Exempt specific GenAI attributes values from custom properties truncation

sdk/monitor/azure-monitor-opentelemetry-exporter/CHANGELOG.md

@@ -9,6 +9,8 @@
 ### Breaking Changes
 
 ### Bugs Fixed
+- Exempt specific GenAI attributes values from custom properties truncation
+  ([#45596](https://github.com/Azure/azure-sdk-for-python/pull/45596))
 - Fix attributes override in Rate Limited Sampler
   ([#45592](https://github.com/Azure/azure-sdk-for-python/pull/45592))
 - Add environment variable to disable/enable custom properties truncation

sdk/monitor/azure-monitor-opentelemetry-exporter/README.md

@@ -139,10 +139,6 @@ All configuration options can be passed through the constructors of exporters th
 * `storage_directory`: Storage directory in which to store retry files. Defaults to `<tempfile.gettempdir()>/Microsoft/AzureMonitor/opentelemetry-python-<your-instrumentation-key>`.
 * `credential`: Token credential, such as ManagedIdentityCredential or ClientSecretCredential, used for [Azure Active Directory (AAD) authentication][aad_for_ai_docs]. Defaults to None. See [samples][exporter_samples] for examples. The credential will be automatically created from the `APPLICATIONINSIGHTS_AUTHENTICATION_STRING` environment variable if not explicitly passed in. See [documentation][aad_env_var_docs] for more.
 
-## Environment Variables
-
-* Set `AZURE_MONITOR_DISABLE_CUSTOM_DIMENSIONS_LIMIT` to `True` to remove the 64kb truncation limit on custom dimensions. Defaults to `False`.
-
 ## Examples
 
 ### Logging (experimental)

sdk/monitor/azure-monitor-opentelemetry-exporter/azure/monitor/opentelemetry/exporter/_constants.py

@@ -355,7 +355,15 @@ class _RP_Names(Enum):
 # Resource attribute applicationId
 _APPLICATION_ID_RESOURCE_KEY = "microsoft.applicationId"
 
-# Custom dimensions limit truncation toggle
-AZURE_MONITOR_DISABLE_CUSTOM_DIMENSIONS_LIMIT = "AZURE_MONITOR_DISABLE_CUSTOM_DIMENSIONS_LIMIT"
+# Gen AI attributes whose value should be exempt from truncation
+_GEN_AI_ATTRIBUTES = (
+    "gen_ai.input.messages",
+    "gen_ai.output.messages",
+    "gen_ai.system_instructions",
+    "gen_ai.tool.definitions",
+    "gen_ai.tool.call.arguments",
+    "gen_ai.tool.call.result",
+    "gen_ai.evaluation.explanation",
+)
 
 # cSpell:disable

sdk/monitor/azure-monitor-opentelemetry-exporter/azure/monitor/opentelemetry/exporter/_utils.py

@@ -29,7 +29,7 @@
     _KUBERNETES_SERVICE_HOST,
     _PYTHON_APPLICATIONINSIGHTS_ENABLE_TELEMETRY,
     _WEBSITE_SITE_NAME,
-    AZURE_MONITOR_DISABLE_CUSTOM_DIMENSIONS_LIMIT,
+    _GEN_AI_ATTRIBUTES,
 )
 from azure.monitor.opentelemetry.exporter._constants import (
     _TYPE_MAP,
@@ -351,9 +351,7 @@ def _is_any_synthetic_source(properties: Optional[Any]) -> bool:
 
 # pylint: disable=W0622
 def _filter_custom_properties(properties: Attributes, filter=None) -> Dict[str, str]:
-    disable_custom_dimensions_limit = (
-        environ.get(AZURE_MONITOR_DISABLE_CUSTOM_DIMENSIONS_LIMIT, "").strip().lower() == "true"
-    )
+    max_length = 64 * 1024
     processed_properties: Dict[str, str] = {}
     if not properties:
         return processed_properties
@@ -362,14 +360,13 @@ def _filter_custom_properties(properties: Attributes, filter=None) -> Dict[str,
         if filter is not None:
             if not filter(key, val):
                 continue
-        # Apply truncation/filtering rules
-        # Max key length is 150
+        # Apply truncation rules
+        # Max key length is 150, value is 64 * 1024
         if not key or len(key) > 150 or val is None:
             continue
-        if disable_custom_dimensions_limit:
+        if key in _GEN_AI_ATTRIBUTES:
             processed_properties[key] = str(val)
         else:
-            max_length = 64 * 1024
             processed_properties[key] = str(val)[:max_length]
     return processed_properties
 

sdk/monitor/azure-monitor-opentelemetry-exporter/tests/test_utils.py

@@ -10,6 +10,7 @@
 from opentelemetry.sdk.resources import Resource
 from azure.monitor.opentelemetry.exporter import _utils
 from azure.monitor.opentelemetry.exporter._generated.exporter.models import TelemetryItem
+from azure.monitor.opentelemetry.exporter._constants import _GEN_AI_ATTRIBUTES
 from opentelemetry.sdk.resources import Resource
 from unittest.mock import patch
 
@@ -53,40 +54,51 @@ def test_filter_custom_properties_drops_invalid_entries(self):
         self.assertEqual(filtered["short"], "ok")
         self.assertNotIn("k" * 151, filtered)
 
-    def test_filter_custom_properties_preserves_large_values_after_disable_limit(self):
-        # Ensure values larger than 64KiB are not truncated when the env variable is set to true
-        enable_values = ["true", "True", "TRUE", "TrUE", "  true  "]
-        large_value = "x" * (64 * 1024 + 1000)
-        properties = {"large_key": large_value}
-
-        for env_value in enable_values:
-            with self.subTest(env_value=env_value):
-                with patch.dict(
-                    "azure.monitor.opentelemetry.exporter._utils.environ",
-                    {"AZURE_MONITOR_DISABLE_CUSTOM_DIMENSIONS_LIMIT": env_value},
-                ):
-                    filtered = _utils._filter_custom_properties(properties)
-                    self.assertIn("large_key", filtered)
-                    self.assertEqual(filtered["large_key"], large_value)
-                    self.assertEqual(len(filtered["large_key"]), 64 * 1024 + 1000)
-
-    def test_filter_custom_properties_preserves_large_values_after_enable_limit(self):
-        # Ensure values larger than 64KiB are not truncated when the env variable is set to false/empty/invalid
-        disable_values = ["", "False", "truthy", "89", "fALSE", " "]
+    def test_custom_properties_gen_ai_attributes_not_truncated(self):
+        # All values in _GEN_AI_ATTRIBUTES should not be truncated even when > 64KiB
         large_value = "x" * (64 * 1024 + 1000)
+        properties = {key: large_value for key in _GEN_AI_ATTRIBUTES}
+        filtered = _utils._filter_custom_properties(properties)
+        for key in _GEN_AI_ATTRIBUTES:
+            with self.subTest(key=key):
+                self.assertIn(key, filtered)
+                self.assertEqual(len(filtered[key]), 64 * 1024 + 1000)
+
+    def test_filter_custom_properties_non_gen_ai_truncated_at_64kb(self):
+        # Regular properties exceeding 64KiB should be truncated
         max_length = 64 * 1024
-        properties = {"large_key": large_value}
-
-        for env_value in disable_values:
-            with self.subTest(env_value=env_value):
-                with patch.dict(
-                    "azure.monitor.opentelemetry.exporter._utils.environ",
-                    {"AZURE_MONITOR_DISABLE_CUSTOM_DIMENSIONS_LIMIT": env_value},
-                ):
-                    filtered = _utils._filter_custom_properties(properties)
-                    self.assertIn("large_key", filtered)
-                    self.assertEqual(filtered["large_key"], "x" * max_length)
-                    self.assertEqual(len(filtered["large_key"]), max_length)
+        large_value = "y" * (max_length + 2000)
+        properties = {
+            "span_kind": large_value,
+            "gen_ai.agent.version": large_value,
+            "http.method": large_value,
+            "custom.attribute": large_value,
+        }
+        filtered = _utils._filter_custom_properties(properties)
+        for key in properties:
+            with self.subTest(key=key):
+                self.assertIn(key, filtered)
+                self.assertEqual(len(filtered[key]), max_length)
+
+    def test_filter_custom_properties_mixed_gen_ai_and_regular(self):
+        # Gen AI attributes keep full value, regular ones are truncated
+        max_length = 64 * 1024
+        large_value = "z" * (max_length + 3000)
+        properties = {
+            "gen_ai.input.messages": large_value,
+            "gen_ai.output.messages": large_value,
+            "gen_ai.agent.version": large_value,
+            "span_kind": large_value,
+            "db.statement": large_value,
+        }
+        filtered = _utils._filter_custom_properties(properties)
+        # Gen AI attributes — not truncated
+        self.assertEqual(len(filtered["gen_ai.input.messages"]), max_length + 3000)
+        self.assertEqual(len(filtered["gen_ai.output.messages"]), max_length + 3000)
+        # Regular attributes — truncated
+        self.assertEqual(len(filtered["gen_ai.agent.version"]), max_length)
+        self.assertEqual(len(filtered["span_kind"]), max_length)
+        self.assertEqual(len(filtered["db.statement"]), max_length)
 
     def test_nanoseconds_to_duration(self):
         ns_to_duration = _utils.ns_to_duration