Introduction to AWS Glue
Simple, Flexible, Cost-Effective ETL
Debanjan Saha, General Manager
Amazon Aurora, Amazon RDS MySQL, AWS Glue
August 14, 2017
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Today’s agenda
> Why did we build AWS Glue?
> Main components of AWS Glue
> What did we announce today?
Why would AWS get into the ETL space?
We have lots of ETL partners
Amazon Redshift Partner Page for Data Integration
Fivetran
The problem is
70% of ETL jobs are hand-coded
With no use of ETL tools.
Actually…
It’s over 90% in the cloud
Why do we see so much hand-coding?
Code is flexible Code is powerful
You can unit test You can deploy with other code You know your dev tools
Hand-coding involves a lot of
undifferentiated heavy lifting…
Brittle Error-prone Laborious
► As data formats change ► As you add sources
► As target schemas change ► As data volume grows
AWS Glue automates
the undifferentiated heavy lifting of ETL
Automatically discover and categorize your data making it immediately searchable
Discover
and queryable across data sources
Develop Generate code to clean, enrich, and reliably move data between various data
sources; you can also use their favorite tools to build ETL jobs
Run your jobs on a serverless, fully managed, scale-out environment. No compute
Deploy
resources to provision or manage.
AWS Glue: Components
Hive Metastore compatible with enhanced functionality
Crawlers automatically extracts metadata and creates tables
Integrated with Amazon Athena, Amazon Redshift Spectrum
Data Catalog
Auto-generates ETL code
Build on open frameworks – Python and Spark
Developer-centric – editing, debugging, sharing
Job Authoring
Run jobs on a serverless Spark platform
Provides flexible scheduling
Handles dependency resolution, monitoring and alerting
Job Execution
Common use cases for AWS Glue
Understand your data assets
Instantly query your data lake on Amazon S3
ETL data into your data warehouse
Build event-driven ETL pipelines
Main components of AWS Glue
AWS Glue Data Catalog
Discover and organize your data
Glue data catalog
Manage table metadata through a Hive metastore API or Hive SQL.
Supported by tools like Hive, Presto, Spark etc.
We added a few extensions:
Search over metadata for data discovery
Connection info – JDBC URLs, credentials
Classification for identifying and parsing files
Versioning of table metadata as schemas evolve and other metadata are updated
Populate using Hive DDL, bulk import, or automatically through Crawlers.
Data Catalog: Crawlers
Crawlers automatically build your Data Catalog and keep it in sync
Automatically discover new data, extracts schema definitions
• Detect schema changes and version tables
• Detect Hive style partitions on Amazon S3
Built-in classifiers for popular types; custom classifiers using Grok expressions
Run ad hoc or on a schedule; serverless – only pay when crawler runs
AWS Glue Data Catalog
Bring in metadata from a variety of data sources (Amazon S3, Amazon Redshift, etc.) into a single
categorized list that is searchable
Data Catalog: Table details
Table properties
Nested fields
Data statistics
Table schema
Data Catalog: Version control
Compare schema versions List of table versions
Data Catalog: Detecting partitions
S3 bucket hierarchy Table definition
Column Type
sim=.93 month=Nov
month str
date str
sim=.99 date=10 … sim=.95 date=15 col 1 int
col 2 float
file 1 … file N file 1 … file N
Estimate schema similarity among files at each level to
handle semi-structured logs, schema evolution…
Data Catalog: Automatic partition detection
Table
partitions
Automatically register available partitions
Job authoring in AWS Glue
Python code generated by AWS Glue
You have choices on
how to get started Connect a notebook or IDE to AWS Glue
Existing code brought into AWS Glue
Job authoring: Automatic code generation
1. Customize the mappings
2. Glue generates transformation graph and Python code
3. Connect your notebook to development endpoints to customize your code
Job authoring: ETL code
Human-readable, editable, and portable PySpark code
Flexible: Glue’s ETL library simplifies manipulating complex, semi-structured data
Customizable: Use native PySpark, import custom libraries, and/or leverage Glue’s libraries
Collaborative: share code snippets via GitHub, reuse code across jobs
Job Authoring: Glue Dynamic Frames
Like Spark’s Data Frames, but better for:
Dynamic frame schema • Cleaning and (re)-structuring semi-structured
data sets, e.g. JSON, Avro, Apache logs ...
No upfront schema needed:
• Infers schema on-the-fly, enabling transformations
A B1 B2 C D[ ] in a single pass
Easy to handle the unexpected:
X Y
• Tracks new fields, and inconsistent changing data
types with choices, e.g. integer or string
• Automatically mark and separate error records
Job Authoring: Glue transforms
Adaptive and flexible
project cast separate into cols
ResolveChoice() B B B B B B B
C
Apply Mapping() A
A X Y
X Y
Job authoring: Relationalize() transform
Semi-structured schema Relational schema
A B B C.X C.Y FK
PK Offset Value
A B B C D[ ]
X Y
• Transforms and adds new columns, types, and tables on-the-fly
• Tracks keys and foreign keys across runs
• SQL on the relational schema is orders of magnitude faster than JSON processing
Job authoring: Glue transformations
Prebuilt transformation: Click and
add to your job with simple
configuration
Spigot writes sample data from
DynamicFrame to S3 in JSON format
Expanding… more transformations
to come
Job authoring: Write your own scripts
Convert to a Spark Data Frame
for complex SQL-based ETL
Convert back to Glue Dynamic Frame
for semi-structured processing and
AWS Glue connectors
Import custom libraries required by your code
Job authoring: Developer endpoints
Glue Spark environment
Remote Interpreter
interpreter server
Environment to iteratively develop and test ETL code.
Connect your IDE or notebook (e.g. Zeppelin) to a Glue development endpoint.
When you are satisfied with the results you can create an ETL job that runs your code.
Job Authoring: Leveraging the community
No need to start from scratch.
Use Glue samples stored in Github to share, reuse,
contribute: https://github.com/awslabs/aws-glue-samples
• Migration scripts to import existing Hive Metastore data
into AWS Glue Data Catalog
• Examples of how to use Dynamic Frames and
Relationalize() transform
• Examples of how to use arbitrary PySpark code with
Glue’s Python ETL library
Download Glue’s Python ETL library to start developing code
in your IDE: https://github.com/awslabs/aws-glue-libs
Orchestration and resource management
Fully managed, serverless job execution
Job execution: Scheduling and monitoring
Compose jobs globally with event-
based dependencies Marketing: Ad-spend by
customer segment
Easy to reuse and leverage work across Event Based
organization boundaries Lambda Trigger
Data
Multiple triggering mechanisms based
Schedule-based: e.g., time of day
Event-based: e.g., job completion
Schedule
On-demand: e.g., AWS Lambda Weekly Data
sales based Central: ROI by
More coming soon: Data Catalog based customer
events, S3 notifications and Amazon segment
CloudWatch events Sales: Revenue by
customer segment
Logs and alerts are available in
Amazon CloudWatch
Job execution: Job bookmarks
Option Behavior
Glue keeps track of data that has already Enable Pick up from where you left off
been processed by a previous run of an Ignore and process the entire dataset
Disable
ETL job. This persisted state information is every time
called a bookmark. Temporarily disable advancing the
Pause
bookmark
For example, you get new files everyday
in your S3 bucket. By default, AWS Glue Data objects
keeps track of which files have been
successfully processed by the job to
prevent data duplication. Marketing: Ad-spend by customer segment
Job execution: Serverless
Compute instances
There is no need to provision, configure, or
manage servers
Auto-configure VPC and role-based access
Customers can specify the capacity that
gets allocated to each job
Automatically scale resources (on post-GA
roadmap)
You pay only for the resources you
consume while consuming them
Customer VPC Customer VPC
AWS Glue pricing examples
AWS Glue pricing
ETL jobs, development endpoints, and crawlers
Compute based usage: $0.44 per DPU-Hour, 1 minute increments.10-minute minimum
A single DPU Unit = 4 vCPU and 16 GB of memory
Data Catalog Storage:
Free for the first million objects stored
Data Catalog usage: $1 per 100,000 objects, per month, stored above 1M
Data Catalog Requests:
Free for the first million requests per month
$1 per million requests above 1M
Glue ETL pricing example
Consider an ETL job that ran for 10 minutes on a 6 DPU environment.
The price of 1 DPU-Hour in US East (N. Virginia) is $0.44.
The cost for this job run = 6 DPUs * (10/60) hour * $0.44 per DPU-Hour or $0.44.
Now consider you provision a development endpoint to debug the code for this job
and keep the development endpoint active for 24 min.
Each development endpoint is provisioned with 5 DPUs
The cost to use the development endpoint = 5 DPUs * (24/ 60) hour * 0.44 per DPU-Hour or $0.88.
Glue Data Catalog pricing example
Let’s consider that you store 1 million tables in your Data Catalog in a given
month and make 1 million requests to access these tables.
You pay $0 for using data catalog. You are covered under the Data Catalog free tier.
Now consider your requests double to 2 million requests.
You will only be paying for one million requests above the free tier, which is $1
If you use crawlers to find new tables and they run for 30 min and use 2 DPUs.
You will pay for 2 DPUs * (30/60) hour * $0.44 per DPU-Hour or $0.44.
Your total monthly bill = $0 + $1 + $0.44 or $1.44
AWS Glue regional availability
AWS Glue regional availability plan
Planned schedule Regions
At launch US East (N. Virginia)
Q3 2017 US East (Ohio), US West (Oregon)
Q4 2017 EU (Ireland), Asia Pacific (Tokyo), Asia Pacific (Sydney)
2018 Rest of the public regions
Q&A
Thank you!
