Iceberg AWS Integrations

Iceberg provides integration with different AWS services through the iceberg-aws module. This section describes how to use Iceberg with AWS.

Enabling AWS Integration

The iceberg-aws module is bundled with Spark and Flink engine runtimes for all versions from 0.11.0 onwards. However, the AWS clients are not bundled so that you can use the same client version as your application. You will need to provide the AWS v2 SDK because that is what Iceberg depends on. You can choose to use the AWS SDK bundle, or individual AWS client packages (Glue, S3, DynamoDB, KMS, STS) if you would like to have a minimal dependency footprint.

All the default AWS clients use the URL Connection HTTP Client for HTTP connection management. This dependency is not part of the AWS SDK bundle and needs to be added separately. To choose a different HTTP client library such as Apache HTTP Client, see the section client customization for more details.

For example, to use AWS features with Spark 3 and AWS clients version 2.15.40, you can start the Spark SQL shell with:


spark-sql --packages $DEPENDENCIES \
    --conf spark.sql.catalog.my_catalog=org.apache.iceberg.spark.SparkCatalog \
    --conf spark.sql.catalog.my_catalog.warehouse=s3://my-bucket/my/key/prefix \
    --conf \
    --conf \
    --conf spark.sql.catalog.my_catalog.lock.table=myGlueLockTable

As you can see, In the shell command, we use --packages to specify the additional AWS bundle and HTTP client dependencies with their version as 2.15.40.

For integration with other engines such as Flink, please read their engine documentation pages that explain how to load a custom catalog.

Glue Catalog

Iceberg enables the use of AWS Glue as the Catalog implementation. When used, an Iceberg namespace is stored as a Glue Database, an Iceberg table is stored as a Glue Table, and every Iceberg table version is stored as a Glue TableVersion. You can start using Glue catalog by specifying the catalog-impl as, just like what is shown in the enabling AWS integration section above. More details about loading the catalog can be found in individual engine pages, such as Spark and Flink.

Glue Catalog ID

There is a unique Glue metastore in each AWS account and each AWS region. By default, GlueCatalog chooses the Glue metastore to use based on the user’s default AWS client credential and region setup. You can specify the Glue catalog ID through catalog property to point to a Glue catalog in a different AWS account. The Glue catalog ID is your numeric AWS account ID. If the Glue catalog is in a different region, you should configure you AWS client to point to the correct region, see more details in AWS client customization.

Skip Archive

By default, Glue stores all the table versions created and user can rollback a table to any historical version if needed. However, if you are streaming data to Iceberg, this will easily create a lot of Glue table versions. Therefore, it is recommended to turn off the archive feature in Glue by setting glue.skip-archive to true. For more details, please read Glue Quotas and the UpdateTable API.

DynamoDB for Commit Locking

Glue does not have a strong guarantee over concurrent updates to a table. Although it throws ConcurrentModificationException when detecting two processes updating a table at the same time, there is no guarantee that one update would not clobber the other update. Therefore, DynamoDB can be used for Glue, so that for every commit, GlueCatalog first obtains a lock using a helper DynamoDB table and then try to safely modify the Glue table.

This feature requires the following lock related catalog properties:

  1. Set lock-impl as
  2. Set lock.table as the DynamoDB table name you would like to use. If the lock table with the given name does not exist in DynamoDB, a new table is created with billing mode set as pay-per-request.

Other lock related catalog properties can also be used to adjust locking behaviors such as heartbeat interval. For more details, please refer to Lock catalog properties.

Warehouse Location

Similar to all other catalog implementations, warehouse is a required catalog property to determine the root path of the data warehouse in storage. By default, Glue only allows a warehouse location in S3 because of the use of S3FileIO. To store data in a different local or cloud store, Glue catalog can switch to use HadoopFileIO or any custom FileIO by setting the io-impl catalog property. Details about this feature can be found in the custom FileIO section.

Table Location

By default, the root location for a table my_table of namespace my_ns is at my-warehouse-location/my-ns.db/my-table. This default root location can be changed at both namespace and table level.

To use a different path prefix for all tables under a namespace, use AWS console or any AWS Glue client SDK you like to update the locationUri attribute of the corresponding Glue database. For example, you can update the locationUri of my_ns to s3://my-ns-bucket, then any newly created table will have a default root location under the new prefix. For instance, a new table my_table_2 will have its root location at s3://my-ns-bucket/my_table_2.

To use a completely different root path for a specific table, set the location table property to the desired root path value you want. For example, in Spark SQL you can do:

CREATE TABLE my_catalog.my_ns.my_table (
    id bigint,
    data string,
    category string)
USING iceberg
OPTIONS ('location'='s3://my-special-table-bucket')
PARTITIONED BY (category);

For engines like Spark that supports the LOCATION keyword, the above SQL statement is equivalent to:

CREATE TABLE my_catalog.my_ns.my_table (
    id bigint,
    data string,
    category string)
USING iceberg
LOCATION 's3://my-special-table-bucket'
PARTITIONED BY (category);

S3 FileIO

Iceberg allows users to write data to S3 through S3FileIO. GlueCatalog by default uses this FileIO, and other catalogs can load this FileIO using the io-impl catalog property.

Progressive Multipart Upload

S3FileIO implements a customized progressive multipart upload algorithm to upload data. Data files are uploaded by parts in parallel as soon as each part is ready, and each file part is deleted as soon as its upload process completes. This provides maximized upload speed and minimized local disk usage during uploads. Here are the configurations that users can tune related to this feature:

Property Default Description
s3.multipart.num-threads the available number of processors in the system number of threads to use for uploading parts to S3 (shared across all output streams)
s3.multipart.part-size-bytes 32MB the size of a single part for multipart upload requests
s3.multipart.threshold 1.5 the threshold expressed as a factor times the multipart size at which to switch from uploading using a single put object request to uploading using multipart upload
s3.staging-dir property value the directory to hold temporary files

S3 Server Side Encryption

S3FileIO supports all 3 S3 server side encryption modes:

To enable server side encryption, use the following configuration properties:

Property Default Description
s3.sse.type none none, s3, kms or custom
s3.sse.key aws/s3 for kms type, null otherwise A KMS Key ID or ARN for kms type, or a custom base-64 AES256 symmetric key for custom type.
s3.sse.md5 null If SSE type is custom, this value must be set as the base-64 MD5 digest of the symmetric key to ensure integrity.

S3 Access Control List

S3FileIO supports S3 access control list (ACL) for detailed access control. User can choose the ACL level by setting the s3.acl property. For more details, please read S3 ACL Documentation.

Object Store File Layout

S3 and many other cloud storage services throttle requests based on object prefix. This means data stored in a traditional Hive storage layout has bad read and write throughput since data files of the same partition are placed under the same prefix. Iceberg by default uses the Hive storage layout, but can be switched to use a different ObjectStoreLocationProvider. In this mode, a hash string is added to the beginning of each file path, so that files are equally distributed across all prefixes in an S3 bucket. This results in minimized throttling and maximized throughput for S3-related IO operations. Here is an example Spark SQL command to create a table with this feature enabled:

CREATE TABLE my_catalog.my_ns.my_table (
    id bigint,
    data string,
    category string)
USING iceberg
PARTITIONED BY (category);

For more details, please refer to the LocationProvider Configuration section.

S3 Strong Consistency

In November 2020, S3 announced strong consistency for all read operations, and Iceberg is updated to fully leverage this feature. There is no redundant consistency wait and check which might negatively impact performance during IO operations.

Hadoop S3A FileSystem

Before S3FileIO was introduced, many Iceberg users choose to use HadoopFileIO to write data to S3 through the S3A FileSystem. As introduced in the previous sections, S3FileIO adopts latest AWS clients and S3 features for optimized security and performance, and is thus recommend for S3 use cases rather than the S3A FileSystem.

S3FileIO writes data with s3:// URI scheme, but it is also compatible with schemes written by the S3A FileSystem. This means for any table manifests containing s3a:// or s3n:// file paths, S3FileIO is still able to read them. This feature allows people to easily switch from S3A to S3FileIO.

If for any reason you have to use S3A, here are the instructions:

  1. To store data using S3A, specify the warehouse catalog property to be an S3A path, e.g. s3a://my-bucket/my-warehouse
  2. For HiveCatalog, to also store metadata using S3A, specify the Hadoop config property hive.metastore.warehouse.dir to be an S3A path.
  3. Add hadoop-aws as a runtime dependency of your compute engine.
  4. Configure AWS settings based on hadoop-aws documentation (make sure you check the version, S3A configuration varies a lot based on the version you use).

AWS Client Customization

Many organizations have customized their way of configuring AWS clients with their own credential provider, access proxy, retry strategy, etc. Iceberg allows users to plug in their own implementation of by setting the client.factory catalog property.

Cross-Account and Cross-Region Access

It is a common use case for organizations to have a centralized AWS account for Glue metastore and S3 buckets, and use different AWS accounts and regions for different teams to access those resources. In this case, a cross-account IAM role is needed to access those centralized resources. Iceberg provides an AWS client factory AssumeRoleAwsClientFactory to support this common use case. This also serves as an example for users who would like to implement their own AWS client factory.

This client factory has the following configurable catalog properties:

Property Default Description
client.assume-role.arn null, requires user input ARN of the role to assume, e.g. arn:aws:iam::123456789:role/myRoleToAssume
client.assume-role.region null, requires user input All AWS clients except the STS client will use the given region instead of the default region chain
client.assume-role.external-id null An optional external ID
client.assume-role.timeout-sec 1 hour Timeout of each assume role session. At the end of the timeout, a new set of role session credentials will be fetched through a STS client.

By using this client factory, an STS client is initialized with the default credential and region to assume the specified role. The Glue, S3 and DynamoDB clients are then initialized with the assume-role credential and region to access resources. Here is an example to start Spark shell with this client factory:

spark-sql --packages org.apache.iceberg:iceberg-spark3-runtime:0.11.0, \
    --conf spark.sql.catalog.my_catalog=org.apache.iceberg.spark.SparkCatalog \
    --conf spark.sql.catalog.my_catalog.warehouse=s3://my-bucket/my/key/prefix \    
    --conf \
    --conf \
    --conf spark.sql.catalog.my_catalog.client.assume-role.arn=arn:aws:iam::123456789:role/myRoleToAssume \
    --conf spark.sql.catalog.my_catalog.client.assume-role.region=ap-northeast-1

Run Iceberg on AWS

Amazon EMR can provision clusters with Spark (EMR 6 for Spark 3, EMR 5 for Spark 2), Hive, Flink, Trino that can run Iceberg.

Amazon Kinesis Data Analytics provides a platform to run fully managed Apache Flink applications. You can include Iceberg in your application Jar and run it in the platform.