Delete By Query API

edit

The simplest usage of _delete_by_query just performs a deletion on every document that match a query. Here is the API:

POST twitter/_delete_by_query
{
  "query": { 
    "match": {
      "message": "some message"
    }
  }
}

The query must be passed as a value to the query key, in the same way as the Search API. You can also use the q parameter in the same way as the search api.

That will return something like this:

{
  "took" : 147,
  "timed_out": false,
  "deleted": 119,
  "batches": 1,
  "version_conflicts": 0,
  "noops": 0,
  "retries": {
    "bulk": 0,
    "search": 0
  },
  "throttled_millis": 0,
  "requests_per_second": -1.0,
  "throttled_until_millis": 0,
  "total": 119,
  "failures" : [ ]
}

_delete_by_query gets a snapshot of the index when it starts and deletes what it finds using internal versioning. That means that you’ll get a version conflict if the document changes between the time when the snapshot was taken and when the delete request is processed. When the versions match the document is deleted.

Since internal versioning does not support the value 0 as a valid version number, documents with version equal to zero cannot be deleted using _delete_by_query and will fail the request.

During the _delete_by_query execution, multiple search requests are sequentially executed in order to find all the matching documents to delete. Every time a batch of documents is found, a corresponding bulk request is executed to delete all these documents. In case a search or bulk request got rejected, _delete_by_query relies on a default policy to retry rejected requests (up to 10 times, with exponential back off). Reaching the maximum retries limit causes the _delete_by_query to abort and all failures are returned in the failures of the response. The deletions that have been performed still stick. In other words, the process is not rolled back, only aborted. While the first failure causes the abort, all failures that are returned by the failing bulk request are returned in the failures element; therefore it’s possible for there to be quite a few failed entities.

If you’d like to count version conflicts rather than cause them to abort then set conflicts=proceed on the url or "conflicts": "proceed" in the request body.

Back to the API format, you can limit _delete_by_query to a single type. This will only delete tweet documents from the twitter index:

POST twitter/tweet/_delete_by_query?conflicts=proceed
{
  "query": {
    "match_all": {}
  }
}

It’s also possible to delete documents of multiple indexes and multiple types at once, just like the search API:

POST twitter,blog/tweet,post/_delete_by_query
{
  "query": {
    "match_all": {}
  }
}

If you provide routing then the routing is copied to the scroll query, limiting the process to the shards that match that routing value:

POST twitter/_delete_by_query?routing=1
{
  "query": {
    "range" : {
        "age" : {
           "gte" : 10
        }
    }
  }
}

By default _delete_by_query uses scroll batches of 1000. You can change the batch size with the scroll_size URL parameter:

POST twitter/_delete_by_query?scroll_size=5000
{
  "query": {
    "term": {
      "user": "kimchy"
    }
  }
}

URL Parameters

edit

In addition to the standard parameters like pretty, the Delete By Query API also supports refresh, wait_for_completion, wait_for_active_shards, and timeout.

Sending the refresh will refresh all shards involved in the delete by query once the request completes. This is different than the Delete API’s refresh parameter which causes just the shard that received the delete request to be refreshed.

If the request contains wait_for_completion=false then Elasticsearch will perform some preflight checks, launch the request, and then return a task which can be used with Tasks APIs to cancel or get the status of the task. Elasticsearch will also create a record of this task as a document at .tasks/task/${taskId}. This is yours to keep or remove as you see fit. When you are done with it, delete it so Elasticsearch can reclaim the space it uses.

wait_for_active_shards controls how many copies of a shard must be active before proceeding with the request. See here for details. timeout controls how long each write request waits for unavailable shards to become available. Both work exactly how they work in the Bulk API.

requests_per_second can be set to any positive decimal number (1.4, 6, 1000, etc) and throttles rate at which _delete_by_query issues batches of delete operations by padding each batch with a wait time. The throttling can be disabled by setting requests_per_second to -1.

The throttling is done by waiting between batches so that scroll that _delete_by_query uses internally can be given a timeout that takes into account the padding. The padding time is the difference between the batch size divided by the requests_per_second and the time spent writing. By default the batch size is 1000, so if the requests_per_second is set to 500:

target_time = 1000 / 500 per second = 2 seconds
wait_time = target_time - write_time = 2 seconds - .5 seconds = 1.5 seconds

Since the batch is issued as a single _bulk request large batch sizes will cause Elasticsearch to create many requests and then wait for a while before starting the next set. This is "bursty" instead of "smooth". The default is -1.

Response body

edit

The JSON response looks like this:

{
  "took" : 639,
  "deleted": 0,
  "batches": 1,
  "version_conflicts": 2,
  "retries": 0,
  "throttled_millis": 0,
  "failures" : [ ]
}
took
The number of milliseconds from start to end of the whole operation.
deleted
The number of documents that were successfully deleted.
batches
The number of scroll responses pulled back by the the delete by query.
version_conflicts
The number of version conflicts that the delete by query hit.
retries
The number of retries that the delete by query did in response to a full queue.
throttled_millis
Number of milliseconds the request slept to conform to requests_per_second.
failures
Array of all indexing failures. If this is non-empty then the request aborted because of those failures. See conflicts for how to prevent version conflicts from aborting the operation.

Works with the Task API

edit

You can fetch the status of any running delete-by-query requests with the Task API:

GET _tasks?detailed=true&actions=*/delete/byquery

The responses looks like:

{
  "nodes" : {
    "r1A2WoRbTwKZ516z6NEs5A" : {
      "name" : "r1A2WoR",
      "transport_address" : "127.0.0.1:9300",
      "host" : "127.0.0.1",
      "ip" : "127.0.0.1:9300",
      "attributes" : {
        "testattr" : "test",
        "portsfile" : "true"
      },
      "tasks" : {
        "r1A2WoRbTwKZ516z6NEs5A:36619" : {
          "node" : "r1A2WoRbTwKZ516z6NEs5A",
          "id" : 36619,
          "type" : "transport",
          "action" : "indices:data/write/delete/byquery",
          "status" : {    
            "total" : 6154,
            "updated" : 0,
            "created" : 0,
            "deleted" : 3500,
            "batches" : 36,
            "version_conflicts" : 0,
            "noops" : 0,
            "retries": 0,
            "throttled_millis": 0
          },
          "description" : ""
        }
      }
    }
  }
}

this object contains the actual status. It is just like the response json with the important addition of the total field. total is the total number of operations that the reindex expects to perform. You can estimate the progress by adding the updated, created, and deleted fields. The request will finish when their sum is equal to the total field.

With the task id you can look up the task directly:

GET /_tasks/taskId:1

The advantage of this API is that it integrates with wait_for_completion=false to transparently return the status of completed tasks. If the task is completed and wait_for_completion=false was set on it then it’ll come back with results or an error field. The cost of this feature is the document that wait_for_completion=false creates at .tasks/task/${taskId}. It is up to you to delete that document.

Works with the Cancel Task API

edit

Any Delete By Query can be canceled using the Task Cancel API:

POST _tasks/task_id:1/_cancel

The task_id can be found using the tasks API above.

Cancellation should happen quickly but might take a few seconds. The task status API above will continue to list the task until it is wakes to cancel itself.

Rethrottling

edit

The value of requests_per_second can be changed on a running delete by query using the _rethrottle API:

POST _delete_by_query/task_id:1/_rethrottle?requests_per_second=-1

The task_id can be found using the tasks API above.

Just like when setting it on the _delete_by_query API requests_per_second can be either -1 to disable throttling or any decimal number like 1.7 or 12 to throttle to that level. Rethrottling that speeds up the query takes effect immediately but rethrotting that slows down the query will take effect on after completing the current batch. This prevents scroll timeouts.

Manually slicing

edit

Delete-by-query supports Sliced Scroll allowing you to manually parallelize the process relatively easily:

POST twitter/_delete_by_query
{
  "slice": {
    "id": 0,
    "max": 2
  },
  "query": {
    "range": {
      "likes": {
        "lt": 10
      }
    }
  }
}
POST twitter/_delete_by_query
{
  "slice": {
    "id": 1,
    "max": 2
  },
  "query": {
    "range": {
      "likes": {
        "lt": 10
      }
    }
  }
}

Which you can verify works with:

GET _refresh
POST twitter/_search?size=0&filter_path=hits.total
{
  "query": {
    "range": {
      "likes": {
        "lt": 10
      }
    }
  }
}

Which results in a sensible total like this one:

{
  "hits": {
    "total": 0
  }
}

Automatic slicing

edit

You can also let delete-by-query automatically parallelize using Sliced Scroll to slice on _uid:

POST twitter/_delete_by_query?refresh&slices=5
{
  "query": {
    "range": {
      "likes": {
        "lt": 10
      }
    }
  }
}

Which you also can verify works with:

POST twitter/_search?size=0&filter_path=hits.total
{
  "query": {
    "range": {
      "likes": {
        "lt": 10
      }
    }
  }
}

Which results in a sensible total like this one:

{
  "hits": {
    "total": 0
  }
}

Adding slices to _delete_by_query just automates the manual process used in the section above, creating sub-requests which means it has some quirks:

  • You can see these requests in the Tasks APIs. These sub-requests are "child" tasks of the task for the request with slices.
  • Fetching the status of the task for the request with slices only contains the status of completed slices.
  • These sub-requests are individually addressable for things like cancellation and rethrottling.
  • Rethrottling the request with slices will rethrottle the unfinished sub-request proportionally.
  • Canceling the request with slices will cancel each sub-request.
  • Due to the nature of slices each sub-request won’t get a perfectly even portion of the documents. All documents will be addressed, but some slices may be larger than others. Expect larger slices to have a more even distribution.
  • Parameters like requests_per_second and size on a request with slices are distributed proportionally to each sub-request. Combine that with the point above about distribution being uneven and you should conclude that the using size with slices might not result in exactly size documents being `_delete_by_query`ed.
  • Each sub-requests gets a slightly different snapshot of the source index though these are all taken at approximately the same time.

Picking the number of slices

edit

At this point we have a few recommendations around the number of slices to use (the max parameter in the slice API if manually parallelizing):

  • Don’t use large numbers. 500 creates fairly massive CPU thrash.
  • It is more efficient from a query performance standpoint to use some multiple of the number of shards in the source index.
  • Using exactly as many shards as are in the source index is the most efficient from a query performance standpoint.
  • Indexing performance should scale linearly across available resources with the number of slices.
  • Whether indexing or query performance dominates that process depends on lots of factors like the documents being reindexed and the cluster doing the reindexing.