Fix model serving telemetry_config never converging and silently drop…

.nextchanges/bundles/model-serving-telemetry-config-converges.md

@@ -0,0 +1 @@
+Fixed `telemetry_config` on model serving endpoints never converging on the direct engine: the field is only accepted on create and is never returned by the serving endpoints GET API, so every subsequent `bundle plan` reported the endpoint as a perpetual update. Changing `telemetry_config` on a deployed endpoint no longer plans an update that cannot be applied, since the API has no method to update it.

acceptance/bundle/invariant/configs/model_serving_endpoint_telemetry_config.yml.tmpl

@@ -0,0 +1,10 @@
+bundle:
+  name: test-bundle-$UNIQUE_NAME
+
+resources:
+  model_serving_endpoints:
+    foo:
+      name: test-endpoint-$UNIQUE_NAME
+      telemetry_config:
+        inference_table_config:
+          sampling_fraction: 0.5

acceptance/bundle/invariant/test.toml

@@ -48,6 +48,7 @@ EnvMatrix.INPUT_CONFIG = [
     "model.yml.tmpl",
     "model_with_permissions.yml.tmpl",
     "model_serving_endpoint.yml.tmpl",
+    "model_serving_endpoint_telemetry_config.yml.tmpl",
     "pipeline.yml.tmpl",
     "pipeline_allow_duplicate_names.yml.tmpl",
     "pipeline_apply_policy_default_values.yml.tmpl",

acceptance/bundle/resources/model_serving_endpoints/drift/telemetry_config/databricks.yml

@@ -0,0 +1,23 @@
+bundle:
+  name: test-bundle
+
+resources:
+  model_serving_endpoints:
+    endpoint1:
+      name: test-endpoint
+      config:
+        served_entities:
+          - name: prod
+            external_model:
+              name: gpt-4o-mini
+              provider: openai
+              task: llm/v1/chat
+              openai_config:
+                openai_api_key_plaintext: sk-test-plaintext-key
+        traffic_config:
+          routes:
+            - served_model_name: prod
+              traffic_percentage: 100
+      telemetry_config:
+        inference_table_config:
+          sampling_fraction: 0.5

acceptance/bundle/resources/model_serving_endpoints/drift/telemetry_config/output.txt

@@ -0,0 +1,95 @@
+
+=== Initial deployment
+>>> [CLI] bundle deploy
+Uploading bundle files to /Workspace/Users/[USERNAME]/.bundle/test-bundle/default/files...
+Deploying resources...
+Updating deployment state...
+Deployment complete!
+
+=== Plan is a no-op even though telemetry_config does not round-trip
+>>> [CLI] bundle plan
+Plan: 0 to add, 0 to change, 0 to delete, 1 unchanged
+
+=== telemetry_config is skipped (confirms the matched rules)
+>>> [CLI] bundle plan --output json
+{
+  "config.served_entities[0].external_model.openai_config.openai_api_key_plaintext": {
+    "action": "skip",
+    "reason": "input_only",
+    "old": "sk-test-plaintext-key",
+    "new": "sk-test-plaintext-key",
+    "remote": ""
+  },
+  "description": {
+    "action": "skip",
+    "reason": "empty",
+    "remote": ""
+  },
+  "route_optimized": {
+    "action": "skip",
+    "reason": "empty",
+    "remote": false
+  },
+  "telemetry_config": {
+    "action": "skip",
+    "reason": "input_only",
+    "old": {
+      "inference_table_config": {
+        "sampling_fraction": 0.5
+      }
+    },
+    "new": {
+      "inference_table_config": {
+        "sampling_fraction": 0.5
+      }
+    }
+  }
+}
+
+=== Changing telemetry_config plans nothing: it is create-only
+>>> [CLI] bundle plan
+Plan: 0 to add, 0 to change, 0 to delete, 1 unchanged
+
+=== Redeploy is a no-op (no config/ai-gateway/tags update calls)
+>>> [CLI] bundle deploy
+Uploading bundle files to /Workspace/Users/[USERNAME]/.bundle/test-bundle/default/files...
+Deploying resources...
+Updating deployment state...
+Deployment complete!
+
+>>> print_requests.py //serving-endpoints
+{
+  "method": "POST",
+  "path": "/api/2.0/serving-endpoints",
+  "body": {
+    "config": {
+      "served_entities": [
+        {
+          "external_model": {
+            "name": "gpt-4o-mini",
+            "openai_config": {
+              "openai_api_key_plaintext": "sk-test-plaintext-key"
+            },
+            "provider": "openai",
+            "task": "llm/v1/chat"
+          },
+          "name": "prod"
+        }
+      ],
+      "traffic_config": {
+        "routes": [
+          {
+            "served_model_name": "prod",
+            "traffic_percentage": 100
+          }
+        ]
+      }
+    },
+    "name": "test-endpoint",
+    "telemetry_config": {
+      "inference_table_config": {
+        "sampling_fraction": 0.5
+      }
+    }
+  }
+}

acceptance/bundle/resources/model_serving_endpoints/drift/telemetry_config/script

@@ -0,0 +1,18 @@
+echo "*" > .gitignore
+
+title "Initial deployment"
+trace $CLI bundle deploy
+
+title "Plan is a no-op even though telemetry_config does not round-trip"
+trace $CLI bundle plan | contains.py "Plan: 0 to add, 0 to change, 0 to delete, 1 unchanged"
+
+title "telemetry_config is skipped (confirms the matched rules)"
+trace $CLI bundle plan --output json | jq '.plan[].changes'
+
+title "Changing telemetry_config plans nothing: it is create-only"
+update_file.py databricks.yml "sampling_fraction: 0.5" "sampling_fraction: 0.9"
+trace $CLI bundle plan | contains.py "Plan: 0 to add, 0 to change, 0 to delete, 1 unchanged"
+
+title "Redeploy is a no-op (no config/ai-gateway/tags update calls)"
+trace $CLI bundle deploy
+trace print_requests.py //serving-endpoints

acceptance/bundle/resources/model_serving_endpoints/drift/telemetry_config/test.toml

@@ -0,0 +1,4 @@
+RecordRequests = true
+
+# Scope this regression to the direct engine where the classification lives.
+EnvMatrix.DATABRICKS_BUNDLE_ENGINE = ["direct"]

bundle/direct/dresources/resources.yml

@@ -279,6 +279,9 @@ resources:
       # Accepted on write but not returned by GET.
       - field: config.served_entities[*].burst_scaling_enabled
         reason: input_only
+      # Remote always reads back empty, so without this it never converges.
+      - field: telemetry_config
+        reason: input_only
       # Write-only secrets: the backend stores them and returns the reference field, not the plaintext.
       - field: config.served_entities[*].external_model.ai21labs_config.ai21labs_api_key_plaintext
         reason: input_only
@@ -309,6 +312,9 @@ resources:
         reason: no_update_api
       - field: rate_limits
         reason: not_implemented
+      # Only exists on CreateServingEndpoint; no update API accepts it.
+      - field: telemetry_config
+        reason: no_update_api
     backend_defaults:
       # https://github.com/databricks/terraform-provider-databricks/blob/4eba541abe1a9f50993ea7b9dd83874207e224a1/serving/resource_model_serving.go#L383
       # common.CustomizeSchemaPath(m, "config", "served_entities", "name").SetComputed()