Migrating to 8.15

edit

This section discusses the changes that you need to be aware of when migrating your application to Elasticsearch 8.15.

See also What’s new in 8.16 and Release notes.

Coming in 8.15.0.

Breaking changes

edit

The following changes in Elasticsearch 8.15 might affect your applications and prevent them from operating normally. Before upgrading to 8.15, review these changes and take the described steps to mitigate the impact.

Cluster and node setting changes

edit
Change skip_unavailable remote cluster setting default value to true

Details
The default value of the skip_unavailable setting is now set to true. All existing and future remote clusters that do not define this setting will use the new default. This setting only affects cross-cluster searches using the _search or _async_search API.

Impact
Unavailable remote clusters in a cross-cluster search will no longer cause the search to fail unless skip_unavailable is configured to be false in elasticsearch.yml or via the _cluster/settings API. Unavailable clusters with skip_unavailable=true (either explicitly or by using the new default) are marked as SKIPPED in the search response metadata section and do not fail the entire search. If users want to ensure that a search returns a failure when a particular remote cluster is not available, skip_unavailable must be now be set explicitly.

Rollup changes

edit
Disallow new rollup jobs in clusters with no rollup usage

Details
The put rollup API will fail with an error when a rollup job is created in a cluster with no rollup usage

Impact
Clusters with no rollup usage (either no rollup job or index) can not create new rollup jobs

REST API changes

edit
Interpret ?timeout=-1 as infinite ack timeout

Details
Today Elasticsearch accepts the parameter ?timeout=-1 in many APIs, but interprets this to mean the same as ?timeout=0. From 8.15 onwards ?timeout=-1 will mean to wait indefinitely, aligning the behaviour of this parameter with other similar parameters such as ?master_timeout.

Impact
Use ?timeout=0 to force relevant operations to time out immediately instead of ?timeout=-1

Replace model_id with inference_id in GET inference API

Details
From 8.15 onwards the Get inference API response will return an inference_id field instead of a model_id.

Impact
If your application uses the model_id in a GET inference API response, switch it to use inference_id instead.

Deprecations

edit

The following functionality has been deprecated in Elasticsearch 8.15 and will be removed in a future version. While this won’t have an immediate impact on your applications, we strongly encourage you to take the described steps to update your code after upgrading to 8.15.

To find out if you are using any deprecated functionality, enable deprecation logging.

Cluster and node setting deprecations

edit
Deprecate absolute size values for indices.breaker.total.limit setting

Details
Previously, the value of indices.breaker.total.limit could be specified as an absolute size in bytes. This setting controls the overal amount of memory the server is allowed to use before taking remedial actions. Setting this to a specific number of bytes led to strange behaviour when the node maximum heap size changed because the circut breaker limit would remain unchanged. This would either leave the value too low, causing part of the heap to remain unused; or it would leave the value too high, causing the circuit breaker to be ineffective at preventing OOM errors. The only reasonable behaviour for this setting is that it scales with the size of the heap, and so absolute byte limits are now deprecated.

Impact
Users must change their configuration to specify a percentage instead of an absolute number of bytes for indices.breaker.total.limit, or else accept the default, which is already specified as a percentage.

REST API deprecations

edit
Deprecate text_expansion and weighted_tokens queries

Details
The text_expansion and weighted_tokens queries have been replaced by sparse_vector.

Impact
Please update your existing text_expansion and weighted_tokens queries to use sparse_vector.

Deprecate using slm privileges to access ilm

Details
The read_slm privilege can get the ILM status, and the manage_slm privilege can start and stop ILM. Access to these APIs should be granted using the read_ilm and manage_ilm privileges instead. Access to ILM APIs will be removed from SLM privileges in a future major release, and is now deprecated.

Impact
Users that need access to the ILM status API should now use the read_ilm privilege. Users that need to start and stop ILM, should use the manage_ilm privilege.