Cannot update min_number_of_allocations for ELSER inference endpoint.

[vladislav-marochkin]

Elasticsearch Version

8.18.1

Installed Plugins

No response

Java Version

OpenJDK 24

OS Version

Ubuntu with the Linux kernel version 5.4.0-208-generic

Problem Description

Expected behavior:
I expect to be able to update the min_number_of_allocations parameter for an existing ELSER inference endpoint with adaptive allocations enabled, using the _update API. The update should apply the new value without error, as long as the request is valid and consistent with the endpoint’s configuration.

Actual behavior:
When attempting to update the min_number_of_allocations parameter via the _update API, the request fails with a parsing error related to num_allocations. The error message indicates a failure to parse [num_allocations] in the update request, even though only min_number_of_allocations and max_number_of_allocations are being set under adaptive_allocations.

If I try to explicitly include num_allocations in the request, the update fails with a validation error stating that [number_of_allocations] cannot be set if adaptive allocations is enabled.

This makes it impossible to update min_number_of_allocations for an existing ELSER inference endpoint with adaptive allocations enabled, even though this should be a supported operation.

---

We are using ECK

---

Steps to Reproduce

1. Create ELSER inference endpoint.

PUT _inference/elser_model_2_search
{
      "task_type": "sparse_embedding",
      "service": "elasticsearch",
      "service_settings": {
        "num_threads": 1,
        "model_id": ".elser_model_2_linux-x86_64",
        "adaptive_allocations": {
          "enabled": true,
          "min_number_of_allocations": 1,
          "max_number_of_allocations": 10
        }
      },
      "chunking_settings": {
        "strategy": "sentence",
        "max_chunk_size": 250,
        "sentence_overlap": 1
      }
}

2. Try to update "min_number_of_allocations".

PUT _inference/elser_model_2_search/_update
{
      "task_type": "sparse_embedding",
      "service_settings": {
        "num_threads": 1,
        "model_id": ".elser_model_2_linux-x86_64",
        "adaptive_allocations": {
          "enabled": true,
          "min_number_of_allocations": 0,
          "max_number_of_allocations": 10
        }
      }
}

This returns:

{
  "error": {
    "root_cause": [
      {
        "type": "status_exception",
        "reason": """Failed to parse [num_allocations] of update request [{      "task_type": "sparse_embedding",      "service_settings": {        "num_threads": 1,        "model_id": ".elser_model_2_linux-x86_64",        "adaptive_allocations": {          "enabled": true,          "min_number_of_allocations": 0,          "max_number_of_allocations": 10        }      }}
]"""
      }
    ],
    "type": "status_exception",
    "reason": """Failed to parse [num_allocations] of update request [{      "task_type": "sparse_embedding",      "service_settings": {        "num_threads": 1,        "model_id": ".elser_model_2_linux-x86_64",        "adaptive_allocations": {          "enabled": true,          "min_number_of_allocations": 0,          "max_number_of_allocations": 10        }      }}
]"""
  },
  "status": 400
}

3. Try to include "num_allocations" with the update request.

PUT _inference/elser_model_2_search/_update
{
      "task_type": "sparse_embedding",
      "service_settings": {
        "num_threads": 1,
        "num_allocations": 1,
        "model_id": ".elser_model_2_linux-x86_64",
        "adaptive_allocations": {
          "enabled": true,
          "min_number_of_allocations": 0,
          "max_number_of_allocations": 10
        }
      }
}

This returns:

{
  "error": {
    "root_cause": [
      {
        "type": "illegal_argument_exception",
        "reason": "Validation Failed: 1: [number_of_allocations] cannot be set if adaptive allocations is enabled;"
      }
    ],
    "type": "illegal_argument_exception",
    "reason": "Validation Failed: 1: [number_of_allocations] cannot be set if adaptive allocations is enabled;"
  },
  "status": 400
}

Logs (if relevant)

No response

[l-trotta]

To add more info, I tested this in 9.0.1 and the behavior is the same. It could be a validation issue: validation for UpdateTrainedModelDeploymentAction doesn't require num_allocations, but TransportUpdateInferenceModelAction seems to require it.

[elasticsearchmachine]

Pinging @elastic/ml-core (Team:ML)

[prwhelan]

There is a current workaround in this version of Elasticsearch where you can update via the Trained Model API: https://www.elastic.co/docs/api/doc/elasticsearch/operation/operation-ml-update-trained-model-deployment

PUT _inference/elser_model_2_search creates a Trained Model with deployment id elser_model_2_search, so you should be able to use:

POST _ml/trained_models/elser_model_2_search/deployment/_update
{
  "adaptive_allocations": {
    "enabled": true,
    "min_number_of_allocations": 0,
    "max_number_of_allocations": 10
  }
}

In 9.2, PR #130940 should fix this so that we can update it through the inference API. Specifically, if a Trained Model was created via the Inference API, it can be updated via the Inference API.