Simplifying Apache Cascading

Ming Yuan / Alyssa Romeo
Capital One
May 24th, 2016
Simplifying Apache Cascading

2
Apache Cascading
• Open source framework implementing the “chain of responsibility” design pattern
• Abstraction over MapReduce, Tez, or Flink processing engine when transforming big
data sets on Hadoop
• APIs for constructing and executing data-processing flows

3
PDS Framework on Cascading
A light-weight layer on top of Apache Cascading to
– Manage metadata for inputs and outputs in properties files
– Define data processing rules in properties files
– Support development in a parallel manner
– Make testing easier and more flexible PDS Framework

4
Case Studies
Source code Directly use Cascading After rewritten on the framework
TranOptimizerTrxnDtl.java 473 134
TrxnDtlTransformation.java 278 81
PlanTypeCdeCalculation.java 152 144
MyMain.java 12
Total 903 371
Source code Directly use Cascading After rewritten on the framework
PmsmJoin.java 210 87
JoinFunc.java 257 38
MyMain.java 12
Total 467 137
Cascading application 1 – 60% code reduction
Cascading application 2 – 70% code reduction

5
Root configuration
Data Processing Step
Sources SinkData-Processing Rules
Schema file Schema fileProcessing rules

6
Managing Multiple Steps on the Framework
1
2 3
Processing rules
4
5
Root configuration Schema files
6
Processing rules
Transformation step
Transformation step
Application ControllerApplication Initiator
5

7
Root Configuration
Root file entries configure application level components, including
– Hadoop configurations
– Global configuration entries for the application
– Definitions for File Taps (location and schema)
– Definitions for Hive Taps
ATPT_SCHEME_PATH=/devl/rwa/prtnrshp/prtntshp_whirl/whirl_atpt_mntry_dq_schema.txt
ATPT_RETAIN_FIELDS_PATH=/devl/rwa/prtnrshp/prtnrshp_whirl_trxn_optmzn/ATPT_retain_schema.txt
ATPT_DATA_PATH=/devl/rwa/prtnrshp/prtnrshp_whirl_trxn_optmzn/whirl_atpt_mntry_vldtd_hive_extract_us
ATGT_SCHEME_PATH=/devl/rwa/prtnrshp/prtntshp_whirl/whirl_atgt_mntry_dq_schema.txt
ATGT_RETAIN_FIELDS_PATH=/devl/rwa/prtnrshp/prtnrshp_whirl_trxn_optmzn/ATGT_retain_schema.txt
ATGT_DATA_PATH=/devl/rwa/prtnrshp/prtnrshp_whirl_trxn_optmzn/whirl_atgt_mntry_vldtd_hive_extract_us
HADOOP_PROPS_PATH=/devl/rwa/prtnrshp/prtnrshp_whirl_trxn_optmzn/hadoop.properties
FIRST_HIVE_TAP=/devl/rwa/prtnrshp/prtnrshp_whirl_trxn_optmzn/hiveone.properties
SECOND_HIVE_TAP=/devl/rwa/prtnrshp/prtnrshp_whirl_trxn_optmzn/hivetwo.properties
Root configuration

9
Schema Configuration – HiveTap
Schema file
DATA_BASE=dhdp_coaf
APP_COLUMN_NAMES=app_id, created_dt,…
APP_COLUMN_TYPES=Bigint, String, …
TABLE=MyTable
PARTITION_KEYS=odate
SER_LIB=org.apache.hadoop… (optional, by default it is ParquetHiveSerDe)
APP_PATH=hdfs://….
HiveTap hiveTap = getHiveTapFromConfig(“SECOND_HIVE_TAP”, sinkMode, booleanValue);

10
Data Processing Rules
• Processing rules are documented as properties
• Out-of-box macros define the transformation logic
• Framework translates the processing rules to Cascading API calls on the fly
ARRMT_ID_CHAIN obj(atpt_chain)
TRXN_SEQ_NUM atpt_mt_hi_tran_trk_id
POST_DT str(atpt_mt_posting_date)
TRXN_CD int(atpt_mt_txn_code)
AGT_ID substr(atpt_mt_hi_rep_id, 2, 4)
result.set(outputFields.getPos("ARRMT_ID_CHAIN"), argument.getObject(new Fields ("atpt_chain")));
result.set(outputFields.getPos("TRXN_SEQ_NUM"), argument.getObject(new Fields ("atpt_mt_hi_tran_trk_id")));
result.set(outputFields.getPos("POST_DT"), argument.getString(new Fields ("atpt_mt_posting_date")));
result.set(outputFields.getPos("TRXN_CD"), argument.getInteger(new Fields ("atpt_mt_txn_code")));
result.set(outputFields.getPos("AGT_ID"), argument.getString(new Fields ("atpt_mt_hi_rep_id")).substring(2,4));
Processing rules

11
Data Processing Rules -- Macros
Macro Names Syntax Functionality
obj TARGET obj(SOURCE) result.set(outputFields.getPos("TARGET"), argument.getObject(new Fields("SOURCE")));
default TARGET SOURCE result.set(outputFields.getPos("TARGET"), argument.getObject(new Fields ("SOURCE")));
as-is TARGET asis(default) result.set(outputFields.getPos("TARGET"), default));
string TARGET str(SOURCE) result.set(outputFields.getPos("TARGET"), argument.getString(new Fields ("SOURCE")));
int TARGET int(SOURCE) result.set(outputFields.getPos("TARGET"), argument.getInteger(new Fields("SOURCE")));
sub-string TARGET substr(SOURCE, 2, 4)
result.set(outputFields.getPos("TARGET"), argument.getString(new Fields
("SOURCE")).subString(2,4);
replace
TARGET replace(SOURCE, A,
B, C, D, default)
String rawValue = argument.getString(new Fields ("SOURCE"));
if (A equalto rawValue) then
result.set(outputFields.getPos("TARGET"), B);
if (C equalto rawValue) then
result.set(outputFields.getPos("TARGET"), D);
result.set(outputFields.getPos("TARGET"), “default”);
replace null
TARGET repnull(SOURCE,
default)
if (null equalto rawValue) then
result.set(outputFields.getPos("TARGET"), “default”);
else
result.set(outputFields.getPos("TARGET"), rawValue);
replace null
with
whitespace
TARGET repnullws(SOURCE)
result.set(outputFields.getPos("TARGET"), " ");
else
result.set(outputFields.getPos("TARGET"), rawValue);
not null TARGET notnull(SOURCE)
throw RuntimeException;
else result.set(outputFields.getPos("TARGET"), rawValue);
convert date
TARGET dateconv(SOURCE,
yyyymmdd, dd-mm-yyyy)
targetValue = rawValue from yyyymmdd to dd-mm-yyyy;
result.set(outputFields.getPos("TARGET"), targetValue);
move decimal TARGET movedeci(SOURCE,-2)
String rawValue = argument.getDouble(new Fields ("SOURCE"));
result.set(outputFields.getPos("TARGET"), rawValue / (10 ^ -2));

12
Exception Handling
“Whenever an operation fails and throws an exception, if there is an
associated trap, the offending Tuple is saved to the resource specified by the
trap Tap.”
-- Cascading documentation
FlowDef flowDef = FlowDef.flowDef().addSource(ipAmcpPipe, ipAmcpInTap)
.addSource(ipAtptPipe, ipAtptInTap)
.addTailSink(transformPipe, outTap)
.addTrap(ipAtptPipe, badRecordsTap);
}

13
How to Adopt the Framework
• Create a root configuration file
• Create a schema file for each input and output (or reuse DQ schema files)
• Define processing rules
• Add all of the files to HDFS
• Subclass the PDSBaseFuntion per processing step
@Override
protected void operate(FlowProcess flowProcess, FunctionCall<Tuple> functionCall) {
this.populateTupleSet(functionCall);
TupleEntry argument = functionCall.getArguments();
Tuple result = functionCall.getContext();
Fields outputFields = functionCall.getDeclaredFields();
result.set(outputFields.getPos("CHK_NUM"), check_number_calculation(argument));
functionCall.getOutputCollector().add(result);
}
@Override
protected String getConfigPath() {
return “/path/to/rulesfile”;
}

14
How to Adopt the Framework
• Subclass the PDSBaseDriver class and implement the “transform” method
• Create a “main” class
• Run tests
public class TestHarness {
public static void main(String[] args) {
new MyDriverImp().process("/path/to/rootconfig");
}
}
@Override
protected FlowDef transform() {
Fields pmamfields = getFieldsFromConfigEntry("PMAM_SCHEME_PATH");
String apparrFilePath = this.getFromConfigure("OUTPUT_DATA_PATH");
Tap pmsmTap = new Hfs( this.getTextDelimitedFromConfig("PMSM_SCHEME_PATH", null, false,
fieldDelimiter), apparrFilePath);
FlowDef flowDef = FlowDef.flowDef()
.addSource(ipAmcpPipe, ipAmcpInTap)
.addTailSink(transformPipe, outTap)
.addTrap(ipAtptPipe, badRecordsTap);
return flowDef;
}
Key Words in the
Root config file

15
Conclusion
• Benefits
– Reduce the total effort of developing and testing Cascading applications
• Provide a re-usable layer to reduce the amount of “plumbing” code
• Make Cascading modules configurable
– Improve the code quality
• Modularize Cascading applications and support best practices in Java coding
• Support additional features (such as logging and exception handling)
– Build an open architecture for future extension and integration
• Technical specification
– Compatible with JDK 1.5 and above; Jar file was compiled with JDK 1.7
– Tested with Cascading 2.5

16
For questions, please reach out to Ming.Yuan@capitalone.com

17
Appendix: PDSBaseDriver Class
Method Functionality Override
process(String path)
This method takes the path to the root configuration file, initializes all required
configurations, invokes "transform()" in its subclass, and executes Cascading
flows.
No
init(String path)
This method takes the path to the root configuration file, parses the file, and
stores configuration entries accordingly.
No
getFromConfig(String key)
This method takes a String-typed key, and returns a string-typed value is the key
has been used in the root configuration file. It returns null, otherwise.
No
getFieldsFromConfigEntry(
String key)
This method takes a String-typed key. In the root configuration file, if the key has
been assigned to a path to a schema file, the method returns a Fields object
based on all column names in the schema file. This Fields will be automatically
cached.
No
getFieldsFromConfigEntry(
String key, String[]
appendences)
This method takes a String-typed key in the root configuration file. If the key has
been assigned to a path to a schema file, it returns a Fields object with all
column names in the schema file and all names in the input string array. This
Fields will NOT be cached.
No
getTextDelimitedFromConfig(
String key,
String[] appendences,
boolean hasHeader,
String delimiter)
This methods creates and returns a TextDelimited object from a configuration
key in the root configuration files. You can use the second parameter to append
any column names programmatically. The third and forth parameters are for
input/output files.
No
transform()
Subclass should build a Flowdef object with application processing flow in this
method.
Yes

18
Appendix: PDSBaseFunction Class
Method Functionality Override
prepare(FlowProcess f,
OperationCall<Tuple> call)
This method overrides the same function in Cascading BaseOperation class. No
cleanup(FlowProcess f,
OperationCall<Tuple> call)
This method overrides the same function in Cascading BaseOperation class. No
init(String key, String
filePath)
This method parses a mapping rules file, and initializes the PDSBaseFunction
object.
No
populateTupleSet(
FunctionCall<Tuple> call)
This method populates values in its output tuple based on input values and pre-
defined processing rules.
No
getConfigPath() This method returns the path pointing to processing rules file in HDFS. Yes
Operate(
FlowProcess flowProcess,
FunctionCall<Tuple>
functionCall)
This methods should invoke the "populateTupleSet()" method in order to
execute pre-defined transformation rules, and it should invoke any additional
custom transformation methods for complex logic.
Yes

19
Appendix: Class Diagram
* Yellow color indicates components from Cascading package

Simplifying Apache Cascading

Related slideshows

More Related Content

Simplifying Apache Cascading