Skip to main content

Removing Metadata from DataHub

There are a two ways to delete metadata from DataHub:

  1. Delete metadata attached to entities by providing a specific urn or filters that identify a set of entities
  2. Delete metadata created by a single ingestion run

To follow this guide you need to use DataHub CLI.

Read on to find out how to perform these kinds of deletes.

Note: Deleting metadata should only be done with care. Always use --dry-run to understand what will be deleted before proceeding. Prefer soft-deletes (--soft) unless you really want to nuke metadata rows. Hard deletes will actually delete rows in the primary store and recovering them will require using backups of the primary metadata store. Make sure you understand the implications of issuing soft-deletes versus hard-deletes before proceeding.

info

Deleting metadata using DataHub's CLI and GraphQL API is a simple, systems-level action. If you attempt to delete an Entity with children, such as a Domain, it will not delete those children, you will instead need to delete each child by URN in addition to deleting the parent.

Delete By Urn

To delete all the data related to a single entity, run

Soft Delete (the default)

This sets the Status aspect of the entity to Removed, which hides the entity and all its aspects from being returned by the UI.

datahub delete --urn "<my urn>"

or

datahub delete --urn "<my urn>" --soft

Hard Delete

This physically deletes all rows for all aspects of the entity. This action cannot be undone, so execute this only after you are sure you want to delete all data associated with this entity.

datahub delete --urn "<my urn>" --hard

As of datahub v0.8.35 doing a hard delete by urn will also provide you with a way to remove references to the urn being deleted across the metadata graph. This is important to use if you don't want to have ghost references in your metadata model and want to save space in the graph database. For now, this behaviour must be opted into by a prompt that will appear for you to manually accept or deny.

You can optionally add -n or --dry-run to execute a dry run before issuing the final delete command. You can optionally add -f or --force to skip confirmations You can optionally add --only-soft-deleted flag to remove soft-deleted items only.

:::note

Make sure you surround your urn with quotes! If you do not include the quotes, your terminal may misinterpret the command._

:::

If you wish to hard-delete using a curl request you can use something like below. Replace the URN with the URN that you wish to delete

curl "http://localhost:8080/entities?action=delete" -X POST --data '{"urn": "urn:li:dataset:(urn:li:dataPlatform:hive,fct_users_deleted,PROD)"}'

Delete by filters

_Note: All these commands below support the soft-delete option (-s/--soft) as well as the dry-run option (-n/--dry-run).

Delete all Datasets from the Snowflake platform

datahub delete --entity_type dataset --platform snowflake

Delete all containers for a particular platform

datahub delete --entity_type container --platform s3

Delete all datasets in the DEV environment

datahub delete --env DEV --entity_type dataset

Delete all Pipelines and Tasks in the DEV environment

datahub delete --env DEV --entity_type "dataJob"
datahub delete --env DEV --entity_type "dataFlow"

Delete all bigquery datasets in the PROD environment

datahub delete --env PROD --entity_type dataset --platform bigquery

Delete all looker dashboards and charts

datahub delete --entity_type dashboard --platform looker
datahub delete --entity_type chart --platform looker

Delete all datasets that match a query

datahub delete --entity_type dataset --query "_tmp"

Rollback Ingestion Run

The second way to delete metadata is to identify entities (and the aspects affected) by using an ingestion run-id. Whenever you run datahub ingest -c ..., all the metadata ingested with that run will have the same run id.

To view the ids of the most recent set of ingestion batches, execute

datahub ingest list-runs

That will print out a table of all the runs. Once you have an idea of which run you want to roll back, run

datahub ingest show --run-id <run-id>

to see more info of the run.

Alternately, you can execute a dry-run rollback to achieve the same outcome.

datahub ingest rollback --dry-run --run-id <run-id>

Finally, once you are sure you want to delete this data forever, run

datahub ingest rollback --run-id <run-id>

to rollback all aspects added with this run and all entities created by this run. This deletes both the versioned and the timeseries aspects associated with these entities.

Unsafe Entities and Rollback

NOTE: Preservation of unsafe entities has been added in datahub 0.8.32. Read on to understand what it means and how it works.

In some cases, entities that were initially ingested by a run might have had further modifications to their metadata (e.g. adding terms, tags, or documentation) through the UI or other means. During a roll back of the ingestion that initially created these entities (technically, if the key aspect for these entities are being rolled back), the ingestion process will analyse the metadata graph for aspects that will be left "dangling" and will:

  1. Leave these aspects untouched in the database, and soft-delete the entity. A re-ingestion of these entities will result in this additional metadata becoming visible again in the UI, so you don't lose any of your work.
  2. The datahub cli will save information about these unsafe entities as a CSV for operators to later review and decide on next steps (keep or remove).

The rollback command will report how many entities have such aspects and save as a CSV the urns of these entities under a rollback reports directory, which defaults to rollback_reports under the current directory where the cli is run, and can be configured further using the --reports-dir command line arg.

The operator can use datahub get --urn <> to inspect the aspects that were left behind and either keep them (do nothing) or delete the entity (and its aspects) completely using datahub delete --urn <urn> --hard. If the operator wishes to remove all the metadata associated with these unsafe entities, they can re-issue the rollback command with the --nuke flag.