Min_number_of_allocations for .elser-2-elasticsearch reverted to 0 and inference failures during bulk indexing

Elastic Stack Elasticsearch
1

We are using the preconfigured ELSER inference endpoint .elser-2-elasticsearch with a semantic_text field.

During a full bulk indexing operation, some documents fail with:

inference_exception: Exception when running inference id
[.elser-2-elasticsearch] on field [prism_semantic]

The entire document fails to index when this occurs.

We previously configured the ELSER deployment with:

"adaptive_allocations": {
  "enabled": true,
  "min_number_of_allocations": 1
}

However we later found that the deployment is back to:

"adaptive_allocations": {
  "enabled": true,
  "min_number_of_allocations": 0,
  "max_number_of_allocations": 32
}

Both GET _inference/.elser-2-elasticsearch and GET _ml/trained_models/.elser-2-elasticsearch/_stats now show:

"adaptive_allocations": {
  "enabled": true,
  "min_number_of_allocations": 0,
  "max_number_of_allocations": 32
}

Is it expected for the preconfigured .elser-2-elasticsearch endpoint to reset/recreate its deployment with min_number_of_allocations: 0?

If we need ELSER to remain available during large bulk indexing operations, what is the recommended way to persist min_number_of_allocations: 1?

We are using ECH 8.19.12