How AdMobius uses Cascading in
AdTech Stack
Jyotirmoy Sundi
Sr Data Engineer in Lotame
(Acquired by LOTAME on March, 2014)
What does AdMobius do

AdMobius is a Mobile Audience Management
Platform (MAMP). It helps advertiser identify
mobile audiences by demographics and interest
through standard, custom, private segments
and reach them at scale.
Target effectively across all platforms in multiple devices
Laptop
Mobile
Ipod
Ipad
Wearables
Topics

Device graph building and scoring device links

Cascading Taps for Hive, MySQL, HBase

Modularized Testing

Optimal Config Setups

Running in YARN

Conclusion
AdMobius Stack
Cascading | Hive | Hbase | GiraphCascading | Hive | Hbase | Giraph
Hadoop | (Experimental Spark)Hadoop | (Experimental Spark)
RackspaceRackspace
YARN | MR1YARN | MR1
Custom WorkflowsCustom Workflows

Why Cascading
− Easy custom aggregators.
• In the existing MR framework it was very difficult
to write a series of complex aggregated logic and
run them in scale before making sure of its
correctness. You can do that in hive by UDFs or
UDAFs but we found it much easier in Cascading.
− Easy for Java Developers to understand
• visualize and write complicated workflows though
the concept of pipes, taps, tuples.
Workflow for audience profile scoring
Driven
https://driven.cascading.io/index.html#/apps/D818DD
Audience Profiling

Cascading is used to do
− complex aggregations
− create the device multi-dimensional vectors
− device pair scoring based on the vectors
− rule engine based filters

Size
− Total number of mobile devices ~ 2.7B
− ~500M devices in Giraph computation.
Example: Parallel aggregation of values across multiple fields.
Aggregations

No need to know group modes like in UDAF

Buffer

use for more complex grouping
operations

output multiple tuples per group

Aggregator (simple aggregations, prebuilt
aggregators like SumBy, CountBy)
public class MinGraphScoring extends BaseOperation implements Buffer{
@Override
public void operate(FlowProcess flowProcess, BufferCall bufferCall) {
Iterator<TupleEntry> arguments = bufferCall.getArgumentsIterator();
Graph g = new Graph();
while( arguments.hasNext() )
{
TupleEntry tpe = arguments.next();
ByteBuffer b = ByteBuffer.wrap((byte[])tpe.getObject("field1"););//use kyro
serialization
g.put(b)
}
Node[] nodes = g.nodes;
//For each pair of nodes : i,j {
double minmaxscore = scoring(g,i,j)
Tuple t1 = new Tuple(nodes[i].id ,nodes[j].id ,minmaxscore);
bufferCall.getOutputCollector().add(t1);
}
}
public class PotentialMatchAggregator extends
BaseOperation<PotentialMatchAggregator.IDList> implements
Aggregator<PotentialMatchAggregator.IDList> {
start(FlowProcess flowProcess, AggregatorCall<IDList> aggregatorCall) {
IDList idList = new IDList();
aggregatorCall.setContext(idList);
}
aggregate(FlowProcess flowProcess, AggregatorCall<IDList> aggregatorCall)
{
TupleEntry arguments = aggregatorCall.getArguments();
IDList idList = aggregatorCall.getContext();
idList.updateDev(amid, match);
}
complete(FlowProcess flowProcess, AggregatorCall<IDList> aggregatorCall)
{
IDList idList = aggregatorCall.getContext();
…...
}
Joins

CoGroup:

two pipes cant fit into memory

HashJoin

when one of the pipes fit into memory
Pipe jointermsPipe = new HashJoin(termsPipe, new
Fields("term_token"),dictionary, new Fields("word"), new
Fields("app","term_token","score","d_count","index","word"), new
InnerJoin());

CustomJoins and BloomJoin
Custom Src/Sink Taps

Cascading has good support to read/write to/from different form of
data sources. Slight tuning or change might be required but most of
code already exists.
− Hive (with different file formats), HBase, MySQL
− http://www.cascading.org/extensions/
− Set proper Config parameters while reading from source tap,
example while reading from Hbase Tap,
String tableName = "device_ids";
String[] familyNames = new String[] { "id:type1", "id:type2",
“id:type3”,...”id:typen” };
Scan scan = new Scan();
scan.setCacheBlocks(false);
scan.setCaching(10000);
scan.setBatch(10000);
Hive Src TapsExampleWorkflow.java
Tap dmTap = new HiveTableTap(HiveTableTap.SchemeType.SEQUENCE_FILE, admoFPbase, admoFPBasePartitions, dmFullFilter);
HiveTableTap.java
public class HiveTableTap extends GlobHfs {
static Scheme getScheme(SchemeType st) {
if(st.equals(SchemeType.SEQUENCE_FILE))
return new AdmobiusWritableSequenceFile(new Fields("value"), BytesWritable.class);
else if(st.equals(SchemeType.TEXT_TSV))
return new TextDelimited();
else
return null;
}
…..
}
Hive Sink Taps
ExampleWorkflow.java
Tap srcDstIdsSinkTap = new Hfs(new AdmobiusWritableSequenceFile(new Fields("value"), (Class<? extends Writable>)
Text.class),"/tmp/srcDstIdsSinkTap" , SinkMode.REPLACE);
HiveTableTap.java
public class HiveTableTap extends GlobHfs {
static Scheme getScheme(SchemeType st) {
if(st.equals(SchemeType.SEQUENCE_FILE))
return new AdmobiusWritableSequenceFile(new Fields("value"), BytesWritable.class);
else if(st.equals(SchemeType.TEXT_TSV))
return new TextDelimited();
else
return null;
}
…..
}
conf.setOutputFormat( SequenceFileOutputFormat.class );
valueValue = (Writable) (new Text(tupleEntry.getObject( 0 ).toString().getBytes()));
Hive table
CREATE TABLE CASCADING_HIVE_INTER
(
admo_id string,
segments string
)
PARTITIONED BY ( batch_id STRING )
ROW FORMAT DELIMITED FIELDS TERMINATED BY 't'
STORED AS SEQUENCEFILE
Good Practices

Use Checkpointing optimally

Use subassemblies instead of rewriting logic.
For further control pass additional parameters
to subassemblies.

Use Compression and SequenceFile() in sink
taps to chain multiple cascading workflows.

Use Failure Traps to filter faulty records.

Avoid creating too small or too long workflows.
Chain them in Oozie or similar workflow
management engines
− Example: workflows with 10-20 MR jobs are good
Some Properties for Optimal Performance
Problems with improper configuration
1. Set compression parameters : Jobs would run slow and
may take sometime double the time. Set the correct
compression Type based on cluster configs
2. mapred.reduce.tasks : Its required to be set manually
depending on the size of your job. Keeping it too low would
slow down reducer jobs.
3. small file issue : The input split files read by mappers
would be too small eventually bringing up more mappers
then required.
4. Any custom configuration parameters : You should set it
here and use getProperty to access them anywhere in the
data workflow
properties.setProperty("min_cutoff_score", "0.7");
FlowConnector flowConnector = new HadoopFlowConnector(properties);
Running in Yarn

Yarn deployment is smooth with cascading 2.5
− Make sure the config properties are set as per
YARN as they are different from MR1.
− While running in in workflow engines like oozie ,
make sure properties are set for
• mapred.job.classpath.files and mapred.cache.file
are set with all dependency files in colon
separated formatted
Cascading DSLs in other languages
Scalding (Scala)
PyCascading (Python)
cascading.jruby (Jruby)
Cascalog (Closure)

Thank you for your time

Q & A

Cascading talk in Etsy (http://www.meetup.com/cascading/events/169390262/)

  • 1.
    How AdMobius usesCascading in AdTech Stack Jyotirmoy Sundi Sr Data Engineer in Lotame (Acquired by LOTAME on March, 2014)
  • 2.
    What does AdMobiusdo  AdMobius is a Mobile Audience Management Platform (MAMP). It helps advertiser identify mobile audiences by demographics and interest through standard, custom, private segments and reach them at scale.
  • 3.
    Target effectively acrossall platforms in multiple devices Laptop Mobile Ipod Ipad Wearables
  • 4.
    Topics  Device graph buildingand scoring device links  Cascading Taps for Hive, MySQL, HBase  Modularized Testing  Optimal Config Setups  Running in YARN  Conclusion
  • 5.
    AdMobius Stack Cascading |Hive | Hbase | GiraphCascading | Hive | Hbase | Giraph Hadoop | (Experimental Spark)Hadoop | (Experimental Spark) RackspaceRackspace YARN | MR1YARN | MR1 Custom WorkflowsCustom Workflows
  • 6.
     Why Cascading − Easycustom aggregators. • In the existing MR framework it was very difficult to write a series of complex aggregated logic and run them in scale before making sure of its correctness. You can do that in hive by UDFs or UDAFs but we found it much easier in Cascading. − Easy for Java Developers to understand • visualize and write complicated workflows though the concept of pipes, taps, tuples.
  • 7.
    Workflow for audienceprofile scoring
  • 8.
  • 10.
    Audience Profiling  Cascading isused to do − complex aggregations − create the device multi-dimensional vectors − device pair scoring based on the vectors − rule engine based filters  Size − Total number of mobile devices ~ 2.7B − ~500M devices in Giraph computation.
  • 11.
    Example: Parallel aggregationof values across multiple fields.
  • 12.
    Aggregations  No need toknow group modes like in UDAF  Buffer  use for more complex grouping operations  output multiple tuples per group  Aggregator (simple aggregations, prebuilt aggregators like SumBy, CountBy)
  • 13.
    public class MinGraphScoringextends BaseOperation implements Buffer{ @Override public void operate(FlowProcess flowProcess, BufferCall bufferCall) { Iterator<TupleEntry> arguments = bufferCall.getArgumentsIterator(); Graph g = new Graph(); while( arguments.hasNext() ) { TupleEntry tpe = arguments.next(); ByteBuffer b = ByteBuffer.wrap((byte[])tpe.getObject("field1"););//use kyro serialization g.put(b) } Node[] nodes = g.nodes; //For each pair of nodes : i,j { double minmaxscore = scoring(g,i,j) Tuple t1 = new Tuple(nodes[i].id ,nodes[j].id ,minmaxscore); bufferCall.getOutputCollector().add(t1); } }
  • 14.
    public class PotentialMatchAggregatorextends BaseOperation<PotentialMatchAggregator.IDList> implements Aggregator<PotentialMatchAggregator.IDList> { start(FlowProcess flowProcess, AggregatorCall<IDList> aggregatorCall) { IDList idList = new IDList(); aggregatorCall.setContext(idList); } aggregate(FlowProcess flowProcess, AggregatorCall<IDList> aggregatorCall) { TupleEntry arguments = aggregatorCall.getArguments(); IDList idList = aggregatorCall.getContext(); idList.updateDev(amid, match); } complete(FlowProcess flowProcess, AggregatorCall<IDList> aggregatorCall) { IDList idList = aggregatorCall.getContext(); …... }
  • 15.
    Joins  CoGroup:  two pipes cantfit into memory  HashJoin  when one of the pipes fit into memory Pipe jointermsPipe = new HashJoin(termsPipe, new Fields("term_token"),dictionary, new Fields("word"), new Fields("app","term_token","score","d_count","index","word"), new InnerJoin());  CustomJoins and BloomJoin
  • 16.
    Custom Src/Sink Taps  Cascadinghas good support to read/write to/from different form of data sources. Slight tuning or change might be required but most of code already exists. − Hive (with different file formats), HBase, MySQL − http://www.cascading.org/extensions/ − Set proper Config parameters while reading from source tap, example while reading from Hbase Tap, String tableName = "device_ids"; String[] familyNames = new String[] { "id:type1", "id:type2", “id:type3”,...”id:typen” }; Scan scan = new Scan(); scan.setCacheBlocks(false); scan.setCaching(10000); scan.setBatch(10000);
  • 17.
    Hive Src TapsExampleWorkflow.java TapdmTap = new HiveTableTap(HiveTableTap.SchemeType.SEQUENCE_FILE, admoFPbase, admoFPBasePartitions, dmFullFilter); HiveTableTap.java public class HiveTableTap extends GlobHfs { static Scheme getScheme(SchemeType st) { if(st.equals(SchemeType.SEQUENCE_FILE)) return new AdmobiusWritableSequenceFile(new Fields("value"), BytesWritable.class); else if(st.equals(SchemeType.TEXT_TSV)) return new TextDelimited(); else return null; } ….. }
  • 18.
    Hive Sink Taps ExampleWorkflow.java TapsrcDstIdsSinkTap = new Hfs(new AdmobiusWritableSequenceFile(new Fields("value"), (Class<? extends Writable>) Text.class),"/tmp/srcDstIdsSinkTap" , SinkMode.REPLACE); HiveTableTap.java public class HiveTableTap extends GlobHfs { static Scheme getScheme(SchemeType st) { if(st.equals(SchemeType.SEQUENCE_FILE)) return new AdmobiusWritableSequenceFile(new Fields("value"), BytesWritable.class); else if(st.equals(SchemeType.TEXT_TSV)) return new TextDelimited(); else return null; } ….. } conf.setOutputFormat( SequenceFileOutputFormat.class ); valueValue = (Writable) (new Text(tupleEntry.getObject( 0 ).toString().getBytes()));
  • 19.
    Hive table CREATE TABLECASCADING_HIVE_INTER ( admo_id string, segments string ) PARTITIONED BY ( batch_id STRING ) ROW FORMAT DELIMITED FIELDS TERMINATED BY 't' STORED AS SEQUENCEFILE
  • 20.
    Good Practices  Use Checkpointingoptimally  Use subassemblies instead of rewriting logic. For further control pass additional parameters to subassemblies.  Use Compression and SequenceFile() in sink taps to chain multiple cascading workflows.  Use Failure Traps to filter faulty records.  Avoid creating too small or too long workflows. Chain them in Oozie or similar workflow management engines − Example: workflows with 10-20 MR jobs are good
  • 21.
    Some Properties forOptimal Performance
  • 22.
    Problems with improperconfiguration 1. Set compression parameters : Jobs would run slow and may take sometime double the time. Set the correct compression Type based on cluster configs 2. mapred.reduce.tasks : Its required to be set manually depending on the size of your job. Keeping it too low would slow down reducer jobs. 3. small file issue : The input split files read by mappers would be too small eventually bringing up more mappers then required. 4. Any custom configuration parameters : You should set it here and use getProperty to access them anywhere in the data workflow properties.setProperty("min_cutoff_score", "0.7"); FlowConnector flowConnector = new HadoopFlowConnector(properties);
  • 23.
    Running in Yarn  Yarndeployment is smooth with cascading 2.5 − Make sure the config properties are set as per YARN as they are different from MR1. − While running in in workflow engines like oozie , make sure properties are set for • mapred.job.classpath.files and mapred.cache.file are set with all dependency files in colon separated formatted
  • 24.
    Cascading DSLs inother languages Scalding (Scala) PyCascading (Python) cascading.jruby (Jruby) Cascalog (Closure)
  • 25.
     Thank you foryour time  Q & A