SlideShare a Scribd company logo

The Mechanics of Testing Large Data Pipelines (QCon London 2016)

Mathieu Bastian, profile picture
Mathieu Bastian

The document discusses strategies and tools for testing large data pipelines, emphasizing the importance of automated testing for code and data validation to mitigate bugs and failures. It outlines various testing challenges, including data variability and the need for realistic data, while providing techniques for creating test environments and utilizing schemas for validation. Finally, it highlights validation strategies for both input and output data to ensure robust data pipelines and reviews specific tools for testing Hadoop jobs and Pig scripts.

Technology
Related topics:
Data Mining Insights • Data Analysis Insights • Software Testing Insights • Insights on Software Development
1 of 74
Downloaded 113 times
1
2
3
4
Most read
5
Most read
6
7
8
9
10
11
12
13
14
Most read
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
THE MECHANICS OF TESTING LARGE DATA PIPELINES MATHIEU BASTIAN Head of Data Engineering, GetYourGuide @mathieubastian www.linkedin.com/in/mathieubastian QCon London 2015
Outline ▸ Motivating example ▸ Challenges ▸ Testing strategies ▸ Validation Strategies ▸ Tools Integration Tests ArchitectureUnit Test
Data Pipelines often start simple
Users E-commerce website Search App Views Offline Dashboard Search Metrics Views HDFS They have one use-case and one developer
But there are many other use- cases Recommender Systems Anomaly Detection Search Ranking A/B Testing Spam Detection Sentiment Analysis Topic Detection Trending TagsQuery Expansion Customer Churn Prediction Related searches Fraud Prediction Bidding Prediction Machine Translation Signal Processing Content Curation Sentiment Analysis Image recognition Optimal pricing Location normalization Standardization Funnel Analysis
additional events and logsDevelopers add Users E-commerce website Search App Clicks Views Offline Dashboard Search Metrics Clicks Views HDFS
third-party dataDevelopers add Users E-commerce website 3rd parties Search App Clicks Views A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views A/B Logs HDFS
search ranking predictionDevelopers add Users E-commerce website 3rd parties Search App Clicks Views A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views Training data Training & validation Model Clicks Views Features transformation A/B Logs HDFS
personalized user featuresDevelopers add Users E-commerce website 3rd parties Search App Clicks Views ProfilesUser Database A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views Training data Training & validation Model Clicks Views Profiles Features transformation A/B Logs HDFS
query extensionDevelopers add Users E-commerce website 3rd parties Search App Clicks Views ProfilesUser Database A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views Training data Training & validation Model Clicks Views Profiles Features transformation A/B Logs Filter queries Query extension RDBMS Views Training data HDFS
Developers add recommender system Users E-commerce website 3rd parties Search App Clicks Views ProfilesUser Database A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views Training data Training & validation Model Clicks Views Profiles Features Features transformation Features NoSQL Compute recommendations A/B Logs Filter queries Query extension RDBMS Views Training data HDFS
Data Pipelines can grow very large
That is a lot of code and data
Code contain bugs Industry Average: about 15 - 50 errors per 1000 lines of delivered code.
Data will change Industry Average: ?
Embrace automated testing of code validation of data
Because it delivers ▸ Testing ▸ Tested code has less bugs ▸ Gives the confidence to iterate quickly ▸ Scales well to multiple developers ▸ Validation ▸ Reduce manual testing ▸ Avoid catastrophic failures
But it’s challenging ▸ Testing ▸ Need data to test "realistically" ▸ Not running locally, can be expensive ▸ Tooling weaknesses ▸ Validation ▸ Data sources out of our control ▸ Difficult to test machine learning models
Reality check Source: @SteveGodwin, QCon London 2016
Manual testing Waiting Coding Looking at logs Code Upload Run workflow Look at logs ▸ Time Spent
Testing strategies
Prepare environment ▸ Care about tests from the start of your project ▸ All jobs should be functions (output only depends on input) ▸ Safe to re-run the job ▸ Does the input data still exists? ▸ Would it push partial results? ▸ Centralize configurations and no hard-coded paths ▸ Version code and timestamp data
Unit test locally ▸ Test locally each individual job ▸ Tests its good code ▸ Tests expected failures ▸ Need to overcome challenges with fake data creation ▸ Complex structures and numerous data sources ▸ Too small to be meaningful ▸ Need to specify a different configuration
Build from schemas Fake data creation based on schemas. Compare: Customer c = Customer.newBuilder().
 setId(42).
 setInterests(Arrays.asList(new Interest[]{
 Interest.newBuilder().setId(0).setName("Ping-Pong").build()
 Interest.newBuilder().setId(1).setName(“Pizza").build()})) .build(); vs Map<String, Object> c = new HashMap<>();
 c.put("id", 42);
 Map<String, Object> i1 = new HashMap<>();
 i1.put("id", 0);
 i1.put("name", "Ping-Pong");
 Map<String, Object> i2 = new HashMap<>();
 i2.put("id", 1);
 i2.put("name", "Pizza");
 c.put("interests", Arrays.asList(new Map[] {i1, i2}));
Build from schemas Avro Schema example { "type": "record", "name": "Customer", "fields": [{ "name": "id", "type": "int" }, { "name": "interests", "type": { "type": "array", "items": { "name": "Interest", "type": "record", "fields": [{ "name": "id", "type": "int" }, { "name": "name", "type": ["string", "null"] }] } } } ] } nullable field
Complex generators ▸ Developed in the field of property-based testing //Small Even Number Generator val smallEvenInteger = Gen.choose(0,200) suchThat (_ % 2 == 0) ▸ Goal is to simulate, not sample real data ▸ Define complex random generators that match properties (e.g. frequency) ▸ Can go beyond unit-testing and generate complex domain models ▸ https://www.scalacheck.org/ for Scala/Java is a good starting point for examples
Integration test on sample data ▸ Integration test the entire workflow ▸ File paths ▸ Configuration ▸ Evaluate performance ▸ Sample data ▸ Large enough to be meaningful ▸ Small enough to speed-up testing JOB A JOB B JOB C JOB D
Validation strategies
Where it fail Control Difficulty Model biases Bug Noisy data Schema changes Missing data
Input and output validation Make the pipeline robust by validating inputs and outputs Input Input Input Workflow Production ValidationValidation
Input Validation
Input data validation Input data validation is a key component of pipeline robustness. The goal is to test the entry points of our system for data quality. ETL RDBMS NOSQL EVENTS TWITTER DATA PIPELINE
Why it matters ▸ Bad input data will most likely degrade the output ▸ It likely will fail silently ▸ Because data will change ▸ Data migrations: maintenance, cluster update, new infrastructure ▸ Events change due to product evolution ▸ Data dependencies updated
Input data validation ▸ Validation code should ▸ Detect pathological data and fail early ▸ Deal with expected data variability ▸ Example issues: ▸ Missing values, encoding issues, etc. ▸ Schema changes ▸ Duplicates rows ▸ Data order changes
Pathological data ▸ Value ▸ Validity depends on a single, independent value. ▸ Easy to validate on streams of data ▸ Dataset ▸ Validity depends on the entire dataset ▸ More difficult to validate as it needs a window of data
Metadata validation Analyzing metadata is the quickest way to validate input data ▸ Number of records and file sizes ▸ Hadoop/Spark counters ▸ Number of map/reduce records, size ▸ Record-level custom counters ▸ Average text length ▸ Task-level custom counters ▸ Min/Max/Median values
Hadoop/Spark counters Results can be accessed programmatically and checked
Control inputs with Schemas ▸ CSVs aren’t robust to change, use Schemas ▸ Makes expected data explicit and easy to test against ▸ Gives basic validation for free with binary serialization (e.g. Avro, Thrift, Protocol Buffer) ▸ Typed (integer, boolean, lists etc.) ▸ Specify if value is optional ▸ Schemas can be evolved without breaking compatibility
Output Validation
Why it matters ▸ Humans makes mistake, we need a safeguard ▸ Rolling back data is often complex ▸ Bad output propagates to downstream systems Example with a recommender system // One recommendation set per user { "userId": 42, "recommendations": [{ "itemId": 1456, "score": 0.9 }, { "itemId": 4232, "score": 0.1 }], "model": "test01" }
Check for anomalies Simple strategies similar to input data validation ▸ Record level (e.g. values within bounds) ▸ Dataset level (e.g. counts, order) Challenges around relevance evaluation ▸ When supervised, use a validation dataset and threshold accuracy ▸ Introduce hypothetical examples
Incremental update as validation Join with the previous “best" output ▸ Allows fine comparisons ▸ Incremental framework can be extended to ▸ Only recompute recommendations that have changed ▸ Produce variations metric between different models Daily Recommendations Compute daily recommendationsHDFS Recommendations Yesterday Recommendations Join with previous result
External validation Even in automated environment it is possible to validate with humans ▸ Example: Search ranking evaluation ▸ Solution: Crowdsourcing ▸ Complex validation that requires training ▸ Can be automated through APIs
Mitigate risk with A/B testing Gradually rolling out data products improvements reduces the need for complex output validation ▸ Experiment can be controlled online or offline ▸ Online: Push multiple set of recommendations (1 per model) ▸ Offline: Split users and push unique set of recommendations userId -> [{
 "model": "test01",
 "recommendations": [{...}]
 }, {
 "model": "test02",
 "recommendations": [{...}]
 }] A B userId -> {
 "model": "test01",
 "recommendations": [{...}]
 }
Mitigate risk with A/B testing Important ▸ Log model variation downstream in logs ▸ Encapsulate model logic FEATURE 1-A MODEL A MODEL B FEATURE 1 FEATURE 2 MODEL A MODEL B A BA B FEATURE 1-B FEATURE 2-A FEATURE 2-B
Appendix
Hadoop Testing
Two ways to test Hadoop jobs ▸ MRUnit ▸ Java library to test MapReduce jobs in a simulated environment ▸ Last release June 2014 ▸ MiniCluster ▸ Utility to locally run a fully-functional Hadoop cluster in a test environment ▸ Ships with Hadoop itself
MiniMRCluster ▸ Advantages ▸ Behaves like a real cluster, including setup and configuration ▸ Can be used to test multiple jobs (integration testing) ▸ Disadvantages ▸ Very slow compared to unit testing Java code
MRUnit ▸ Advantages ▸ Faster ▸ Less boilerplate code ▸ Disadvantages ▸ Need to replicate job configuration ▸ Only built to test map and reduce functions ▸ Difficult to make it work with custom input formats (e.g. Avro)
MiniMRCluster setup* Setup MR cluster and obtain FileSystem @BeforeClass
 public void setup() {
 Configuration dfsConf = new Configuration();
 dfsConf.set(MiniDFSCluster.HDFS_MINIDFS_BASEDIR, new File("./target/ hdfs/").getAbsolutePath());
 _dfsCluster = new MiniDFSCluster.Builder(dfsConf).numDataNodes(1).build();
 _dfsCluster.waitClusterUp();
 _fileSystem = _dfsCluster.getFileSystem();
 
 YarnConfiguration yarnConf = new YarnConfiguration();
 yarnConf.setFloat(YarnConfiguration.NM_MAX_PER_DISK_UTILIZATION_PERCENTAGE, 99.0f);
 yarnConf.setInt(YarnConfiguration.RM_SCHEDULER_MINIMUM_ALLOCATION_MB, 64);
 yarnConf.setClass(YarnConfiguration.RM_SCHEDULER, FifoScheduler.class, ResourceScheduler.class);
 _mrCluster = new MiniMRYarnCluster(getClass().getName(), taskTrackers);
 yarnConf.set("fs.defaultFS", _fileSystem.getUri().toString());
 _mrCluster.init(yarnConf);
 _mrCluster.start();
 } * Hadoop version used 2.7.2
Keep the test file clean of boilerplate code Best is to wrap the start/stop code into a TestBase class /**
 * Default constructor with one task tracker and one node.
 */
 public TestBase() { ... }
 
 @BeforeClass
 public void startCluster() throws IOException { ... }
 
 @AfterClass
 public void stopCluster() throws IOException { ... }
 
 /**
 * Returns the Filesystem in use.
 *
 * @return the filesystem used by Hadoop.
 */
 protected FileSystem getFileSystem() {
 return _fileSystem;
 }
Initialize and clean HDFS before/after each test Clean up and initialize file system before each test private final Path _inputPath = new Path("/input");
 private final Path _cachePath = new Path("/cache");
 private final Path _outputPath = new Path("/output");
 
 @BeforeMethod
 public void beforeMethod(Method method) throws IOException {
 getFileSystem().delete(_inputPath, true);
 getFileSystem().mkdirs(_inputPath);
 getFileSystem().delete(_cachePath, true);
 }
 
 @AfterMethod
 public void afterMethod(Method method) throws IOException {
 getFileSystem().delete(_inputPath, true);
 getFileSystem().delete(_cachePath, true);
 getFileSystem().delete(_outputPath, true);
 }
Run MiniCluster Test Clean up and initialize file system before each test @Test
 public void testBasicWordCountJob() throws IOException, InterruptedException, ClassNotFoundException {
 writeWordCountInput();
 configureAndRunJob(new BasicWordCountJob(), "BasicWordCountJob", _inputPath, _outputPath);
 checkWordCountOutput();
 } private void configureAndRunJob(AbstractJob job, String name, Path inputPath, Path outputPath) throws IOException, ClassNotFoundException, InterruptedException {
 Properties _props = new Properties();
 _props.setProperty("input.path", inputPath.toString());
 _props.setProperty("output.path", outputPath.toString());
 job.setProperties(_props);
 job.setName(name);
 job.run();
 }
MRUnit setup Setup MapDriver and ReduceDriver BasicWordCountJob.Map mapper;
 BasicWordCountJob.Reduce reducer;
 MapDriver<LongWritable, Text, Text, IntWritable> mapDriver;
 ReduceDriver<Text, IntWritable, Text, IntWritable> reduceDriver;
 
 @BeforeClass
 public void setup() {
 mapper = new BasicWordCountJob.Map();
 mapDriver = MapDriver.newMapDriver(mapper);
 reducer = new BasicWordCountJob.Reduce();
 reduceDriver = ReduceDriver.newReduceDriver(reducer);
 }
Run MRUnit test Set Input/Output and run Test @Test
 public void testMapper() throws IOException {
 mapDriver.withInput(new LongWritable(0), new Text("banana pear banana"));
 mapDriver.withOutput(new Text("banana"), new IntWritable(1));
 mapDriver.withOutput(new Text("pear"), new IntWritable(1));
 mapDriver.withOutput(new Text("banana"), new IntWritable(1));
 mapDriver.runTest();
 }
 
 @Test
 public void testReducer() throws IOException {
 reduceDriver.withInput(new Text("banana"), Arrays.asList(new IntWritable(1), new IntWritable(1)));
 reduceDriver.withInput(new Text("pear"), Arrays.asList(new IntWritable(1)));
 reduceDriver.withOutput(new Text("banana"), new IntWritable(2));
 reduceDriver.withOutput(new Text("pear"), new IntWritable(1));
 reduceDriver.runTest();
 }
Most common pitfall ▸ With both MiniMRCluster and MRUnit one spend most of the time ▸ Creating fake input data ▸ Verifying output data ▸ Solutions ▸ Use rich data structures format (e.g. Avro, Thrift) ▸ Use automated Java classes generation
Other common pitfalls ▸ MiniMRCluster ▸ Enable Hadoop INFO logging so you can see real job failure causes ▸ Beware of partitioning or sorting issues unrevealed when testing with too few rows and number of nodes ▸ The API has changed over the years, difficult to find examples ▸ MRUnit ▸ Custom serialization issues (e.g. Avro, Thrift)
Pig Testing
Introducing PigUnit ▸ PigUnit ▸ Official library to unit tests Pig script ▸ Ships with Pig (latest version 0.15.0) ▸ The principle is easy 1. Generate test data 2. Run script with PigUnit 3. Verify output ▸ Runs locally but can be run on a cluster too
Script example WordCount example ‣ Input and output are standard formats ‣ Uses variables $input and $output text = LOAD '$input' USING TextLoader();
 
 flattened = FOREACH text GENERATE flatten(TOKENIZE((chararray)$0)) as word;
 grouped = GROUP flattened by word;
 result = FOREACH grouped GENERATE group, (int)COUNT($1) AS cnt;
 sorted = ORDER result BY cnt DESC;
 
 STORE sorted INTO '$output' USING PigStorage('t');
PigTestBase Create PigTest object protected final FileSystem _fileSystem;
 
 protected PigTestBase() {
 System.setProperty("udf.import.list", StringUtils.join(Arrays.asList("oink.", "org.apache.pig.piggybank."), ":"));
 fileSystem = FileSystem.get(new Configuration());
 }
 
 /**
 * Creates a new <em>PigTest</em> instance ready to be used.
 *
 * @param scriptFile the path to the Pig script file
 * @param inputs the Pig arguments
 * @return new PigTest instance
 */
 protected PigTest newPigTest(String scriptFile, String[] inputs) {
 PigServer pigServer = new PigServer(ExecType.LOCAL);
 Cluster pigCluster = new Cluster(pigServer.getPigContext());
 return new PigTest(scriptFile, inputs, pigServer, pigCluster);
 }
Test using aliases getAlias() allows to obtain the data anywhere in the script @Test
 public void testWordCountAlias() throws IOException, ParseException {
 //Write input data
 BufferedWriter writer = new BufferedWriter(new FileWriter(new File("input.txt")));
 writer.write("banana pear banana");
 writer.close();
 
 PigTest t = newPigTest("pig/src/main/pig/wordcount_text.pig", new String[] {"input=input.txt", "output=result.csv"});
 
 Iterator<Tuple> tuples = t.getAlias("sorted");
 Assert.assertTrue(tuples.hasNext());
 Tuple tuple = tuples.next();
 Assert.assertEquals(tuple.get(0), "banana");
 Assert.assertEquals(tuple.get(1), 2);
 Assert.assertTrue(tuples.hasNext());
 tuple = tuples.next();
 Assert.assertEquals(tuple.get(0), "pear");
 Assert.assertEquals(tuple.get(1), 1);
 }
Test using mock and assert ▸ mockAlias allows to substitute input data ▸ assertOutput allows to compare String output data @Test
 public void testWordCountMock() throws IOException, ParseException {
 //Write input data
 BufferedWriter writer = new BufferedWriter(new FileWriter(new File("input.txt")));
 writer.write("banana pear banana");
 writer.close();
 
 PigTest t = newPigTest("pig/src/main/pig/wordcount_text.pig", new String[] {"input=input.txt", "output=null"});
 t.runScript();
 t.assertOutputAnyOrder("sorted", new String[]{"(banana,2)", "(pear,1)"});
 }
Both of these tools have limitations ▸ Built around standard input and output (Text, CSVs etc.) ▸ Realistically most of our data is in other formats (e.g. Avro, Thrift, JSON) ▸ Does not test the STORE function (e.g. schema errors) ▸ getAlias() is especially difficult to use ▸ Need to remember field position: tuple.get(0) ▸ assertOutput() only allows String comparison ▸ Cumbersome to write complex structures (e.g. bags of bags)
Example with Avro input/output ▸ Focus on testing script’s output ▸ Difficulty is to generate dummy Avro data and compare result text = LOAD '$input' USING AvroStorage();
 
 flattened = FOREACH text GENERATE flatten(TOKENIZE(body)) as word;
 grouped = GROUP flattened by word;
 result = FOREACH grouped GENERATE group AS word, (int)COUNT($1) AS cnt;
 sorted = ORDER result BY cnt DESC;
 
 STORE result INTO '$output' USING AvroStorage(); ▸ By default, PigUnit doesn’t execute the STORE, but it can be overridden pigTest.unoverride("STORE");
Simple utility classes for Avro ▸ BasicAvroWriter ▸ Writes Avro file on disk based on a schema ▸ Supports GenericRecord and SpecificRecord ▸ BasicAvroReader ▸ Reads Avro file, the schema heads the file ▸ Also supports GenericRecord and SpecificRecord
Test with Avro GenericRecord ▸ Create Schema with SchemaBuilder, write data, run script, read result and compare @Test
 public void testWordCountGenericRecord() throws IOException, ParseException {
 Schema schema = SchemaBuilder.builder().record("record").fields().
 name("text").type().stringType().noDefault().endRecord();
 GenericRecord genericRecord = new GenericData.Record(schema);
 genericRecord.put("text", "banana apple banana");
 
 BasicAvroWriter writer = new BasicAvroWriter(new Path(new File("input.avro").getAbsolutePath()), schema, getFileSystem());
 writer.append(genericRecord);
 
 PigTest t = newPigTest("pig/src/main/pig/wordcount_avro.pig", new String[] {"input=input.avro", "output=sorted.avro"});
 t.unoverride("STORE");
 t.runScript();
 
 //Check output
 BasicAvroReader reader = new BasicAvroReader(new Path(new File("sorted.avro").getAbsolutePath()), getFileSystem());
 Map<Utf8, GenericRecord> result = reader.readAndMapAll("word");
 Assert.assertEquals(result.size(), 2);
 Assert.assertEquals(result.get(new Utf8("banana")).get("cnt"), 2);
 Assert.assertEquals(result.get(new Utf8("apple")).get("cnt"), 1);
 }
Test with Avro SpecificRecord ▸ Use InputRecord and OutputRecord generated Java classes, write data, run script, read result and compare @Test
 public void testWordCountSpecificRecord() throws IOException, ParseException {
 InputRecord input = InputRecord.newBuilder().setText("banana apple banana").build();
 
 BasicAvroWriter<InputRecord> writer = new BasicAvroWriter<InputRecord>(new Path(new File("input.avro").getAbsolutePath()), input.getSchema(), getFileSystem());
 writer.writeAll(input);
 
 PigTest t = newPigTest("pig/src/main/pig/wordcount_avro.pig", new String[] {"input=input.avro", "output=sorted.avro"});
 t.unoverride("STORE");
 t.runScript();
 
 //Check output
 BasicAvroReader<OutputRecord> reader = new BasicAvroReader<OutputRecord>(new Path(new File("sorted.avro").getAbsolutePath()), getFileSystem());
 List<OutputRecord> result = reader.readAll();
 Assert.assertEquals(result.size(), 2);
 Assert.assertEquals(result.get(0), OutputRecord.newBuilder().setWord("banana").setCount(2).build());
 Assert.assertEquals(result.get(1), OutputRecord.newBuilder().setWord("apple").setCount(1).build());
 }
Common pitfalls ▸ PigUnit ▸ Mocking capabilities are very limited ▸ Overhead of 1-5 seconds per script ▸ Cryptic error messages sometimes (NullPointerException) ▸ Pig UDFs ▸ Can be tested independently
Spark Testing
Spark Testing Base Base classes to use when writing tests with Spark ▸ https://github.com/holdenk/spark-testing-base ▸ Functionalities ▸ Provides SparkContext ▸ Utilities to compare RDDs and DataFrames ▸ Simulate how Streaming works ▸ Includes cool RDD and DataFrames generator
Thank You! We are hiring! http://careers.getyourguide.com/
Extra Resources ▸ https://github.com/miguno/avro-hadoop-starter ▸ http://www.michael-noll.com/blog/2013/07/04/using-avro-in-mapreduce-jobs-with-hadoop-pig-hive/ ▸ http://blog.cloudera.com/blog/2015/09/making-apache-spark-testing-easy-with-spark-testing-base/ ▸ http://www.slideshare.net/hkarau/effective-testing-for-spark-programs-strata-ny-2015 ▸ http://avro.apache.org/docs/current/ ▸ http://www.confluent.io/blog/schema-registry-kafka-stream-processing-yes-virginia-you-really-need-one ▸ http://mkuthan.github.io/blog/2015/03/01/spark-unit-testing/ ▸ https://engineering.linkedin.com/distributed-systems/log-what-every-software-engineer-should-know-about- real-time-datas-unifying

More Related Content

PDF
Test strategies for data processing pipelines
Lars Albertsson
 
PDF
Engineering data quality
Lars Albertsson
 
PDF
Effective testing for spark programs Strata NY 2015
Holden Karau
 
PDF
Testing data streaming applications
Lars Albertsson
 
PPTX
Introduction to Azure Databricks
James Serra
 
PDF
Building a Data Pipeline using Apache Airflow (on AWS / GCP)
Yohei Onishi
 
PDF
Productionzing ML Model Using MLflow Model Serving
Databricks
 
PDF
Integration between Filebeat and logstash
DaeMyung Kang
 
Test strategies for data processing pipelines
Lars Albertsson
 
Engineering data quality
Lars Albertsson
 
Effective testing for spark programs Strata NY 2015
Holden Karau
 
Testing data streaming applications
Lars Albertsson
 
Introduction to Azure Databricks
James Serra
 
Building a Data Pipeline using Apache Airflow (on AWS / GCP)
Yohei Onishi
 
Productionzing ML Model Using MLflow Model Serving
Databricks
 
Integration between Filebeat and logstash
DaeMyung Kang
 

What's hot (20)

PPTX
Dynamic Rule-based Real-time Market Data Alerts
Flink Forward
 
PPTX
A Beginner's Guide to Building Data Pipelines with Luigi
Growth Intelligence
 
PDF
DVC - Git-like Data Version Control for Machine Learning projects
Francesco Casalegno
 
PDF
Load Data Fast!
Karwin Software Solutions LLC
 
PDF
What is new in Apache Hive 3.0?
DataWorks Summit
 
PPTX
Databricks Fundamentals
Dalibor Wijas
 
PDF
Upgrade from MySQL 5.7 to MySQL 8.0
Olivier DASINI
 
PDF
Learn to Use Databricks for the Full ML Lifecycle
Databricks
 
PDF
MySQL InnoDB Cluster - A complete High Availability solution for MySQL
Olivier DASINI
 
PDF
Percona Live 2022 - MySQL Architectures
Frederic Descamps
 
PDF
From Query Plan to Query Performance: Supercharging your Apache Spark Queries...
Databricks
 
PDF
MLFlow: Platform for Complete Machine Learning Lifecycle
Databricks
 
PDF
Importance of ML Reproducibility & Applications with MLfLow
Databricks
 
PDF
TiDB for Big Data
PingCAP
 
PDF
Natural Language Processing with Graph Databases and Neo4j
William Lyon
 
PDF
The Full MySQL and MariaDB Parallel Replication Tutorial
Jean-François Gagné
 
PDF
Spark's Role in the Big Data Ecosystem (Spark Summit 2014)
Databricks
 
PPTX
Spark SQL Tutorial | Spark SQL Using Scala | Apache Spark Tutorial For Beginn...
Simplilearn
 
PPTX
Databricks on AWS.pptx
Wasm1953
 
PPTX
Building a modern data warehouse
James Serra
 
Dynamic Rule-based Real-time Market Data Alerts
Flink Forward
 
A Beginner's Guide to Building Data Pipelines with Luigi
Growth Intelligence
 
DVC - Git-like Data Version Control for Machine Learning projects
Francesco Casalegno
 
Load Data Fast!
Karwin Software Solutions LLC
 
What is new in Apache Hive 3.0?
DataWorks Summit
 
Databricks Fundamentals
Dalibor Wijas
 
Upgrade from MySQL 5.7 to MySQL 8.0
Olivier DASINI
 
Learn to Use Databricks for the Full ML Lifecycle
Databricks
 
MySQL InnoDB Cluster - A complete High Availability solution for MySQL
Olivier DASINI
 
Percona Live 2022 - MySQL Architectures
Frederic Descamps
 
From Query Plan to Query Performance: Supercharging your Apache Spark Queries...
Databricks
 
MLFlow: Platform for Complete Machine Learning Lifecycle
Databricks
 
Importance of ML Reproducibility & Applications with MLfLow
Databricks
 
TiDB for Big Data
PingCAP
 
Natural Language Processing with Graph Databases and Neo4j
William Lyon
 
The Full MySQL and MariaDB Parallel Replication Tutorial
Jean-François Gagné
 
Spark's Role in the Big Data Ecosystem (Spark Summit 2014)
Databricks
 
Spark SQL Tutorial | Spark SQL Using Scala | Apache Spark Tutorial For Beginn...
Simplilearn
 
Databricks on AWS.pptx
Wasm1953
 
Building a modern data warehouse
James Serra
 
Ad

Viewers also liked (6)

PDF
10 ways to stumble with big data
Lars Albertsson
 
PDF
Data pipelines from zero to solid
Lars Albertsson
 
PDF
Building Scalable Data Pipelines - 2016 DataPalooza Seattle
Evan Chan
 
PDF
Test strategies for data processing pipelines, v2.0
Lars Albertsson
 
PPTX
Building a unified data pipeline in Apache Spark
DataWorks Summit
 
PDF
Building a Data Pipeline from Scratch - Joe Crobak
Hakka Labs
 
10 ways to stumble with big data
Lars Albertsson
 
Data pipelines from zero to solid
Lars Albertsson
 
Building Scalable Data Pipelines - 2016 DataPalooza Seattle
Evan Chan
 
Test strategies for data processing pipelines, v2.0
Lars Albertsson
 
Building a unified data pipeline in Apache Spark
DataWorks Summit
 
Building a Data Pipeline from Scratch - Joe Crobak
Hakka Labs
 
Ad

Similar to The Mechanics of Testing Large Data Pipelines (QCon London 2016) (20)

PDF
The Mechanics of Testing Large Data Pipelines
C4Media
 
PPTX
Big Data Testing : Automate theTesting of Hadoop, NoSQL & DWH without Writing...
RTTS
 
PPT
Test Automation Best Practices (with SOA test approach)
Leonard Fingerman
 
PDF
From Relational Database Management to Big Data: Solutions for Data Migration...
Cognizant
 
PPTX
Query Wizards - data testing made easy - no programming
RTTS
 
PDF
StarWest 2019 - End to end testing: Stupid or Legit?
mabl
 
PPTX
Big Data Testing: Ensuring MongoDB Data Quality
RTTS
 
PPTX
Why You Need to STOP Using Spreadsheets for Audit Analysis
CaseWare IDEA
 
PDF
How to Automate your Enterprise Application / ERP Testing
RTTS
 
PPT
Data Collection Process And Integrity
Gerrit Klaschke, CSM
 
PPTX
Techniques for effective test data management in test automation.pptx
Knoldus Inc.
 
PDF
Knowledge discovery claudiad amato
SSSW
 
PDF
Data Driven Testing Is More Than an Excel File
Mehmet Gök
 
PDF
Tutorial Knowledge Discovery
SSSW
 
PDF
All You Need To Know About Big Data Testing - Bahaa Al Zubaidi.pdf
Bahaa Al Zubaidi
 
PPTX
Presentation Title
butest
 
PDF
Oleksii Pavlenko : Streamlining AI Integration: Organizing Teams and Processe...
Lviv Startup Club
 
PPTX
Data Warehouse Testing in the Pharmaceutical Industry
RTTS
 
PPT
Automated Testing with Databases
elliando dias
 
PDF
End to-end test automation at scale
mabl
 
The Mechanics of Testing Large Data Pipelines
C4Media
 
Big Data Testing : Automate theTesting of Hadoop, NoSQL & DWH without Writing...
RTTS
 
Test Automation Best Practices (with SOA test approach)
Leonard Fingerman
 
From Relational Database Management to Big Data: Solutions for Data Migration...
Cognizant
 
Query Wizards - data testing made easy - no programming
RTTS
 
StarWest 2019 - End to end testing: Stupid or Legit?
mabl
 
Big Data Testing: Ensuring MongoDB Data Quality
RTTS
 
Why You Need to STOP Using Spreadsheets for Audit Analysis
CaseWare IDEA
 
How to Automate your Enterprise Application / ERP Testing
RTTS
 
Data Collection Process And Integrity
Gerrit Klaschke, CSM
 
Techniques for effective test data management in test automation.pptx
Knoldus Inc.
 
Knowledge discovery claudiad amato
SSSW
 
Data Driven Testing Is More Than an Excel File
Mehmet Gök
 
Tutorial Knowledge Discovery
SSSW
 
All You Need To Know About Big Data Testing - Bahaa Al Zubaidi.pdf
Bahaa Al Zubaidi
 
Presentation Title
butest
 
Oleksii Pavlenko : Streamlining AI Integration: Organizing Teams and Processe...
Lviv Startup Club
 
Data Warehouse Testing in the Pharmaceutical Industry
RTTS
 
Automated Testing with Databases
elliando dias
 
End to-end test automation at scale
mabl
 

Recently uploaded (20)

PDF
7 ChatGPT Prompts to Help You Define Your Ideal Customer Profile.pdf
SOFTTECHHUB
 
PDF
Electronic commerce courselecture one. Pdf
OmerMohamed64
 
PDF
Bridging biosciences and deep learning for revolutionary discoveries: a compr...
IAESIJAI
 
PDF
TokAI - TikTok AI Agent : The First AI Application That Analyzes 10,000+ Vira...
SOFTTECHHUB
 
PDF
Dropbox Q2 2025 Financial Results & Investor Presentation
Dropbox
 
PDF
Blue Purple Modern Animated Computer Science Presentation.pdf.pdf
Akar16
 
PDF
Encapsulation theory and applications.pdf
gurumoop
 
PPTX
Understanding_Digital_Forensics_Presentation.pptx
ImranKhan423233
 
PDF
Approach and Philosophy of On baking technology
30anthuong17
 
PPTX
Cloud computing and distributed systems.
kkkkkhan
 
PDF
Agricultural_Statistics_at_a_Glance_2022_0.pdf
SandeepSingh286037
 
PDF
Per capita expenditure prediction using model stacking based on satellite ima...
IAESIJAI
 
PDF
cuic standard and advanced reporting.pdf
SimoneTitaniumTomase
 
PPTX
VMware vSphere Foundation How to Sell Presentation-Ver1.4-2-14-2024.pptx
blackmambaettijean
 
PDF
Network Security Unit 5.pdf for BCA BBA.
Serpent6
 
PPTX
Detection-First SIEM: Rule Types, Dashboards, and Threat-Informed Strategy
ReZa AdineH
 
DOCX
The AUB Centre for AI in Media Proposal.docx
GAONE9
 
PDF
Advanced methodologies resolving dimensionality complications for autism neur...
IAESIJAI
 
PPTX
MYSQL Presentation for SQL database connectivity
Swati270511
 
PDF
Review of recent advances in non-invasive hemoglobin estimation
IAESIJAI
 
7 ChatGPT Prompts to Help You Define Your Ideal Customer Profile.pdf
SOFTTECHHUB
 
Electronic commerce courselecture one. Pdf
OmerMohamed64
 
Bridging biosciences and deep learning for revolutionary discoveries: a compr...
IAESIJAI
 
TokAI - TikTok AI Agent : The First AI Application That Analyzes 10,000+ Vira...
SOFTTECHHUB
 
Dropbox Q2 2025 Financial Results & Investor Presentation
Dropbox
 
Blue Purple Modern Animated Computer Science Presentation.pdf.pdf
Akar16
 
Encapsulation theory and applications.pdf
gurumoop
 
Understanding_Digital_Forensics_Presentation.pptx
ImranKhan423233
 
Approach and Philosophy of On baking technology
30anthuong17
 
Cloud computing and distributed systems.
kkkkkhan
 
Agricultural_Statistics_at_a_Glance_2022_0.pdf
SandeepSingh286037
 
Per capita expenditure prediction using model stacking based on satellite ima...
IAESIJAI
 
cuic standard and advanced reporting.pdf
SimoneTitaniumTomase
 
VMware vSphere Foundation How to Sell Presentation-Ver1.4-2-14-2024.pptx
blackmambaettijean
 
Network Security Unit 5.pdf for BCA BBA.
Serpent6
 
Detection-First SIEM: Rule Types, Dashboards, and Threat-Informed Strategy
ReZa AdineH
 
The AUB Centre for AI in Media Proposal.docx
GAONE9
 
Advanced methodologies resolving dimensionality complications for autism neur...
IAESIJAI
 
MYSQL Presentation for SQL database connectivity
Swati270511
 
Review of recent advances in non-invasive hemoglobin estimation
IAESIJAI
 

The Mechanics of Testing Large Data Pipelines (QCon London 2016)

  • 1. THE MECHANICS OF TESTING LARGE DATA PIPELINES MATHIEU BASTIAN Head of Data Engineering, GetYourGuide @mathieubastian www.linkedin.com/in/mathieubastian QCon London 2015
  • 2. Outline ▸ Motivating example ▸ Challenges ▸ Testing strategies ▸ Validation Strategies ▸ Tools Integration Tests ArchitectureUnit Test
  • 3. Data Pipelines often start simple
  • 4. Users E-commerce website Search App Views Offline Dashboard Search Metrics Views HDFS They have one use-case and one developer
  • 5. But there are many other use- cases Recommender Systems Anomaly Detection Search Ranking A/B Testing Spam Detection Sentiment Analysis Topic Detection Trending TagsQuery Expansion Customer Churn Prediction Related searches Fraud Prediction Bidding Prediction Machine Translation Signal Processing Content Curation Sentiment Analysis Image recognition Optimal pricing Location normalization Standardization Funnel Analysis
  • 6. additional events and logsDevelopers add Users E-commerce website Search App Clicks Views Offline Dashboard Search Metrics Clicks Views HDFS
  • 7. third-party dataDevelopers add Users E-commerce website 3rd parties Search App Clicks Views A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views A/B Logs HDFS
  • 8. search ranking predictionDevelopers add Users E-commerce website 3rd parties Search App Clicks Views A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views Training data Training & validation Model Clicks Views Features transformation A/B Logs HDFS
  • 9. personalized user featuresDevelopers add Users E-commerce website 3rd parties Search App Clicks Views ProfilesUser Database A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views Training data Training & validation Model Clicks Views Profiles Features transformation A/B Logs HDFS
  • 10. query extensionDevelopers add Users E-commerce website 3rd parties Search App Clicks Views ProfilesUser Database A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views Training data Training & validation Model Clicks Views Profiles Features transformation A/B Logs Filter queries Query extension RDBMS Views Training data HDFS
  • 11. Developers add recommender system Users E-commerce website 3rd parties Search App Clicks Views ProfilesUser Database A/B LogsMobile Analytics Offline Dashboard Search Metrics Clicks Views Training data Training & validation Model Clicks Views Profiles Features Features transformation Features NoSQL Compute recommendations A/B Logs Filter queries Query extension RDBMS Views Training data HDFS
  • 12. Data Pipelines can grow very large
  • 13. That is a lot of code and data
  • 14. Code contain bugs Industry Average: about 15 - 50 errors per 1000 lines of delivered code.
  • 16. Embrace automated testing of code validation of data
  • 17. Because it delivers ▸ Testing ▸ Tested code has less bugs ▸ Gives the confidence to iterate quickly ▸ Scales well to multiple developers ▸ Validation ▸ Reduce manual testing ▸ Avoid catastrophic failures
  • 18. But it’s challenging ▸ Testing ▸ Need data to test "realistically" ▸ Not running locally, can be expensive ▸ Tooling weaknesses ▸ Validation ▸ Data sources out of our control ▸ Difficult to test machine learning models
  • 20. Manual testing Waiting Coding Looking at logs Code Upload Run workflow Look at logs ▸ Time Spent
  • 22. Prepare environment ▸ Care about tests from the start of your project ▸ All jobs should be functions (output only depends on input) ▸ Safe to re-run the job ▸ Does the input data still exists? ▸ Would it push partial results? ▸ Centralize configurations and no hard-coded paths ▸ Version code and timestamp data
  • 23. Unit test locally ▸ Test locally each individual job ▸ Tests its good code ▸ Tests expected failures ▸ Need to overcome challenges with fake data creation ▸ Complex structures and numerous data sources ▸ Too small to be meaningful ▸ Need to specify a different configuration
  • 24. Build from schemas Fake data creation based on schemas. Compare: Customer c = Customer.newBuilder().
 setId(42).
 setInterests(Arrays.asList(new Interest[]{
 Interest.newBuilder().setId(0).setName("Ping-Pong").build()
 Interest.newBuilder().setId(1).setName(“Pizza").build()})) .build(); vs Map<String, Object> c = new HashMap<>();
 c.put("id", 42);
 Map<String, Object> i1 = new HashMap<>();
 i1.put("id", 0);
 i1.put("name", "Ping-Pong");
 Map<String, Object> i2 = new HashMap<>();
 i2.put("id", 1);
 i2.put("name", "Pizza");
 c.put("interests", Arrays.asList(new Map[] {i1, i2}));
  • 25. Build from schemas Avro Schema example { "type": "record", "name": "Customer", "fields": [{ "name": "id", "type": "int" }, { "name": "interests", "type": { "type": "array", "items": { "name": "Interest", "type": "record", "fields": [{ "name": "id", "type": "int" }, { "name": "name", "type": ["string", "null"] }] } } } ] } nullable field
  • 26. Complex generators ▸ Developed in the field of property-based testing //Small Even Number Generator val smallEvenInteger = Gen.choose(0,200) suchThat (_ % 2 == 0) ▸ Goal is to simulate, not sample real data ▸ Define complex random generators that match properties (e.g. frequency) ▸ Can go beyond unit-testing and generate complex domain models ▸ https://www.scalacheck.org/ for Scala/Java is a good starting point for examples
  • 27. Integration test on sample data ▸ Integration test the entire workflow ▸ File paths ▸ Configuration ▸ Evaluate performance ▸ Sample data ▸ Large enough to be meaningful ▸ Small enough to speed-up testing JOB A JOB B JOB C JOB D
  • 29. Where it fail Control Difficulty Model biases Bug Noisy data Schema changes Missing data
  • 30. Input and output validation Make the pipeline robust by validating inputs and outputs Input Input Input Workflow Production ValidationValidation
  • 32. Input data validation Input data validation is a key component of pipeline robustness. The goal is to test the entry points of our system for data quality. ETL RDBMS NOSQL EVENTS TWITTER DATA PIPELINE
  • 33. Why it matters ▸ Bad input data will most likely degrade the output ▸ It likely will fail silently ▸ Because data will change ▸ Data migrations: maintenance, cluster update, new infrastructure ▸ Events change due to product evolution ▸ Data dependencies updated
  • 34. Input data validation ▸ Validation code should ▸ Detect pathological data and fail early ▸ Deal with expected data variability ▸ Example issues: ▸ Missing values, encoding issues, etc. ▸ Schema changes ▸ Duplicates rows ▸ Data order changes
  • 35. Pathological data ▸ Value ▸ Validity depends on a single, independent value. ▸ Easy to validate on streams of data ▸ Dataset ▸ Validity depends on the entire dataset ▸ More difficult to validate as it needs a window of data
  • 36. Metadata validation Analyzing metadata is the quickest way to validate input data ▸ Number of records and file sizes ▸ Hadoop/Spark counters ▸ Number of map/reduce records, size ▸ Record-level custom counters ▸ Average text length ▸ Task-level custom counters ▸ Min/Max/Median values
  • 37. Hadoop/Spark counters Results can be accessed programmatically and checked
  • 38. Control inputs with Schemas ▸ CSVs aren’t robust to change, use Schemas ▸ Makes expected data explicit and easy to test against ▸ Gives basic validation for free with binary serialization (e.g. Avro, Thrift, Protocol Buffer) ▸ Typed (integer, boolean, lists etc.) ▸ Specify if value is optional ▸ Schemas can be evolved without breaking compatibility
  • 40. Why it matters ▸ Humans makes mistake, we need a safeguard ▸ Rolling back data is often complex ▸ Bad output propagates to downstream systems Example with a recommender system // One recommendation set per user { "userId": 42, "recommendations": [{ "itemId": 1456, "score": 0.9 }, { "itemId": 4232, "score": 0.1 }], "model": "test01" }
  • 41. Check for anomalies Simple strategies similar to input data validation ▸ Record level (e.g. values within bounds) ▸ Dataset level (e.g. counts, order) Challenges around relevance evaluation ▸ When supervised, use a validation dataset and threshold accuracy ▸ Introduce hypothetical examples
  • 42. Incremental update as validation Join with the previous “best" output ▸ Allows fine comparisons ▸ Incremental framework can be extended to ▸ Only recompute recommendations that have changed ▸ Produce variations metric between different models Daily Recommendations Compute daily recommendationsHDFS Recommendations Yesterday Recommendations Join with previous result
  • 43. External validation Even in automated environment it is possible to validate with humans ▸ Example: Search ranking evaluation ▸ Solution: Crowdsourcing ▸ Complex validation that requires training ▸ Can be automated through APIs
  • 44. Mitigate risk with A/B testing Gradually rolling out data products improvements reduces the need for complex output validation ▸ Experiment can be controlled online or offline ▸ Online: Push multiple set of recommendations (1 per model) ▸ Offline: Split users and push unique set of recommendations userId -> [{
 "model": "test01",
 "recommendations": [{...}]
 }, {
 "model": "test02",
 "recommendations": [{...}]
 }] A B userId -> {
 "model": "test01",
 "recommendations": [{...}]
 }
  • 45. Mitigate risk with A/B testing Important ▸ Log model variation downstream in logs ▸ Encapsulate model logic FEATURE 1-A MODEL A MODEL B FEATURE 1 FEATURE 2 MODEL A MODEL B A BA B FEATURE 1-B FEATURE 2-A FEATURE 2-B
  • 48. Two ways to test Hadoop jobs ▸ MRUnit ▸ Java library to test MapReduce jobs in a simulated environment ▸ Last release June 2014 ▸ MiniCluster ▸ Utility to locally run a fully-functional Hadoop cluster in a test environment ▸ Ships with Hadoop itself
  • 49. MiniMRCluster ▸ Advantages ▸ Behaves like a real cluster, including setup and configuration ▸ Can be used to test multiple jobs (integration testing) ▸ Disadvantages ▸ Very slow compared to unit testing Java code
  • 50. MRUnit ▸ Advantages ▸ Faster ▸ Less boilerplate code ▸ Disadvantages ▸ Need to replicate job configuration ▸ Only built to test map and reduce functions ▸ Difficult to make it work with custom input formats (e.g. Avro)
  • 51. MiniMRCluster setup* Setup MR cluster and obtain FileSystem @BeforeClass
 public void setup() {
 Configuration dfsConf = new Configuration();
 dfsConf.set(MiniDFSCluster.HDFS_MINIDFS_BASEDIR, new File("./target/ hdfs/").getAbsolutePath());
 _dfsCluster = new MiniDFSCluster.Builder(dfsConf).numDataNodes(1).build();
 _dfsCluster.waitClusterUp();
 _fileSystem = _dfsCluster.getFileSystem();
 
 YarnConfiguration yarnConf = new YarnConfiguration();
 yarnConf.setFloat(YarnConfiguration.NM_MAX_PER_DISK_UTILIZATION_PERCENTAGE, 99.0f);
 yarnConf.setInt(YarnConfiguration.RM_SCHEDULER_MINIMUM_ALLOCATION_MB, 64);
 yarnConf.setClass(YarnConfiguration.RM_SCHEDULER, FifoScheduler.class, ResourceScheduler.class);
 _mrCluster = new MiniMRYarnCluster(getClass().getName(), taskTrackers);
 yarnConf.set("fs.defaultFS", _fileSystem.getUri().toString());
 _mrCluster.init(yarnConf);
 _mrCluster.start();
 } * Hadoop version used 2.7.2
  • 52. Keep the test file clean of boilerplate code Best is to wrap the start/stop code into a TestBase class /**
 * Default constructor with one task tracker and one node.
 */
 public TestBase() { ... }
 
 @BeforeClass
 public void startCluster() throws IOException { ... }
 
 @AfterClass
 public void stopCluster() throws IOException { ... }
 
 /**
 * Returns the Filesystem in use.
 *
 * @return the filesystem used by Hadoop.
 */
 protected FileSystem getFileSystem() {
 return _fileSystem;
 }
  • 53. Initialize and clean HDFS before/after each test Clean up and initialize file system before each test private final Path _inputPath = new Path("/input");
 private final Path _cachePath = new Path("/cache");
 private final Path _outputPath = new Path("/output");
 
 @BeforeMethod
 public void beforeMethod(Method method) throws IOException {
 getFileSystem().delete(_inputPath, true);
 getFileSystem().mkdirs(_inputPath);
 getFileSystem().delete(_cachePath, true);
 }
 
 @AfterMethod
 public void afterMethod(Method method) throws IOException {
 getFileSystem().delete(_inputPath, true);
 getFileSystem().delete(_cachePath, true);
 getFileSystem().delete(_outputPath, true);
 }
  • 54. Run MiniCluster Test Clean up and initialize file system before each test @Test
 public void testBasicWordCountJob() throws IOException, InterruptedException, ClassNotFoundException {
 writeWordCountInput();
 configureAndRunJob(new BasicWordCountJob(), "BasicWordCountJob", _inputPath, _outputPath);
 checkWordCountOutput();
 } private void configureAndRunJob(AbstractJob job, String name, Path inputPath, Path outputPath) throws IOException, ClassNotFoundException, InterruptedException {
 Properties _props = new Properties();
 _props.setProperty("input.path", inputPath.toString());
 _props.setProperty("output.path", outputPath.toString());
 job.setProperties(_props);
 job.setName(name);
 job.run();
 }
  • 55. MRUnit setup Setup MapDriver and ReduceDriver BasicWordCountJob.Map mapper;
 BasicWordCountJob.Reduce reducer;
 MapDriver<LongWritable, Text, Text, IntWritable> mapDriver;
 ReduceDriver<Text, IntWritable, Text, IntWritable> reduceDriver;
 
 @BeforeClass
 public void setup() {
 mapper = new BasicWordCountJob.Map();
 mapDriver = MapDriver.newMapDriver(mapper);
 reducer = new BasicWordCountJob.Reduce();
 reduceDriver = ReduceDriver.newReduceDriver(reducer);
 }
  • 56. Run MRUnit test Set Input/Output and run Test @Test
 public void testMapper() throws IOException {
 mapDriver.withInput(new LongWritable(0), new Text("banana pear banana"));
 mapDriver.withOutput(new Text("banana"), new IntWritable(1));
 mapDriver.withOutput(new Text("pear"), new IntWritable(1));
 mapDriver.withOutput(new Text("banana"), new IntWritable(1));
 mapDriver.runTest();
 }
 
 @Test
 public void testReducer() throws IOException {
 reduceDriver.withInput(new Text("banana"), Arrays.asList(new IntWritable(1), new IntWritable(1)));
 reduceDriver.withInput(new Text("pear"), Arrays.asList(new IntWritable(1)));
 reduceDriver.withOutput(new Text("banana"), new IntWritable(2));
 reduceDriver.withOutput(new Text("pear"), new IntWritable(1));
 reduceDriver.runTest();
 }
  • 57. Most common pitfall ▸ With both MiniMRCluster and MRUnit one spend most of the time ▸ Creating fake input data ▸ Verifying output data ▸ Solutions ▸ Use rich data structures format (e.g. Avro, Thrift) ▸ Use automated Java classes generation
  • 58. Other common pitfalls ▸ MiniMRCluster ▸ Enable Hadoop INFO logging so you can see real job failure causes ▸ Beware of partitioning or sorting issues unrevealed when testing with too few rows and number of nodes ▸ The API has changed over the years, difficult to find examples ▸ MRUnit ▸ Custom serialization issues (e.g. Avro, Thrift)
  • 60. Introducing PigUnit ▸ PigUnit ▸ Official library to unit tests Pig script ▸ Ships with Pig (latest version 0.15.0) ▸ The principle is easy 1. Generate test data 2. Run script with PigUnit 3. Verify output ▸ Runs locally but can be run on a cluster too
  • 61. Script example WordCount example ‣ Input and output are standard formats ‣ Uses variables $input and $output text = LOAD '$input' USING TextLoader();
 
 flattened = FOREACH text GENERATE flatten(TOKENIZE((chararray)$0)) as word;
 grouped = GROUP flattened by word;
 result = FOREACH grouped GENERATE group, (int)COUNT($1) AS cnt;
 sorted = ORDER result BY cnt DESC;
 
 STORE sorted INTO '$output' USING PigStorage('t');
  • 62. PigTestBase Create PigTest object protected final FileSystem _fileSystem;
 
 protected PigTestBase() {
 System.setProperty("udf.import.list", StringUtils.join(Arrays.asList("oink.", "org.apache.pig.piggybank."), ":"));
 fileSystem = FileSystem.get(new Configuration());
 }
 
 /**
 * Creates a new <em>PigTest</em> instance ready to be used.
 *
 * @param scriptFile the path to the Pig script file
 * @param inputs the Pig arguments
 * @return new PigTest instance
 */
 protected PigTest newPigTest(String scriptFile, String[] inputs) {
 PigServer pigServer = new PigServer(ExecType.LOCAL);
 Cluster pigCluster = new Cluster(pigServer.getPigContext());
 return new PigTest(scriptFile, inputs, pigServer, pigCluster);
 }
  • 63. Test using aliases getAlias() allows to obtain the data anywhere in the script @Test
 public void testWordCountAlias() throws IOException, ParseException {
 //Write input data
 BufferedWriter writer = new BufferedWriter(new FileWriter(new File("input.txt")));
 writer.write("banana pear banana");
 writer.close();
 
 PigTest t = newPigTest("pig/src/main/pig/wordcount_text.pig", new String[] {"input=input.txt", "output=result.csv"});
 
 Iterator<Tuple> tuples = t.getAlias("sorted");
 Assert.assertTrue(tuples.hasNext());
 Tuple tuple = tuples.next();
 Assert.assertEquals(tuple.get(0), "banana");
 Assert.assertEquals(tuple.get(1), 2);
 Assert.assertTrue(tuples.hasNext());
 tuple = tuples.next();
 Assert.assertEquals(tuple.get(0), "pear");
 Assert.assertEquals(tuple.get(1), 1);
 }
  • 64. Test using mock and assert ▸ mockAlias allows to substitute input data ▸ assertOutput allows to compare String output data @Test
 public void testWordCountMock() throws IOException, ParseException {
 //Write input data
 BufferedWriter writer = new BufferedWriter(new FileWriter(new File("input.txt")));
 writer.write("banana pear banana");
 writer.close();
 
 PigTest t = newPigTest("pig/src/main/pig/wordcount_text.pig", new String[] {"input=input.txt", "output=null"});
 t.runScript();
 t.assertOutputAnyOrder("sorted", new String[]{"(banana,2)", "(pear,1)"});
 }
  • 65. Both of these tools have limitations ▸ Built around standard input and output (Text, CSVs etc.) ▸ Realistically most of our data is in other formats (e.g. Avro, Thrift, JSON) ▸ Does not test the STORE function (e.g. schema errors) ▸ getAlias() is especially difficult to use ▸ Need to remember field position: tuple.get(0) ▸ assertOutput() only allows String comparison ▸ Cumbersome to write complex structures (e.g. bags of bags)
  • 66. Example with Avro input/output ▸ Focus on testing script’s output ▸ Difficulty is to generate dummy Avro data and compare result text = LOAD '$input' USING AvroStorage();
 
 flattened = FOREACH text GENERATE flatten(TOKENIZE(body)) as word;
 grouped = GROUP flattened by word;
 result = FOREACH grouped GENERATE group AS word, (int)COUNT($1) AS cnt;
 sorted = ORDER result BY cnt DESC;
 
 STORE result INTO '$output' USING AvroStorage(); ▸ By default, PigUnit doesn’t execute the STORE, but it can be overridden pigTest.unoverride("STORE");
  • 67. Simple utility classes for Avro ▸ BasicAvroWriter ▸ Writes Avro file on disk based on a schema ▸ Supports GenericRecord and SpecificRecord ▸ BasicAvroReader ▸ Reads Avro file, the schema heads the file ▸ Also supports GenericRecord and SpecificRecord
  • 68. Test with Avro GenericRecord ▸ Create Schema with SchemaBuilder, write data, run script, read result and compare @Test
 public void testWordCountGenericRecord() throws IOException, ParseException {
 Schema schema = SchemaBuilder.builder().record("record").fields().
 name("text").type().stringType().noDefault().endRecord();
 GenericRecord genericRecord = new GenericData.Record(schema);
 genericRecord.put("text", "banana apple banana");
 
 BasicAvroWriter writer = new BasicAvroWriter(new Path(new File("input.avro").getAbsolutePath()), schema, getFileSystem());
 writer.append(genericRecord);
 
 PigTest t = newPigTest("pig/src/main/pig/wordcount_avro.pig", new String[] {"input=input.avro", "output=sorted.avro"});
 t.unoverride("STORE");
 t.runScript();
 
 //Check output
 BasicAvroReader reader = new BasicAvroReader(new Path(new File("sorted.avro").getAbsolutePath()), getFileSystem());
 Map<Utf8, GenericRecord> result = reader.readAndMapAll("word");
 Assert.assertEquals(result.size(), 2);
 Assert.assertEquals(result.get(new Utf8("banana")).get("cnt"), 2);
 Assert.assertEquals(result.get(new Utf8("apple")).get("cnt"), 1);
 }
  • 69. Test with Avro SpecificRecord ▸ Use InputRecord and OutputRecord generated Java classes, write data, run script, read result and compare @Test
 public void testWordCountSpecificRecord() throws IOException, ParseException {
 InputRecord input = InputRecord.newBuilder().setText("banana apple banana").build();
 
 BasicAvroWriter<InputRecord> writer = new BasicAvroWriter<InputRecord>(new Path(new File("input.avro").getAbsolutePath()), input.getSchema(), getFileSystem());
 writer.writeAll(input);
 
 PigTest t = newPigTest("pig/src/main/pig/wordcount_avro.pig", new String[] {"input=input.avro", "output=sorted.avro"});
 t.unoverride("STORE");
 t.runScript();
 
 //Check output
 BasicAvroReader<OutputRecord> reader = new BasicAvroReader<OutputRecord>(new Path(new File("sorted.avro").getAbsolutePath()), getFileSystem());
 List<OutputRecord> result = reader.readAll();
 Assert.assertEquals(result.size(), 2);
 Assert.assertEquals(result.get(0), OutputRecord.newBuilder().setWord("banana").setCount(2).build());
 Assert.assertEquals(result.get(1), OutputRecord.newBuilder().setWord("apple").setCount(1).build());
 }
  • 70. Common pitfalls ▸ PigUnit ▸ Mocking capabilities are very limited ▸ Overhead of 1-5 seconds per script ▸ Cryptic error messages sometimes (NullPointerException) ▸ Pig UDFs ▸ Can be tested independently
  • 72. Spark Testing Base Base classes to use when writing tests with Spark ▸ https://github.com/holdenk/spark-testing-base ▸ Functionalities ▸ Provides SparkContext ▸ Utilities to compare RDDs and DataFrames ▸ Simulate how Streaming works ▸ Includes cool RDD and DataFrames generator
  • 73. Thank You! We are hiring! http://careers.getyourguide.com/
  • 74. Extra Resources ▸ https://github.com/miguno/avro-hadoop-starter ▸ http://www.michael-noll.com/blog/2013/07/04/using-avro-in-mapreduce-jobs-with-hadoop-pig-hive/ ▸ http://blog.cloudera.com/blog/2015/09/making-apache-spark-testing-easy-with-spark-testing-base/ ▸ http://www.slideshare.net/hkarau/effective-testing-for-spark-programs-strata-ny-2015 ▸ http://avro.apache.org/docs/current/ ▸ http://www.confluent.io/blog/schema-registry-kafka-stream-processing-yes-virginia-you-really-need-one ▸ http://mkuthan.github.io/blog/2015/03/01/spark-unit-testing/ ▸ https://engineering.linkedin.com/distributed-systems/log-what-every-software-engineer-should-know-about- real-time-datas-unifying