[ML] Old trained model deployment got deleted unexpectedly after a new one is added through inference API #137134
Description
Version:
9.2.0
Step to reproduce:
0. Update the scale_to_zero_time setting to 1 minutes
PUT /_cluster/settings
{
"persistent": {
"xpack.ml.trained_models.adaptive_allocations.scale_to_zero_time": "1m"
}
}
- Create an inference endpoint,
PUT _inference/rerank/mytest-old
{
"service": "elasticsearch",
"service_settings": {
"num_threads": 1,
"model_id": ".rerank-v1",
"adaptive_allocations": {
"enabled": true,
"min_number_of_allocations": 0,
"max_number_of_allocations": 2
}
}
}
- After trained model deployed and started (can use
GET _ml/trained_models/_statsto check stats), wait couple minutes until the number_of_allocations turns to 0:"number_of_allocations": 0 - Create another inference endpoint,
PUT _inference/rerank/mytest-new
{
"service": "elasticsearch",
"service_settings": {
"num_threads": 1,
"model_id": ".rerank-v1",
"adaptive_allocations": {
"enabled": true,
"min_number_of_allocations": 0,
"max_number_of_allocations": 2
}
}
}
- then run
GET _ml/trained_models/_stats
the previous mytest-old model deployment got deleted unexpectedly.
Note:
it can be reproduced on 9.1.0, but not on 8.19.5