Running incremental extraction

Any object added to S3, including new buckets and new versions of existing objects, publishes an event to AWS SQS (as configured by the user).

Incremental extraction consumes the events from SQS queue and creates corresponding Atlas entities.

After running the bulk extraction once, you can run incremental extraction to update Atlas metadata based on the events tracked in S3. This method has the benefit of tracking changes that may not be captured in successive bulk extraction, such as when more than one object version is created between extractions or when objects are deleted.

The incremental extraction operation reads all events produced by S3, then filters the list based on the include and exclude configurations for the extractor. The appropriate messages are passed into the Kafka queue and consumed by Atlas. Because the messages sit in the SQS queue, you don't need to run the incremental extraction operation constantly: when it runs, it will go through all the messages waiting in the queue. Run the extractor often enough that the metadata is not stale by the time it processes, but not so often that it takes longer to run than the time between runs. In addition, note that the SQS Queue does not maintain the events forever. By default, the messages stay for 4 days; you can configure this value to hold the events longer to protect against the chance that the extractor isn't able to process events in time.

In addition to the include and exclude lists described in Defining what assets to extract metadata for, the configuration properties relevant to incremental extraction control the connection to the SQS Queue and how long the extraction actively waits for events from S3. You must configure the name of the queue that the events are published on. You can control how long the extractor waits to receive messages from the queue and how many messages to wait for. When configuring the number of messages to collect, remember that the extractor collects all events from the queue, not just the events corresponding to objects in the whitelist.

In addition, you can set the timeout or retry time so the extractor doesn't keep running even when it isn't able to connect to the queue.

You must plan whether to set up a repeating job to run incremental extraction. Assuming you have a lot of changes that you want to track:

  • Incremental extraction takes more time to process the individual changes than running bulk extraction.

  • Incremental extraction maintains a complete history of objects (shows updates and deletes)

  • Batch is faster, but only captures current state and leaves deleted entities in Atlas

In conclusion, if you want to capture the history of S3 objects, run bulk extraction once then run incremental extraction on a regular basis. As you have more experience with the volume of changes and the speed of the extraction, you can determine if you should run the incremental extraction more frequently.

The following command line example runs the incremental extraction. Assuming the mandatory properties are set in the default configuration file, only the parameter to enable incremental mode is required:

/opt/cloudera/parcels/CDH/lib/atlas/extractors/bin/aws-s3-extractor.sh -e INCREMENTAL