Switch branches/tags
Nothing to show
Find file Copy path
5ed3ddf Aug 14, 2017
1 contributor

Users who have contributed to this file

330 lines (220 sloc) 15.3 KB

AWS Glue FAQ, or How to Get Things Done

1. How do I repartition or coalesce my output into more or fewer files?

AWS Glue is based on Apache Spark, which partitions data across multiple nodes to achieve high throughput. When writing data to a file-based sink like Amazon S3, Glue will write a separate file for each partition. In some cases it may be desirable to change the number of partitions, either to change the degree of parallelism or the number of output files.

To change the number of partitions in a DynamicFrame, you can first convert it into a DataFrame and then leverage Apache Spark's partitioning capabilities. For example, the first line of the following snippet converts the DynamicFrame called "datasource0" to a DataFrame and then repartitions it to a single partition. The second line converts it back to a DynamicFrame for further processing in AWS Glue.

# Convert to a dataframe and partition based on "partition_col"
partitioned_dataframe = datasource0.toDF().repartition(1)

# Convert back to a DynamicFrame for further processing.
partitioned_dynamicframe = DynamicFrame.fromDF(partitioned_dataframe, glueContext, "partitioned_df")

You can also pass the name of a column to the repartition method to use that field as a partitioning key. This may help performance in certain cases where there is benefit in co-locating data. Note that while different records with the same value for this column will be assigned to the same partition, there is no guarantee that there will be a separate partition for each distinct value.

2. I have a ChoiceType in my schema. So...

a. How can I convert to a data frame?

Since DataFrames do not have the type flexibility that DynamicFrames do, you have to resolve the choice type in your DynamicFrame before conversion. Glue provides a transformation called ResolveChoice with the following signature:

 ResolveChoice.apply(self, frame, specs = None, choice = "",
                     transformation_ctx = "", info = "",
                     stageThreshold = 0, totalThreshold = 0)

This transformation provides you two general ways to resolve choice types in a DynamicFrame.

  • You can specify a list of (path, action) tuples for each individual choice column, where path is the full path of the column and action is the strategy to resolve the choice in this column.

  • You can give an action for all the potential choice columns in your data using the choice parameter.

The action above is a string, one of four strategies that AWS Glue provides:

  1. cast - When this is specified, the user must specify a type to cast to, such as cast:int.

  2. make_cols This flattens a potential choice. For instance, if col1 is choice<int, string>, then using make_cols creates two columns in the target: col1_int and col1_string.

  3. make_struct This creates a struct containing both choices. For example, if col1 is choice<int, string>, then using make_struct creates a column called struct<int, string> that contains one or the other of the choice types.

  4. project When this is specified, the user must also specify a type. For instance, when project:string is specified for col1 that is choice<int, string>, then the column produced in the target would only contain string values.

Once all the choice types in your DynamicFrame are resolved, you can convert it to a data frame using the 'toDF()' method.

b. How do I write to targets that do not handle ChoiceTypes?

Resolve the choice types as described above and then write the data out using DynamicFrame writers or DataFrame write, depending on your use case.

3. There are some transforms that I cannot figure out.

a. How can I use SQL queries with DynamicFrames?

You can leverage Spark's SQL engine to run SQL queries over your data. If you have a DynamicFrame called my_dynamic_frame, you can use the following snippet to convert the DynamicFrame to a DataFrame, issue a SQL query, and then convert back to a DynamicFrame.

  df = my_dynamic_frame.toDF()
  sql_df = spark.sql("SELECT * FROM temptable")
  new_dynamic_frame = DynamicFrame.fromDF(sql_df, glueContext, "new_dynamic_frame")

Note that we assign the spark variable at the start of generated scripts for convenience, but if you modify your script and delete this variable, you can also reference the SparkSession using glueContext.spark_session.

b. How do I do filtering in DynamicFrames?

DynamicFrames support basic filtering via the SplitRows transformation which partitions it into two new DynamicFrames based on a predicate. For example, the snippet partition my_dynamic_frame into two frames called "adults" and "youths".:

  frame_collection = SplitRows.apply(my_dynamic_frame,
                                     {"age": {">": 21}},
                                     "adults", "youths")

You can access these by indexing into the frame_collection. For instance,
`frame_collection['adults']` returns the DynamicFrame containing all records
with `age > 21`.

 It is possible to perform more sophisticated filtering by converting to a
 DataFrame and then using the filter method. For instance, the query above
 could be expressed as:

 result = my_dynamic_frame.toDF().filter("age > 21")
 new_dynamic_frame = DynamicFrame.fromDF(result, glueContext, "new_dynamic_frame")

 More information about Spark SQL's filter syntax can be found in the [Spark SQL
 Programming Guide](

c. Can I use a lambda function with a DynamicFrame? Lambda functions and other user-defined functions are currently only supported using SparkSQL DataFrames. You can convert a DynamicFrame to a DataFrame using the toDF() method and then specify Python functions (including lambdas) when calling methods like foreach. More information about methods on DataFrames can be found in the Spark SQL Programming Guide or the PySpark Documentation.

d. So, what else can I do with DynamicFrames?

DynamicFrames are designed to provide maximum flexibility when dealing with messy data that may lack a declared schema. Records are represented in a flexible self-describing way that preserves information about schema inconsistencies in the data.

For example, with changing requirements, an address column stored as a string in some records might be stored as a struct in later rows. Rather than failing or falling back to a string, DynamicFrames will track both types and gives users a number of options in how to resolve these inconsistencies, providing fine grain resolution options via the ResolveChoice transforms.

DynamicFrames also provide a number of powerful high-level ETL operations that are not found in DataFrames. For example, the Relationalize transform can be used to flatten and pivot complex nested data into tables suitable for transfer to a relational database. In additon, the ApplyMapping transform supports complex renames and casting in a declarative fashion.

DynamicFrames are also integrated with the AWS Glue Data Catalog, so creating frames from tables is a simple operation. Writing to databases can be done through connections without specifying the password. Moreover, DynamicFrames are integrated with job bookmarks, so running these scripts in the job system can allow the script to implictly keep track of what was read and written.

We are continually adding new transform, so be sure to check our documentation and let us know if there are new transforms that would be useful to you.

4. File Formats?

a. Which file formats do you support for input and for output?

Out of the box we support JSON, CSV, ORC, Parquet, and Avro.

b. What compression types do you support?

We support gzip, bzip2, and lz4.

5. JDBC access?

a. Which JDBC databases do you support?

Out of the box we support Postgres, MySQL, Redshift, and Aurora.

b. Can you support my JDBC driver for database XYZ?

In the case of unsupported databases, we fall back to using Spark. You can specify the driver and driver options using the options fields, and make the driver available using the –extra-jars options in the job arguments.

6. How can I handle multi-line CSV files?

By default we assume that each CSV record is contained on a single line. This allows us to automatically split large files to achieve much better parallelism while reading. If your CSV data does contain multiline fields enclosed in double-quotes, you can set the 'multiLine' table property in the Data Catalog to 'true' to disable splitting.

7. My test connection is not working... What do I do?

AWS Glue uses private IP addresses in the subnet while creating Elastic Network Interface(s) in the customer’s specified VPC/Subnet. Security groups specified in the Connection are applied on each of the ENIs. Check whether your Security Groups allow outbound access and whether they allow connectivity to the database cluster. Also, Spark requires bi-directional connectivity among driver and executor nodes. One of the security groups need to allow ingress rules on all TCP ports. You can prevent it from being open to the world by restricting the source of the Security Group to itself (self-referential security group).

8. I'm getting the error that I have no S3 access... What do I do?

AWS Glue uses private IP addresses in the subnet while creating Elastic Network Interface(s) in customer’s specified VPC/Subnet. Check your VPC route tables to ensure that there is an S3 VPC Endpoint so that traffic does not leave out to the internet. If the S3 buckets that you need to access are in a different region, you need to set up a NAT Gateway (the IP addresses are private).

9. How do I take advantage of JobBookmarks?

a. What is a JobBookmark?

A JobBookmark captures the state of job. It is composed of states for various elements of the job, including sources, transformations, and sinks. Currently we only have implementation for S3 sources and Relationalize. Enabling a JobBookmark ensures that if a job is run again after a previous successful run, it will continue from where it left off. However, if a job is run after a previous failed run, it will process the data that it failed to process in the previous attempt.

b. How do I enable/disable Job Bookmarks?

Bookmarks are optional and can be disabled or suspended and re-enabled in the console.

c. I modified my script... can I reset my JobBookmark?

Yes if you want to reset the Job bookmark, it can be reset in the console, or by calling the ResetJobBookmark API.

d. I made a mistake... can I rewind my JobBookmark?

Currently we don’t support rewinding to any arbitrary state. You can manually clean up the data and reset the bookmark.

10. What is a development endpoint?

A DevEndpoint is used for developing and debugging your ETL scripts. You can connect a Notebook such as Zeppelin, or an IDE, or a terminal to a DevEndpoint, which can then provide interactive development and testing of a pyspark script. Once the script succeeds in the DevEndpoint, you can upload the script to S3 and run it in a Job.

11. Custom Libraries

a. How do I create a Python library and use it with Glue?

You can split your script into multiple scripts and refer to these functions in the main script or within the scripts. You can use the extra python files option in the job to provide a list of files that will be made available to the main script. If you have more than a handful of files or if they are in some hierarchy, you can create a zip archive of the files and just pass the archive as the extra python files in the Job option.

b. *How do I create a Java library and use it with Glue?

It is not a typical use case to write to java function and invoke it in Python. However, it is possible to invoke a java function from Python as follows:

   from py4j.java_gateway import java_import
   java_import(sc._jvm, "com.amazonaws.glue.userjava.UserJava")
   uj = sc._jvm.UserJava()

Provide the jar which has the class com.amazonaws.glue.userjava.UserJava using the extra jars options in the Job argument.

12. GlueContext, SparkContext, and SparkSession, Oh My!

With Spark 2.1, SparkSession is the recommended way to run SQL queries or create temporary table views. Here is an example of a SQL query that uses a SparkSession:

 sql_df = spark.sql("SELECT * FROM temptable")

To simplify using spark for registered jobs in AWS Glue, our code generator initializes the spark session in the spark variable similar to GlueContext and SparkContext.

 spark = glueContext.spark_session

On DevEndpoints, a user can initialize the spark session herself in a similar way.

13. Can I use a graphical tool to build my ETL scripts?

I don't like to write programs, and the console doesn't provide all the transformations I need... Is there any other graphical tool that I can use for building ETL scripts?

We are constantly improving our suite of transformations as well as the ability to graphically construct ETL flows. We would love your feedback on what new transforms you'd like to have and what tools you'd like us to support.

14. Do you support other sources and sinks?

How about DynamoDB, Kinesis, ElasticSearch, and others like those?

We are constantly adding connectors to new data sources. In the meantime, our environment comes with Boto3 pre-installed, so for small data sets you can connect directly to these services using standard API calls through Python. For larger data sets, you can dump their data into S3 periodically and use AWS Glue to schedule jobs to process those data sets.

15. I'm having a problem recrawling an S3 bucket...

It is possible that you might be encountering this problem: Suppose you have an s3 bucket with contents like this:

|--- year=2016
|--- year=2017
|--- unrelated.csv
  • The billing folder contains billing information partitioned by year, while unrelated.csv is a file containing unrelated data.
  • So you created a crawler with target {‘S3 path’ : ‘billing’}, but you were unaware of the unrelated csv file. You expected the crawl to create a single table called billing.
  • But instead, you ended up with three tables named year=2016, year=2017, and unrelated_csv.
  • So, suppose you now exclude the unrelated.csv file and crawl again.
  • Again, you expect that one `billing' table will be created, and the other tables will be deleted... but it doesn't work!

To make the recrawl work properly, you actually have to remove the unrelated.csv file from the bucket-- excluding it will not work.