Building Robust ETL Pipelines with Apache Spark

Building Robust ETL
Pipelines with Apache Spark
Xiao Li
Spark Summit | SF | Jun 2017

2
TEAM
About Databricks
Started Spark project (now Apache Spark) at UC Berkeley in 2009
22
PRODUCT
Unified Analytics Platform
MISSION
Making Big Data Simple

3
About Me
• Apache Spark Committer
• Software Engineer at Databricks
• Ph.D. in University of Florida
• Previously, IBM Master Inventor, QRep, GDPS A/A and STC
• Spark SQL, Database Replication, Information Integration
• Github: gatorsmile

4
Overview
1. What’s an ETL Pipeline?
2. Using Spark SQL for ETL
- Extract: Dealing with Dirty Data (Bad Records or Files)
- Extract: Multi-line JSON/CSV Support
- Transformation: High-order functions in SQL
- Load: Unified write paths and interfaces
3. New Features in Spark 2.3
- Performance (Data Source API v2, Python UDF)

5
What is a Data Pipeline?
1. Sequence of transformations on data
2. Source data is typically semi-structured/unstructured
(JSON, CSV etc.) and structured (JDBC, Parquet, ORC, the
other Hive-serde tables)
3. Output data is integrated, structured and curated.
– Ready for further data processing, analysis and reporting

6
Example of a Data Pipeline
Aggregate Reporting
Applications
ML
Model
Ad-hoc Queries
Database
Cloud
Warehouse
Kafka, Log
Kafka, Log

7
ETL is the First Step in a Data Pipeline
1. ETL stands for EXTRACT, TRANSFORM and LOAD
2. Goal is to clean or curate the data
- Retrieve data from sources (EXTRACT)
- Transform data into a consumable format (TRANSFORM)
- Transmit data to downstream consumers (LOAD)

8
An ETL Query in Apache Spark
spark.read.json("/source/path")
.filter(...)
.agg(...)
.write.mode("append")
.parquet("/output/path")
EXTRACT
TRANSFORM
LOAD

9
An ETL Query in Apache Spark
Extract
EXTRACT
TRANSFORM
LOAD
val csvTable = spark.read.csv("/source/path")
val jdbcTable = spark.read.format("jdbc")
.option("url", "jdbc:postgresql:...")
.option("dbtable", "TEST.PEOPLE")
.load()
csvTable
.join(jdbcTable, Seq("name"), "outer")
.filter("id <= 2999")
.write
.mode("overwrite")
.format("parquet")
.saveAsTable("outputTableName")

10
What’s so hard about ETL
Queries?

11
Why is ETL Hard?
1. Too complex
2. Error-prone
3. Too slow
4. Too expensive
1. Various sources/formats
2. Schema mismatch
3. Different representation
4. Corrupted files and data
5. Scalability
6. Schema evolution
7. Continuous ETL

12
This is why ETL is important
Consumers of this data don’t want to deal with this
messiness and complexity

14
Structured
Streaming
Spark SQL's flexible APIs,
support for a wide
variety of datasources,
build-in support for
structured streaming,
state of art catalyst
optimizer and tungsten
execution engine make it
a great framework for
building end-to-end ETL
pipelines.

15
Data Source Supports
1. Built-in connectors in Spark:
– JSON, CSV, Text, Hive, Parquet, ORC, JDBC
2. Third-party data source connectors:
– https://spark-packages.org
3. Define your own data source connectors by
Data Source APIs
– Ref link: https://youtu.be/uxuLRiNoDio

16
{"a":1, "b":2, "c":3}
{"e":2, "c":3, "b":5}
{"a":5, "d":7}
spark.read
.json("/source/path”)
.printSchema()
Schema Inference – semi-structured files

17
{"a":1, "b":2, "c":3.1}
{"e":2, "c":3, "b":5}
{"a":"5", "d":7}
spark.read
.json("/source/path”)
.printSchema()
Schema Inference – semi-structured files

18
{"a":1, "b":2, "c":3}
{"e":2, "c":3, "b":5}
{"a":5, "d":7}
val schema = new StructType()
.add("a", "int")
.add("b", "int")
spark.read
.json("/source/path")
.schema(schema)
.show()
User-specified Schema

19
{"a":1, "b":2, "c":3}
{"e":2, "c":3, "b":5}
{"a":5, "d":7}
Availability: Apache Spark 2.2
spark.read
.json("/source/path")
.schema("a INT, b INT")
.show()
User-specified DDL-format Schema

20
Corrupt
Files
java.io.IOException. For example, java.io.EOFException: Unexpected end of input
stream at org.apache.hadoop.io.compress.DecompressorStream.decompress
java.lang.RuntimeException: file:/temp/path/c000.json is not a Parquet file (too
small)
spark.sql.files.ignoreCorruptFiles = true
[SPARK-17850] If true, the Spark jobs will
continue to run even when it encounters
corrupt files. The contents that have
been read will still be returned.
Dealing with Bad Data: Skip Corrupt Files

21
Missing or
Corrupt
Records
[SPARK-12833][SPARK-
13764] TextFile formats
(JSON and CSV) support
3 different ParseModes
while reading data:
1. PERMISSIVE
2. DROPMALFORMED
3. FAILFAST
Dealing with Bad Data: Skip Corrupt Records

22
{"a":1, "b":2, "c":3}
{"a":{, b:3}
{"a":5, "b":6, "c":7}
spark.read
.option("mode", "PERMISSIVE")
.option("columnNameOfCorruptRecord", "_corrupt_record")
.json(corruptRecords)
.show() The default can be configured via
spark.sql.columnNameOfCorruptRecord
Json: Dealing with Corrupt Records

23
{"a":1, "b":2, "c":3}
{"a":{, b:3}
{"a":5, "b":6, "c":7}
spark.read
.option("mode", "DROPMALFORMED")
.show()

24
{"a":1, "b":2, "c":3}
{"a":{, b:3}
{"a":5, "b":6, "c":7}
spark.read
.option("mode", "FAILFAST")
.show()
org.apache.spark.sql.catalyst.json
.SparkSQLJsonProcessingException:
Malformed line in FAILFAST mode:
{"a":{, b:3}

25
spark.read
.option("mode", "FAILFAST")
.csv(corruptRecords)
.show()
java.lang.RuntimeException:
Malformed line in FAILFAST mode:
2015,Chevy,Volt
CSV: Dealing with Corrupt Records
year,make,model,comment,blank
"2012","Tesla","S","No comment",
1997,Ford,E350,"Go get one now they",
2015,Chevy,Volt

26
spark.read.
.show()
2015,Chevy,Volt

27
2015,Chevy,Volt
spark.read
.option("header", true)
.show()

28
val schema = "col1 INT, col2 STRING, col3 STRING, col4 STRING, " +
"col5 STRING, __corrupted_column_name STRING"
spark.read
.option("header", true)
.show()

29
2015,Chevy,Volt
spark.read
.option("mode", ”DROPMALFORMED")
.show()

30
Functionality: Better Corruption Handling
badRecordsPath: a user-specified path to store exception files for
recording the information about bad records/files.
- A unified interface for both corrupt records and files
- Enabling multi-phase data cleaning
- DROPMALFORMED + Exception files
- No need an extra column for corrupt records
- Recording the exception data, reasons and time.
Availability: Databricks Runtime 3.0

31
Functionality: Better JSON and CSV Support
[SPARK-18352] [SPARK-19610] Multi-line JSON and CSV Support
- Spark SQL currently reads JSON/CSV one line at a time
- Before 2.2, it requires custom ETL
spark.read
.option(”multiLine",true)
.json(path)
spark.read
.option(”multiLine",true)
.json(path)

32
Transformation: Higher-order Function in SQL
Transformation on complex objects like arrays, maps and
structures inside of columns.
UDF ? Expensive data serialization
tbl_nested
|-- key: long (nullable = false)
|-- values: array (nullable = false)
| |-- element: long (containsNull = false)

33
Transformation: Higher order function in SQL
1) Check for element existence
SELECT EXISTS(values, e -> e > 30) AS v
FROM tbl_nested;
2) Transform an array
SELECT TRANSFORM(values, e -> e * e) AS v
FROM tbl_nested;
tbl_nested
Transformation on complex objects like arrays, maps and
structures inside of columns.

34
4) Aggregate an array
SELECT REDUCE(values, 0, (value, acc) -> value + acc) AS sum
FROM tbl_nested;
Ref Databricks Blog: http://dbricks.co/2rUKQ1A
More cool features available in DB Runtime 3.0: http://dbricks.co/2rhPM4c
Availability: Databricks Runtime 3.0
3) Filter an array
SELECT FILTER(values, e -> e > 30) AS v
FROM tbl_nested;
Transformation: Higher order function in SQL
tbl_nested

35
Users can create Hive-serde tables using
DataframeWriter APIs
New Format in DataframeWriter API
df.write.format("parquet")
.saveAsTable("tab")
df.write.format("hive")
.option("fileFormat", "avro")
.saveAsTable("tab")
CREATE Hive-serde tables CREATE data source tables

36
Unified CREATE TABLE [AS SELECT]
CREATE TABLE t1(a INT, b INT)
USING ORC
USING hive
OPTIONS(fileFormat 'ORC')
CREATE Hive-serde tables CREATE data source tables
STORED AS ORC

37
CREATE [TEMPORARY] TABLE [IF NOT EXISTS]
[db_name.]table_name
USING table_provider
[OPTIONS table_property_list]
[PARTITIONED BY (col_name, col_name, ...)]
[CLUSTERED BY (col_name, col_name, ...)
[SORTED BY (col_name [ASC|DESC], ...)]
INTO num_buckets BUCKETS]
[LOCATION path]
[COMMENT table_comment]
[AS select_statement];
Unified CREATE TABLE [AS SELECT]
Apache Spark preferred syntax

38
Apache Spark 2.3+
Massive focus on building ETL-friendly pipelines

39
[SPARK-15689] Data Source API v2
1. [SPARK-20960] An efficient column batch interface for data
exchanges between Spark and external systems.
o Cost for conversion to and from RDD[Row]
o Cost for serialization/deserialization
o Publish the columnar binary formats
2. Filter pushdown and column pruning
3. Additional pushdown: limit, sampling and so on.
Target: Apache Spark 2.3

40
Performance: Python UDFs
1. Python is the most popular language for ETL
2. Python UDFs are often used to express elaborate data
conversions/transformations
3. Any improvements to python UDF processing will ultimately
improve ETL.
4. Improve data exchange between Python and JVM
5. Block-level UDFs
o Block-level arguments and return types
Target: Apache Spark 2.3

41
Recap
1. What’s an ETL Pipeline?
2. Using Spark SQL for ETL
- Extract: Dealing with Dirty Data (Bad Records or Files)
- Extract: Multi-line JSON/CSV Support
- Transformation: High-order functions in SQL
- Load: Unified write paths and interfaces
3. New Features in Spark 2.3
- Performance (Data Source API v2, Python UDF)

42
UNIFIED ANALYTICS PLATFORM
Try Apache Spark in Databricks!
• Collaborative cloud environment
• Free version (community edition)
4242
DATABRICKS RUNTIME 3.0
• Apache Spark - optimized for the cloud
• Caching and optimization layer - DBIO
• Enterprise security - DBES
Try for free today.
databricks.com

Questions?
Xiao Li (lixiao@databricks.com)

Building Robust ETL Pipelines with Apache Spark

In this document

More Related Content

What's hot

Similar to Building Robust ETL Pipelines with Apache Spark

More from Databricks

Recently uploaded

Building Robust ETL Pipelines with Apache Spark