druid/docs/data-management/delete.md

id	title
delete	Data deletion

By time range, manually

Apache Druid stores data partitioned by time chunk and supports deleting data for time chunks by dropping segments. This is a fast, metadata-only operation.

Deletion by time range happens in two steps:

1. Segments to be deleted must first be marked as "unused". This can happen when a segment is dropped by a drop rule or when you manually mark a segment unused through the Coordinator API or web console. This is a soft delete: the data is not available for querying, but the segment files remains in deep storage, and the segment records remains in the metadata store.
2. Once a segment is marked "unused", you can use a kill task to permanently delete the segment file from deep storage and remove its record from the metadata store. This is a hard delete: the data is unrecoverable unless you have a backup.

For documentation on disabling segments using the Coordinator API, see the Legacy metadata API reference.

A data deletion tutorial is available at Tutorial: Deleting data.

By time range, automatically

Druid supports load and drop rules, which are used to define intervals of time where data should be preserved, and intervals where data should be discarded. Data that falls under a drop rule is marked unused, in the same manner as if you manually mark that time range unused. This is a fast, metadata-only operation.

Data that is dropped in this way is marked unused, but remains in deep storage. To permanently delete it, use a kill task.

Specific records

Druid supports deleting specific records using reindexing with a filter. The filter specifies which data remains after reindexing, so it must be the inverse of the data you want to delete. Because segments must be rewritten to delete data in this way, it can be a time-consuming operation.

For example, to delete records where userName is 'bob' with native batch indexing, use a transformSpec with filter {"type": "not", "field": {"type": "selector", "dimension": "userName", "value": "bob"}}.

To delete the same records using SQL, use REPLACE with WHERE userName <> 'bob'.

To reindex using native batch, use the druid input source. If needed, transformSpec can be used to filter or modify data during the reindexing job. To reindex with SQL, use REPLACE <table> OVERWRITE with SELECT ... FROM <table>. (Druid does not have UPDATE or ALTER TABLE statements.) Any SQL SELECT query can be used to filter, modify, or enrich the data during the reindexing job.

Data that is deleted in this way is marked unused, but remains in deep storage. To permanently delete it, use a kill task.

Entire table

Deleting an entire table works the same way as deleting part of a table by time range. First, mark all segments unused using the Coordinator API or web console. Then, optionally, delete it permanently using a kill task.

Permanently (kill task)

Data that has been overwritten or soft-deleted still remains as segments that have been marked unused. You can use a kill task to permanently delete this data.

The available grammar is:

{
    "type": "kill",
    "id": <task_id>,
    "dataSource": <task_datasource>,
    "interval" : <all_unused_segments_in_this_interval_will_die!>,
    "context": <task context>,
    "batchSize": <optional_batch size>,
    "limit": <the maximum number of segments to delete>
}

Some of the parameters used in the task payload are further explained below:

Parameter	Default	Explanation
batchSize	100	Maximum number of segments that are deleted in one kill batch. Some operations on the Overlord may get stuck while a kill task is in progress due to concurrency constraints (such as in TaskLockbox). Thus, a kill task splits the list of unused segments to be deleted into smaller batches to yield the Overlord resources intermittently to other task operations.
limit	null - no limit	Maximum number of segments for the kill task to delete.

WARNING: The kill task permanently removes all information about the affected segments from the metadata store and deep storage. This operation cannot be undone.