse though.
• Disk space not really and issue here.
We can always get rid of old indexes.
• Problem?
• It takes more Hme to generate the
index than to create the Data File in
Hadoop.
12
Key1 x Size
0
0
0
Key2 x Size
0
0
0
0
KeyN x Size
Key1 Value 1
Key1 Value 2
Key1 Value 3
Key1 Value 4
Key1 Value 5
Key2 Value 1
Key2 Value 2
Key2 Value 3
Key2 Value 4
Key2 Value 5
Key2 Value 6
...
...
...
...
...
...
...
...
...
KeyN Value 1
KeyN Value 2
KeyN Value 3
KeyN Value 4
KeyN Value 5
KeyN Value 6
KeyN Value 7
Index File Data File

13. File Format
• Easy to grep / read, from the
command line.
• Server is easy to implement &
maintain.
• Very fast thanks to the index. Very
sparse though.
• Disk space not really and issue here.
We can always get rid of old indexes.
• Problem?
• It takes more Hme to generate the
index than to create the Data File in
Hadoop.
• Like... 6 Hmes more.
13
Key1 x Size
0
0
0
Key2 x Size
0
0
0
0
KeyN x Size
Key1 Value 1
Key1 Value 2
Key1 Value 3
Key1 Value 4
Key1 Value 5
Key2 Value 1
Key2 Value 2
Key2 Value 3
Key2 Value 4
Key2 Value 5
Key2 Value 6
...
...
...
...
...
...
...
...
...
KeyN Value 1
KeyN Value 2
KeyN Value 3
KeyN Value 4
KeyN Value 5
KeyN Value 6
KeyN Value 7
Index File Data File

14. SoluHon?
• Move to HBase (or another data storage system)
– Chartserver is not simply a key/value store.
– Lots of people in Last.fm want to use different things, for
different reasons.
•Our ops team do not want (!!) to maintain several different NoSql
systems around.
– This will take some Lme, some experimentaLon,
benchmarks and diplomacy.
14

15. 15
Our last meeLng to decide which NoSql database we should use.
Sysadmins dressed in funny yellow ou1it.

16. Requirements for the new file format:
• Binary:
– So it is smaller.
– Store thriZ serialized data.
• Compression friendly.
• Self indexed:
– We do not want an index file anymore.
• Hadoop friendly:
– Generated in Hadoop, we don’t want to preprocess it before serving.
• Java/C++/Python friendly:
– These are the languages used in the Data and M.I.R. teams.
16

17. Requirements for the new file format:
• Binary:
– So it is smaller.
– Store thriZ serialized data.
• Compression friendly:
• Self indexed:
– We do not want an index file anymore.
• Hadoop friendly:
– Generated in Hadoop, we don’t want to preprocess it before serving.
• Java/C++/Python friendly:
– These are the languages used in the Data and M.I.R. teams.
– Yeah, we sLll use C++.
17

18. !
KeyLen (int) ValLen (int) Key (byte[]) Value (byte[])
DATA BLOCK MAGIC (8B)
Key-Value (First)
……
Key-Value (Last)
Data Block 0
Data Block 1
Data Block 2
Meta Block 0
(Optional)
Meta Block 1
(Optional)
User Defined Metadata,
start with METABLOCKMAGIC
KeyLen
(vint)
Key
(byte[])
id
(1B)
ValLen
(vint)
Val
(byte[])
File Info
Size or ItemsNum (int)
LASTKEY (byte[])
AVG_KEY_LEN (int)
AVG_VALUE_LEN (int)
COMPARATOR (className)
Data Index
Meta Index
(Optional)
Index of Data Block 0
…
User Defined
INDEX BLOCK MAGIC (8B)
Index of Meta Block 0
…
Offset(long) DataSize (int) Key (byte[])
KeyLen (vint)
Trailer INDEX BLOCK MAGIC (8B)
Fixed File Trailer
(Go to next picture)
Offset(long) MetaSize (int) MetaNameLen (vint) MetaName (byte[])
3
HFile:
18
by Schubert Zang
hqp://cloudepr.blogspot.com
• Based on Google’s SSTable (From Bigtable)
• Keys and Values are byte strings.
• Keys are ordered.
• Sequence of blocks.
• Block index loaded into memory.
• Can be queried with hbase
org.apache.hadoop.hbase.io.hfile.HFile

19. HFile:
19
// create an HFile reader from a file.
Hfile.Reader reader = new HFile.Reader(fs,
filePath, new SimpleBlockCache(),true);
// load its info into memory.
reader.loadFileInfo();
// get a Scanner
HFileScanner scan = reader.getScanner(true,true);
// create the key we are interested in.
KeyValue kvKey = new KeyValue(Bytes.toBytes(key),
Bytes.toBytes(“f”),...);
// check if the key is in the file.
if (0 != scan.seekTo(kvKey.getKey()) {
log.error(“Couldn’t find the key”);
} else {
log.info(“Value:” +
scan.getKeyValue().getValue());
}

20. 20
Before coding...
some tests.

21. Some tests (generaHng the datasets).
21
Plain text format HFile
5.9 Gb (2.4Gb data, 3.5 Gb Index) 2.8 Gigabytes.
369 minutes (6 hours) 25 minutes (25 minutes)
Plain text ThriZ serialized
Exactly the same contents:
‐ 16.395.747 keys (16 million)
‐ 121.930.516 values (121 million)

22. Some tests (querying randomly).
22
Plain text format HFile
54 seconds. 6.11 seconds.
mean: 54 us mean: 6.11 us
stdev: 403 us stdev: 108 us
max: 72700 us max: 95300 us
min: 40.2 us min: 3.35 us
Querying with 1 million random keys.

23. Some tests (querying in order).
23
Plain text format HFile
10.695 seconds. (3 hours) 3287 seconds. (< 1 hour)
mean: 652 us mean: 201 us
stdev: 3120 us stdev: 2320 us
max: 464000 us max: 468000 us
min: 37.6 us min: 5.29 us
Querying with all the keys (16 million)

24. Some tests (querying randomly).
24

25. Some tests (querying all the keys).
25

26. Merging HFile with Chartserver.
• Changes in the Hadoop programs:
– We just created a new program that translated a Sequence File to an HFile.
– Shamelessly copy & pasted Todd Lipcon’s bulk load tool. [have a look at ‘Bibliography’]
• Changes in Chartserver.
– Know how to load the HFiles.
– Know how to access them.
• Status
– Not in producLon yet.
– Finishing some Junit tests.
26

27. 27
That’s it
Any doubts ?
oh... wait.

28. We are hiring! (http://www.last.fm/about/jobs)
28
Data Scientist
Purpose & Background of Role
We're seeking two top notch data scientists with strong programming skills to join the
small and very enthusiastic data and recommendations team at Last.fm. These two
positions are full-time and based in London.
Are you a superb data analyst as well as a hands-on implementer that understands the
trade-offs of the memory hierarchy and is able to work around constraints in disk speed,
memory size and CPU cycles? Are you familiar with all common data structures and their
complexity? Do you take pride in being clever and solving difficult problems creatively?
Are you full of ideas and always looking for new ways of making use out of data? Are you
an advocate for data-driven development and fully capable of conducting a proper A/B
test? Do you love music?
Requirements:
• Solid background in statistics and computer science
• Highly fluent in Python and either C++ or Java (or both)
• Comfortable with the Unix CLI and shell scripting
• Passion for machine learning and data visualisation
• Proficient with databases, both relational and non-relational
• Experience with Hadoop and analysing terabyte-scale datasets
• Familiar with data-driven development and split testing
• Basic understanding of common web technologies
• Track record in music information retrieval research is a plus

29. We are hiring! (http://www.last.fm/about/jobs)
29
Data Scientist
Purpose & Background of Role
We're seeking two top notch data scientists with strong programming skills to join the
small and very enthusiastic data and recommendations team at Last.fm. These two
positions are full-time and based in London.
Are you a superb data analyst as well as a hands-on implementer that understands the
trade-offs of the memory hierarchy and is able to work around constraints in disk speed,
memory size and CPU cycles? Are you familiar with all common data structures and their
complexity? Do you take pride in being clever and solving difficult problems creatively?
Are you full of ideas and always looking for new ways of making use out of data? Are you
an advocate for data-driven development and fully capable of conducting a proper A/B
test? Do you love music?
Requirements:
• Solid background in statistics and computer science
• Highly fluent in Python and either C++ or Java (or both)
• Comfortable with the Unix CLI and shell scripting
• Passion for machine learning and data visualisation
• Proficient with databases, both relational and non-relational
• Experience with Hadoop and analysing terabyte-scale datasets
• Familiar with data-driven development and split testing
• Basic understanding of common web technologies
• Track record in music information retrieval research is a plus
x 2

30. 30
That’s it
Any doubts ?
marc@last.fm
@lant

31. Bibliography.
• HFile:
– hqp://issues.apache.org/jira/browse/HBASE‐1818
– hqp://cloudepr.blogspot.com/2009/09/hfile‐block‐indexed‐file‐format‐to.html
– hqp://www.larsgeorge.com/2009/10/hbase‐architecture‐101‐storage.html
• Todd Lipcon’s Bulk load tool:
– hQp://hbase.apache.org/docs/r0.89.20100726/bulk‐loads.html
– TRUNK/org/apache/hadoop/hbase/mapreduce/ImportTsv.java
31