Configuring and Managing S3Guard
Data written to Amazon S3 buckets is subject to the "eventual consistency" guarantee provided by Amazon Web Services (AWS), which means that data written to S3 may not be immediately available for queries and listing operations. This can cause failures in multi-step ETL workflows, where data from a previous step is not available to the next step. The S3Guard feature guarantees a consistent view of data stored in Amazon S3 by storing additional metadata in a table residing in an Amazon DynamoDB instance. Depending on the workload, this additional metadata store may also improve performance for Hive, Spark, and Impala jobs.
All processes that modify the S3 bucket that S3Guard is enabled for must use S3Guard. Since S3Guard works by logging metadata changes to an external database, modifying the bucket outside of S3Guard will cause the S3 data and the S3Guard database to go out of sync. This can cause issues such as S3A/S3Guard thinking that files are or are not present despite the bucket having different data.
To enable S3Guard, you set up an Amazon DynamoDB database from Amazon Web Services. Amazon charges an hourly rate for this service based on the capacity you provision.
When the data stored in S3 eventually becomes consistent (usually within 24 hours or less), the S3Guard metadata is no longer required and you can periodically prune the S3Guard metadata stored in the DynamoDB to clear older entries. Pruning can also reduce costs associated with the DynamoDB.
- Credentials for the Amazon S3 bucket.
- An instance of Amazon DynamoDB database provisioned from Amazon Web Services.
- The configured region for the DynamoDB database.
- A cluster managed by Cloudera Manager.
Pruning the S3Guard Metadata
Amazon charges for the amount of data stored in the DynamoDB and the bandwidth used for reads and writes to the database. To optimize costs and improve performance, you can remove stale metadata from the DynamoDB table by running the Prune command. Generally, data written to S3 becomes consistent after 24 hours or less, meaning that you only need to maintain metadata in DynamoDB for about one day. You can monitor the usage of DynamoDB using AWS tools to determine how often and when to prune the table.
Running the Prune command removes all metadata that is older than
the age you specify with the S3Guard Metadata Pruning Age property in
the S3Guard configuration. You can run this command from the Cloudera Manager Admin Console,
or you can create a script to run the Prune command automatically using the Cloudera Manager
API. Cloudera recommends that you run that script using a Linux
cron job or
other scheduling mechanism to regularly prune the metadata.