practical hadoop using pig

Practical Hadoop with Pig

Dave Wellman

#openwest @dwellman

How does it all work?

HDFSHadoop Shell

MR Data Structures Pig Commands

Pig Example

HDFS

HDFS has 3 main actors

The Name Node

The Name Node is “The Conductor”.

It directs the performance of the cluster.

The Data Nodes:

A Data Node stores blocks of data.

Clusters can be contain thousands of Data Nodes.

*Yahoo has a 40,000 node cluster.

The Client

The client is a window to the cluster.

The Name Node

The heart of the System.

The heart of the System.

Maintains a virtual File Directory.

The heart of the System.

Maintains a virtual File Directory.

Tracks all the nodes.

The heart of the System.

Maintains a virtual File Directory.

Tracks all the nodes.

Listens for “heartbeats” and “Block Reports” (more on this later).

The heart of the System.

Maintains a virtual File Directory.

Tracks all the nodes.

Listens for “heartbeats” and “Block Reports” (more on this later).

If the NameNode is down, the cluster is offline.

Storing Data

The Data Nodes

Add a Data Node:

Add a Data Node:

The Data Node says “Hello” to the Name Node.

Add a Data Node:

The Data Node says “Hello” to the Name Node.

The Name Node offers the Data Node a handshake with version requirements.

Add a Data Node:

The Data Node says “Hello” to the Name Node.

The Name Node offers the Data Node a handshake with version requirements.

The Data Node replies back to the Name Node, “Okay”, or “Shuts Down”.

Add a Data Node:

The Data Node says “Hello” to the Name Node.

The Name Node offers the Data Node a handshake with version requirements.

The Data Node replies back to the Name Node, “Okay”, or “Shuts Down”.

The Name Node hands the Data Node a NodeId that it remembers.

.

Add a Data Node:

The Data Node says “Hello” to the Name Node.

The Name Node offers the Data Node a handshake with version requirements.

The Data Node replies back to the Name Node, “Okay”, or “Shuts Down”.

The Name Node hands the Data Node a NodeId that it remembers.

The Data Node is now part of cluster and it checks in with the Name Node every 3 seconds.

Data Node Heartbeat:

Data Node Heartbeat:

The “check-in” is a simple HTTP Request/Response.

Data Node Heartbeat:

The “check-in” is a simple HTTP Request/Response.

This "check-in" is very important communication protocol that guarantees the health of the cluster.

Data Node Heartbeat:

The “check-in” is a simple HTTP Request/Response.

This "check-in" is very important communication protocol that guarantees the health of the cluster.

Block Reports – what data I have and is it okay.

Data Node Heartbeat:

The “check-in” is a simple HTTP Request/Response.

This "check-in" is very important communication protocol that guarantees the health of the cluster.

Block Reports – what data I have and is it okay.

Name Node controls the Data Nodes by issuing orders when they return and report their status.

Data Node Heartbeat:

The “check-in” is a simple HTTP Request/Response.

This "check-in" is very important communication protocol that guarantees the health of the cluster.

Block Reports – what data I have and is it okay.

Name Node controls the Data Nodes by issuing orders when they return and report their status.

Replicate Data, Delete Data, Verify Data

Data Node Heartbeat:

The “check-in” is a simple HTTP Request/Response.

This "check-in" is very important communication protocol that guarantees the health of the cluster.

Block Reports – what data I have and is it okay.

Name Node controls the Data Nodes by issuing orders when they return and report their status.

Replicate Data, Delete Data, Verify Data

Same process for all nodes within a cluster.

Writing Data

The client “tells” the NameNode the virtual directory location for the file.

A64 B64 C28

The client “tells” the NameNode the virtual directory location for the file.

The Client breaks the file into 64MB “blocks”

A64 B64 C28

The client “tells” the NameNode the virtual directory location for the file.

The Client breaks the file into 64MB “blocks”

The client “ask” the NameNode where the blocks go.

A64 B64 C28

A64 B64 C28

The client “tells” the NameNode the virtual directory location for the file.

The Client breaks the file into 64MB “blocks”

The client “ask” the NameNode where the blocks go.

The Client “stream” the blocks, in parallel, to the DataNodes.

A64 B64 C28

The client “tells” the NameNode the virtual directory location for the file.

The Client breaks the file into 64MB “blocks”

The client “ask” the NameNode where the blocks go.

The Client “stream” the blocks, in parallel, to the DataNodes.

DataNode(s) tells the NameNode they have the data via the block report

The client “tells” the NameNode the virtual directory location for the file.

The Client breaks the file into 64MB “blocks”

The client “ask” the NameNode where the blocks go.

The Client “stream” the blocks, in parallel, to the DataNodes.

DataNode(s) tells the NameNode they have the data via the block report

The NameNode tells the DataNode where to replicate the block.

A64 A64

A64

Reading Data

The client tells the NameNode it would like to read a file.

The client tells the NameNode it would like to read a file.

The NameNode reply’s with the list of blocks and the nodes the blocks are on.

A64

B64 C28

The client tells the NameNode it would like to read a file.

The NameNode reply’s with the list of blocks and the nodes the blocks are on.

The client request the first block from a DataNode

B64 C28

A64The client tells the NameNode it would like to read a file.

The NameNode reply’s with the list of blocks and the nodes the blocks are on.

The client request the first block from a DataNode

The client compares the checksum of the block against the manifest from the NameNode.

The client tells the NameNode it would like to read a file.

The NameNode reply’s with the list of blocks and the nodes the blocks are on.

The client request the first block from a DataNode

The client compares the checksum of the block against the manifest from the NameNode.

The client moves on to the next block in the sequence until the file has been read.

B64 C28

A64 B64 C28

Failure Recovery

A Data Node Fails to “check-in”

A64

A Data Node Fails to “check-in”

After 10 minutes the Name Node gives up on that Data Node.

A64X

A Data Node Fails to “check-in”

After 10 minutes the Name Node gives up on that Data Node.

When another node that has blocks originally assigned to the lost node checks-in, the name node sends a block replication command. A64XA64

A64

A Data Node Fails to “check-in”

After 10 minutes the Name Node gives up on that Data Node.

When another node that has blocks originally assigned to the lost node checks-in, the name node sends a block replication command.

The Data Node replicates that block of data. (Just like a write)

A64XA64

A64A64

Interacting with HadoopHDFS Shell Commands

HDFS Shell Commands.

> Hadoop fs –ls <args>

Same as unix or osx ls command.

/user/hadoop/file1 /user/hadoop/file2...

HDFS Shell Commands.

> Hadoop fs –mkdir <path>

Creates directories in HDFS using path.

HDFS Shell Commands.

> hadoop fs -copyFromLocal <localsrc> URI

Copy a file from your client to HDFS.

Similar to put command, except that the source is restricted to a local file reference.

HDFS Shell Commands.

> hadoop fs -cat <path>

Copies source paths to stdout.

HDFS Shell Commands.

> hadoop fs -copyToLocal URI <localdst>

Copy a file from HDFS to your client.

Similar to get command, except that the destination is restricted to a local file reference.

HDFS Shell Commands.

catchgrpchmodchowncopyFromLocalcopyToLocalcpdudusexpungegetgetmergels

lsrmkdirmovefromLocalmvputrmrmrsetrepstattailtesttexttouchz

Map Reduce Data StructuresBasic, Tuples & Bags

Basic Data Types:

Strings, Integers, Doubles, Longs, Byte,

Boolean, etc.

Advanced Data Types:

Tuples and Bags

Tuples are JSON like and simple.

raw_data: {date_time: bytearray,seconds: bytearray

}

Bags hold Tuples and Bags

element: {date_time: bytearray,seconds: bytearraygroup: chararray,ordered_list: {

date: chararray,hour: chararray,score: long

}}

Expert Advice:

Always know your data structures.

They are the foundation for all Map Reduce

operations.

Complex (deep) data structures will kill -9

performance.

Keep them simple!

Processing DataInteracting with Pig using Grunt

GRUNT

Grunt is a command line interface used to

debug pig jobs. Similar to Ruby IRB or

Groovy CLI.

Grunt is your best weapon against bad

pigs.

pig -x local

Grunt> |

GRUNT

Grunt> describe Element

Describe will display the data structure of

an Element

Grunt> dump Element

Dump will display the data represented by

an Element

GRUNT

> describe raw_data

Produces the output:> raw_data: { date_time: bytearray, items: bytearray }

Or in a more human readable form:Raw_data: {

date_time: bytearray,items: bytearray

}

GRUNT

> dump raw_data

You can dump terabytes of data to your screen,

so be careful.

(05/10/2011 20:30:00.0,0)(05/10/2011 20:45:00.0,0)(05/10/2011 21:00:00.0,0)(05/10/2011 21:15:00.0,0)...

Pig ProgramsMap Reduce Made Simple

Most PIG commands are assignments.

• The element names the collection of records that exist out in the cluster.

• It’s not a traditional programming variable. • It describes the data from the operation.• It does not change.

Element = Operation;

The SET command

Used to set a hadoop job variable. Like the name of

your pig job.

SET job.name 'Day over Day - [$input]’;

The REGISTER and DEFINE commands

-- Setup udf jarsREGISTER $jar_prefix/sidekick-hadoop-0.0.1.jar

DEFINE BUCKET_FORMAT_DATE com.sidekick.hadoop.udf.UnixTimeFormatter('MM/dd/yyyy HH:mm', 'HH');

The LOAD USING command

-- load in the data from HDFS

raw_data = LOAD '$input' USING PigStorage('\t') AS (date_time, items);

The FILTER BY command

Selects tuples from a relation based on some condition.

-- filter to the week we wantbroadcast_week = FILTER bucket_list BY (date >= '03-Oct-2011') AND (date <= '10-Oct-2011');

The GROUP BY command

Groups the data in one or multiple relations.

daily_stats = GROUP broadcast_week BY (date,

hour);

The FOREACH command

Generates data transformations based on columns of

data.

bucket_list = FOREACH raw_data GENERATE FLATTEN(DATE_FORMAT_DATE(date_time)) AS

date,MINUTE_BUCKET(date_time) AS hour, MAX_ITEMS(items) AS items;

*DATE_FORMAT_DATE is a user defined function, an advanced topic we’ll come to in a minute.

The GENERATE command

Use the FOREACH GENERATE operation to work with

columns of data.

bucket_list = FOREACH raw_data GENERATE FLATTEN(DATE_FORMAT_DATE(date_time)) AS

date,MINUTE_BUCKET(date_time) AS hour, MAX_ITEMSS(items) AS items;

The FLATTEN command

FLATTEN substitutes the fields of a tuple in place of the

tuple.

traffic_stats = FOREACH daily_stats GENERATE FLATTEN(GROUP),COUNT(broadcast_week) AS cnt,SUM(broadcast_week.items) AS total;

The STORE INTO USING command

Store function determine how data stored after a pig

job.

-- All done, now store it STORE final_results INTO '$output' USING PigStorage();

Demo Time! “Because, it’s all a big lie

until someone demos’ the code.”- Genghis Khan

Thank You.

- Genghis Khan