Glutil

A collection of utilities for managing partitions of tables in the AWS Glue Data Catalog that are built on datasets stored in S3.

Background

AWS's Glue Data Catalog provides an index of the location and schema of your data across AWS data stores and is used to reference sources and targets for ETL jobs in AWS Glue. It is fully-integrated with AWS Athena, an ad-hoc query tool that uses the Hive metastore to build external tables on top of S3 data and PrestoDB to query the data with standard SQL.

Journera heavily uses Kinesis Firehoses to write data from our platform to S3 in near real-time, Athena for ad-hoc analysis of data on S3, and Glue's serverless engine to execute PySpark ETL jobs on S3 data using the tables defined in the Data Catalog. Using the Data Catalog is generally pretty great, but sometimes all of these managed services don't play well together, or a configuration mistake was made (e.g., in a table DDL). For those cases, we have these utilities.

Our original use case for this project was as a Glue Crawler replacement for adding new partitions to tables that don't use Hive-style partitions and for tables built on top of S3 datasets that the Glue Crawler could not successfully parse. For the most part this is a workaround, because of current limitations with the Glue Crawler and Terraform, which does not support configuring Kinesis Firehoses to write JSON data to S3 using formatted prefixes.

Installation

glutil can be installed using pip.

pip install glutil

If you wish to manually install it, you can clone the repository and run

python3 setup.py install

Provided Utilities

There are three main ways to use these utilities, either by using the glutil library in your python code, by using the provided glutil command line script, or as a lambda replacement for a Glue Crawler.

Built-In Assumptions

Because glutil started life as a way to work with Journera-managed data there are still a number of assumptions built in to the code. Ideally these will be removed in the future to enable use with more diverse sets of data.

1. The tables use S3 as their backing data store.

2. All partitions are stored under the table's location.

   For example, if you have a table with the location s3://some-data-bucket/my-table/, glutil will only find partitions located in s3://some-data-bucket/my-table/. Your table location can be as deep or shallow as you want, glutil will operate the same for a table located in s3://bucket/path/to/table/it/goes/here/ and s3://bucket/.

3. Your partition keys are [year, month, day, hour].

4. Your partitions are stored in one of two ways (examples assume your table's location is s3://bucket/table/):

   1. Partitions are stored in key=value form or pathed similarly (examples below):

      s3://bucket/table/YYYY/MM/DD/HH/
      s3://bucket/table/year=YYYY/month=MM/day=DD/hour=HH/
      

   2. You have a single partition key, which is your path with slashes changed into dashes (examples below):

      s3://bucket/table/YYYY/MM/DD/HH/ => partition value of YYYY-MM-DD-HH
      

IAM Permissions

To use glutil you need the following IAM permissions:

• glue:GetDatabase
• glue:GetTable
• glue:GetTables
• glue:BatchCreatePartition
• glue:BatchDeleteTable
• glue:BatchDeletePartition
• glue:GetPartitions
• glue:UpdatePartition
• s3:ListBucket on the buckets containing your data
• s3:GetObject on the buckets containing your data

If you're only using the create-partition lambda, you can get by with only:

• glue:GetDatabase
• glue:GetTable
• glue:BatchCreatePartition
• glue:GetPartitions
• s3:ListBucket
• s3:GetObject

glutil Command Line Interface (CLI)

The glutil CLI includes a number of subcommands for managing partitions and fixing the Glue Data Catalog when things go wrong. Most of the commands were written to fix issues caused by a Glue Crawler gone wrong, moving underlying data, or dealing with newly created data.

For the most part, they operate with the leading principle that any action they take can be reversed (if it was an incorrect action) by running the glutil create-partitions command.

All commands support the --dry-run flag, which will output the command's expected result without modifying the Glue Data Catalog.

Below are short descriptions of the available commands. For larger descriptions and command line arguments, run glutil <command> --help.

glutil create-partitions

create-partitions is the original use case for this code. Running it will search S3 for partitioned data, and will create new partitions for data missing from the Glue Data Catalog.

glutil delete-all-partitions

delete-all-partitions will query the Glue Data Catalog and delete any partitions attached to the specified table. For the most part it is substantially faster to just delete the entire table and recreate it because of AWS batch limits, but sometimes it's harder to recreate than to remove all partitions.

glutil delete-bad-partitions

delete-bad-partitions will remove partitions that meet the following criteria from the catalog:

• Partitions without any data in S3
• Partitions with values that do not match their S3 location (ex. Partition with values [2019 01 02 03] with a location of anything other than s3://table/path/2019/01/02/03/)

In general, if you use glutil create-partitions multiple times and see attempts to create the same partition both times, you should run delete-bad-partitions and try create-partitions again.

glutil delete-missing-partitions

delete-missing-partitions will remove any partition in the Glue Data Catalog without data in S3.

glutil update-partitions

update-partitions should be run after moving your data in S3 and updating your table's location in the catalog. It updates partitions by finding all partitions in S3, and checking if a partition with matching values exists in the catalog. If it finds a matching partition, it updates the existing partition with the new location.

glutil delete-bad-tables

Sometimes when running a Glue Crawler, the crawler doesn't aggregate the data correctly, and instead creates tables for individual partitions. When this happens, it may create a large number of junk tables in the catalog. delete-bad-tables should be run to fix this.

delete-bad-tables deletes any tables in your Glue Data Catalog that meet the following criteria:

• A table with a path that is below another table's path.

  For example, if you have two tables with these paths:

  s3://some-data-bucket/table-path/, and
  s3://some-data-bucket/table-path/another-table/
  

  The table at s3://some-data-bucket/table-path/another-table/ will be deleted.

• A table with the same location as another, with a name that's a superstring of the other's (this is from the Glue Crawler semantic of creating tables which would otherwise have the same name with the name {table}-somelongid).

  For example, if you have the tables foo and foo-buzzer, both with the same location, foo-buzzer will be deleted.

Running create-partitions as a Lambda

Journera's biggest use for this library is as a Glue Crawler replacement for tables and datasets the Glue Crawlers have problems parsing. Information on this lambda can be found in the lambda directory.

Contributing to Glutil

This project was recently open-sourced. As such, please pardon any sharp edges, and let us know about them by creating an issue.

All contributions, bug reports, bug fixes, documentation improvements, enhancements and ideas are welcome.

A detailed overview on how to contribute can be found in the contributing guide.

License

BSD 3