databricks
[ISSUE] Issue with `databricks_model_serving` resource
Configuration
resource "databricks_model_serving" "this" {
name = "e5_${local.STICKY_RANDOM}"
config {
served_entities {
name = "e5_small_v2"
entity_name = "${databricks_catalog.databricks_e5_v2_models.id}.models.e5_small_v2"
entity_version = "2"
workload_size = "Small"
scale_to_zero_enabled = false
}
traffic_config {
routes {
served_model_name = "e5_small_v2"
traffic_percentage = 100
}
}
}
}
Seems like the issue which was resolved here https://github.com/databricks/terraform-provider-databricks/issues/3383 is happening again with databricks provider versions 1.49.0 and above.
Steps to Reproduce
terraform plan works fine but it fails when we try to do terraform apply with the error
cannot update model serving: scaleToZeroEnabled is undefined. Please set it to either true or false
Terraform and provider versions
databricks_model_serving resource is broken since 1.49.0
Is it a regression?
Debug Output
Important Factoids
Would you like to implement a fix?
Running into the same issue:
Error: cannot update model serving: scaleToZeroEnabled is undefined. Please set it to either true or false.
I noticed that it only works when you set it to true.
Does adding back ForceSendFields which was added in the PR https://github.com/databricks/terraform-provider-databricks/commit/dbe732e02b2223d32f085f66bd1c5c0707304236#diff-c4af24d835708d4701b7681a821ffeed5cb6ad8958b7f0d32a65f911bda3e693R78 will fix the issue? It was removed in the last PR. @alexott ?
@Vikranth-Subramanian it's a bit complicated as it depends on a specific model.
Cc: @arpitjasa-db
Hmm @yucheng-qian-db were any changes made recently on the Model Serving API side?
@arpitjasa-db @yucheng-qian-db I reported this error internally few weeks ago...
No change in API, but we'll make this field no longer required to address this issue.
there is one more issue in Documentation
https://github.com/databricks/terraform-provider-databricks/blob/main/internal/service/serving_tf/model.go#L753
scale_to_zero_enabled is no longer required. The issue should be fixed.