sentrivana · Nov 11, 2026 · Oct 26, 2026 · Oct 27, 2026 · Oct 26, 2026 · Oct 28, 2026
@@ -47,7 +47,7 @@ Looking to upgrade from Sentry SDK 2.x to 3.x? Here's a comprehensive list of wh
 - The `RustTracingIntegration` no longer includes the values of tracing fields if `send_default_pii` is enabled. Pass `include_tracing_fields=True` to the integration to include them.
 - Whether the Dramatiq integration attaches message bodies to events is now controlled by `data_collection`'s `queues` setting instead of `http_bodies`.
 - The Celery integration no longer sets the status of task spans to error when a task raises a control flow exception (`Retry`, `Ignore`, or `Reject`).
-- The boto3 integration now names spans `<ServiceID>.<Operation>` (for example, `S3.GetObject` or `DynamoDB.GetItem`), following OpenTelemetry semantic conventions. Previously, spans were named `aws.<service-id>.<Operation>` (for example, `aws.s3.GetObject`). Its spans, including the span for reading a streamed response body, now also have the origin `auto.aws.boto3` instead of `auto.http.boto3`. SQS messaging calls are the exception: `SendMessage` and `SendMessageBatch` are named `send <queue name>`, `ReceiveMessage` is named `receive <queue name>`, and `DeleteMessage` and `DeleteMessageBatch` are named `settle <queue name>`. Update any alerts, dashboards, or `ignore_spans` rules that match on the old names or origin.
+- The boto3 integration now names spans `<ServiceID>.<Operation>` (for example, `S3.GetObject` or `DynamoDB.GetItem`), following OpenTelemetry semantic conventions. Previously, spans were named `aws.<service-id>.<Operation>` (for example, `aws.s3.GetObject`). Its spans, including the span for reading a streamed response body, now also have the origin `auto.aws.boto3` instead of `auto.http.boto3`. SQS and SNS messaging calls are the exception: SQS `SendMessage` and `SendMessageBatch` are named `send <queue name>`, `ReceiveMessage` is named `receive <queue name>`, and `DeleteMessage` and `DeleteMessageBatch` are named `settle <queue name>`; SNS `Publish` and `PublishBatch` are named `send <topic name>`. Update any alerts, dashboards, or `ignore_spans` rules that match on the old names or origin.
 - Enabling an AI agent integration (LangChain, OpenAI Agents, or Pydantic AI) no longer disables the OpenAI, Anthropic, and Google GenAI integrations. Instead, those integrations don't create chat spans for requests made from within an agent framework's chat span, but still create other spans, such as embedding spans. The LangChain integration no longer instruments embeddings directly; embedding spans are now created by the integration of the underlying client library.
 - Whether the MCP integration records tool arguments, tool results, and prompt contents is now controlled by `data_collection`'s new `mcp` setting (`{"inputs": bool, "outputs": bool}`, both `True` by default) instead of the `gen_ai` setting.
 - With `handled_tool_call_exceptions=True` (the default), the Pydantic AI integration now reports an error for every failed tool argument validation attempt, including retries. Previously, only one error was reported per tool call.
@@ -77,7 +77,7 @@ Span ops are now a smaller, framework-neutral set.
 - The Django integration now emits `view.render` instead of `view.response.render`, `view` instead of `view.authenticate`, and `function` instead of `event.django`.
 - The Django, Starlette, Starlite, and Litestar integrations now emit `middleware` instead of `middleware.django`, `middleware.starlette`, `middleware.starlette.receive`, `middleware.starlette.send`, `middleware.starlite`, `middleware.starlite.receive`, `middleware.starlite.send`, `middleware.litestar`, `middleware.litestar.receive`, and `middleware.litestar.send`.
 - The ARQ, Celery, Django, Dramatiq, Huey, RQ, and Ray integrations now emit `queue.publish` and `queue.process` instead of `queue.submit.arq`, `queue.task.arq`, `queue.submit.celery`, `queue.task.celery`, `queue.submit.django`, `queue.task.dramatiq`, `queue.submit.huey`, `queue.task.huey`, `queue.task.rq`, `queue.submit.ray`, and `queue.task.ray`.
-- The Boto3 integration now emits `rpc` instead of `http.client` for AWS API calls. SQS messaging calls emit `queue.publish` (`SendMessage`, `SendMessageBatch`), `queue.receive` (`ReceiveMessage`), and `queue.settle` (`DeleteMessage`, `DeleteMessageBatch`) instead.
+- The Boto3 integration now emits `rpc` instead of `http.client` for AWS API calls. SQS messaging calls emit `queue.publish` (`SendMessage`, `SendMessageBatch`), `queue.receive` (`ReceiveMessage`), and `queue.settle` (`DeleteMessage`, `DeleteMessageBatch`) instead, and SNS `Publish` and `PublishBatch` emit `queue.publish`.
 
 #### GenAI span attributes
 

@@ -489,6 +489,12 @@ class SPANDATA:
     Example: "https://sqs.us-east-1.amazonaws.com/123456789012/orders"
     """
 
+    AWS_SNS_TOPIC_ARN = "aws.sns.topic.arn"
+    """
+    The ARN of the AWS SNS topic the request refers to.
+    Example: "arn:aws:sns:us-east-1:123456789012:alerts"
+    """
+
     CACHE_HIT = "cache.hit"
     """
     A boolean indicating whether the requested data was found in the cache.

@@ -1,36 +1,131 @@
 from typing import TYPE_CHECKING
 
+import sentry_sdk
+from sentry_sdk.consts import OP, SPANDATA
 from sentry_sdk.integrations.boto3._services.base import _ServiceExtension
-from sentry_sdk.integrations.boto3._utils import _get_aws_arn_attributes
+from sentry_sdk.integrations.boto3._utils import (
+    _get_aws_arn_attributes,
+    _get_batch_error_type,
+)
+from sentry_sdk.traces import SpanStatus
 from sentry_sdk.utils import capture_internal_exceptions
 
 if TYPE_CHECKING:
-    from typing import Any
+    from typing import Any, Optional
 
     from sentry_sdk._types import Attributes
     from sentry_sdk.integrations.boto3._context import AwsCallContext
     from sentry_sdk.traces import Span
 
+_PUBLISH_OPERATIONS = ("Publish", "PublishBatch")
+
 
 class _SNSExtension(_ServiceExtension):
+    def get_span_op(self, ctx: "AwsCallContext") -> "Optional[str]":
+        if ctx.operation_name in _PUBLISH_OPERATIONS:
+            return OP.QUEUE_PUBLISH
+        # defaults to `rpc` for other operations.
+        return None
+
+    def get_span_name(
+        self, ctx: "AwsCallContext", attributes: "Attributes"
+    ) -> "Optional[str]":
+        if ctx.operation_name not in _PUBLISH_OPERATIONS:
+            return None
+
+        destination = attributes.get(SPANDATA.MESSAGING_DESTINATION_NAME)
+        if destination is None and SPANDATA.SERVER_ADDRESS in attributes:
+            destination = f"{attributes[SPANDATA.SERVER_ADDRESS]}:{attributes[SPANDATA.SERVER_PORT]}"
+
+        # for example: `send <topic name>` when `messaging.operation.type=send` (always send).
+        # https://opentelemetry.io/docs/specs/semconv/messaging/messaging-spans/#span-name
+        return f"send {destination}" if destination else "send"
+
     def get_request_attributes(self, ctx: "AwsCallContext") -> "Attributes":
         attributes: "Attributes" = {}
+
+        # botocore validates params after this hook, so a malformed value must only
+        # lose the attributes read from it, not the span.
+        target_arn = ctx.params.get("TargetArn")
+        endpoint_arn = ctx.params.get("EndpointArn")
+        topic_arn = None
         with capture_internal_exceptions():
+            # `TargetArn` can be either a topic or a platform endpoint; a topic ARN
+            # should never contain ":endpoint/".
+            if target_arn and ":endpoint/" in target_arn:
+                endpoint_arn, target_arn = target_arn, None
+
+            # `TopicArn` and `ResourceArn` are always topic ARNs.
+            topic_arn = (
+                ctx.params.get("TopicArn")
+                or ctx.params.get("ResourceArn")
+                or target_arn
+            )
+
+            if topic_arn:
+                attributes[SPANDATA.AWS_SNS_TOPIC_ARN] = topic_arn
+                # `arn:aws:sns:eu-north-1:123456789012:alerts` -> `alerts`
+                topic_name = topic_arn.split(":")[-1]
+                attributes[SPANDATA.MESSAGING_DESTINATION_NAME] = topic_name
+            elif ctx.operation_name == "CreateTopic" and "Name" in ctx.params:
+                attributes[SPANDATA.MESSAGING_DESTINATION_NAME] = ctx.params["Name"]
+
+        if ctx.operation_name in _PUBLISH_OPERATIONS:
             attributes.update(
-                _get_aws_arn_attributes(
-                    (
-                        ctx.params.get("EndpointArn"),
-                        ctx.params.get("PlatformApplicationArn"),
-                        ctx.params.get("SubscriptionArn"),
-                        ctx.params.get("TargetArn"),
-                        ctx.params.get("TopicArn"),
-                    ),
-                    ("sns",),
-                )
+                {
+                    SPANDATA.SENTRY_KIND: "producer",
+                    SPANDATA.MESSAGING_SYSTEM: "aws.sns",
+                    SPANDATA.MESSAGING_OPERATION_TYPE: "send",
+                    SPANDATA.MESSAGING_OPERATION_NAME: "publish",
+                }
             )
+
+            # phone numbers and device endpoints are PII.
+            if "PhoneNumber" in ctx.params or endpoint_arn:
+                attributes[SPANDATA.MESSAGING_DESTINATION_NAME] = "[Filtered]"
+            if "PublishBatchRequestEntries" in ctx.params:
+                with capture_internal_exceptions():
+                    attributes[SPANDATA.MESSAGING_BATCH_MESSAGE_COUNT] = len(
+                        ctx.params["PublishBatchRequestEntries"]
+                    )
+
+        # if scrubbing fails, drop all ARN attributes rather than send a raw
+        # endpoint or subscription ARN.
+        with capture_internal_exceptions():
+            # endpoints and subscriptions identify a single device or subscriber.
+            subscription_arn = ctx.params.get("SubscriptionArn")
+            if not sentry_sdk.get_client().options["data_collection"]["user_info"]:
+                if endpoint_arn:
+                    endpoint_arn = endpoint_arn.rsplit("/", 1)[0].replace(
+                        ":endpoint/", ":app/", 1
+                    )
+                if subscription_arn:
+                    subscription_arn = subscription_arn.rsplit(":", 1)[0]
+
+            arns = (
+                ctx.params.get("PlatformApplicationArn"),
+                topic_arn,
+                endpoint_arn,
+                subscription_arn,
+            )
+            attributes.update(_get_aws_arn_attributes(arns, ("sns",)))
         return attributes
 
     def on_success_response(
         self, ctx: "AwsCallContext", response: "Any", span: "Span"
     ) -> None:
+        attributes: "Attributes" = {}
+
+        # the call already succeeded; an unexpected response shape, e.g. from a
+        # `Stubber` or an SNS-compatible service, must only lose these attributes.
+        with capture_internal_exceptions():
+            if ctx.operation_name == "Publish" and "MessageId" in response:
+                attributes[SPANDATA.MESSAGING_MESSAGE_ID] = response["MessageId"]
+
+            error_type = _get_batch_error_type(response)
+            if error_type is not None:
+                span.status = SpanStatus.ERROR
+                attributes[SPANDATA.ERROR_TYPE] = error_type
+
+        span.set_attributes(attributes)
         span.end()