aws-sdk-lambda: improve error handling

use separate error handling for errors raised by the gem itself and
errors raised by the agent.

Errors raised by the gem need to be noticed by the agent (sent to New
Relic) and re-raised to the user's app.

Errors raised by the agent need to be logged but not impact the
returning of the successful gem invocation response.

Co-authored-by: Kayla Reopelle <87386821+kaylareopelle@users.noreply.github.com>

lib/new_relic/agent/instrumentation/aws_sdk_lambda/instrumentation.rb

@@ -31,13 +31,13 @@ def with_tracing(action, *args)
       # prevent additional instrumentation for things like Net::HTTP from
       # creating any segments that may appear as redundant / confusing
       NewRelic::Agent.disable_all_tracing do
+        response = NewRelic::Agent::Tracer.capture_segment_error(segment) { yield }
         begin
-          response = yield
           process_response(response, segment)
-          response
         rescue => e
-          NewRelic::Agent.notice_error(e)
-          raise
+          NewRelic::Agent.logger.error("Error processing aws-sdk-lambda invocation response (action = #{action}): #{e}")
+        ensure
+          response
         end
       end
     ensure

test/multiverse/suites/aws_sdk_lambda/aws_sdk_lambda_instrumentation_test.rb

@@ -38,16 +38,47 @@ def test_client_call_raises_an_exception
     in_transaction do |txn|
       client = Aws::Lambda::Client.new(region: REGION)
       client.config.account_id = AWS_ACCOUNT_ID
-      def client.process_response(*_args); raise 'kaboom'; end
 
-      assert_raises(RuntimeError) { client.invoke(function_name: 'Invoke-Me-And-Explode') }
+      # We need the aws-sdk-lambda client's `yield` to raise an exception for
+      # this test, but that's tricky given that we're monkeypatching the very
+      # client invocation method under test. So let's trust that
+      # capture_segment_error is thoroughly tested elsewhere and stub it to
+      # do its raise-and-notice-an-error thing in a predictable way so that the
+      # aws-sdk-lambda specific stuff can be tested.
+      capture_proc = proc do |segment|
+        error = RuntimeError.new('kaboom')
+        segment.notice_error(error)
+        raise error
+      end
+
+      NewRelic::Agent::Tracer.stub(:capture_segment_error,
+        capture_proc,
+        [NewRelic::Agent::Transaction::ExternalRequestSegment]) do
+        assert_raises(RuntimeError) { client.invoke(function_name: 'Invoke-Me-And-Explode') }
+      end
+
       noticed_error = lambda_segment(txn).noticed_error
 
       assert_equal 'kaboom', noticed_error.message
       assert_equal 'RuntimeError', noticed_error.exception_class_name
     end
   end
 
+  def test_errors_in_processing_the_invocation_response_are_logged_but_not_raised_to_the_user_app
+    in_transaction do |txn|
+      client = Aws::Lambda::Client.new(region: REGION)
+      client.config.account_id = AWS_ACCOUNT_ID
+      def client.process_response(*_args); raise 'kaboom'; end
+
+      NewRelic::Agent.stub(:logger, NewRelic::Agent::MemoryLogger.new) do
+        client.invoke(function_name: 'Invoke-Me-And-Only-The-Agent-Explodes')
+
+        assert NewRelic::Agent.logger.messages.any? { |m| m[1][0].match?(/^Error processing aws-sdk-lambda/) },
+          'Expected to find an error message in the agent logs'
+      end
+    end
+  end
+
   def test_client_response_indicates_an_unhandled_function_error
     function_error = 'Unhandled'
     message = 'oh no'