Handle OutOfMemory runner errors with gVisor

app/controllers/submissions_controller.rb

@@ -99,8 +99,7 @@ def render_file
     end
   end
 
-  # rubocop:disable Metrics/CyclomaticComplexity
-  def run
+  def run # rubocop:disable Metrics/PerceivedComplexity, Metrics/CyclomaticComplexity
     # These method-local socket variables are required in order to use one socket
     # in the callbacks of the other socket. As the callbacks for the client socket
     # are registered first, the runner socket may still be nil.
@@ -199,12 +198,6 @@ def run
           end
         stream = @testrun[:status] == :ok ? :stdout : :stderr
         send_and_store client_socket, {cmd: :write, stream:, data: "#{exit_statement}\n"}
-        if exit_code == 137
-          send_and_store client_socket, {cmd: :status, status: :out_of_memory}
-          @testrun[:status] = :out_of_memory
-        end
-
-        # The client connection will be closed once the file listing finished.
       end
 
       runner_socket.on :files do |files|
@@ -213,30 +206,33 @@ def run
           js_tree = FileTree.new(downloadable_files).to_js_tree
           send_and_store client_socket, {cmd: :files, data: js_tree}
         end
-
-        close_client_connection(client_socket)
       end
     end
     @testrun[:container_execution_time] = durations[:execution_duration]
     @testrun[:waiting_for_container_time] = durations[:waiting_duration]
   rescue Runner::Error::ExecutionTimeout => e
     send_and_store client_socket, {cmd: :status, status: :timeout}
-    close_client_connection(client_socket)
     Rails.logger.debug { "Running a submission timed out: #{e.message}" }
     @testrun[:status] ||= :timeout
     @testrun[:output] = "timeout: #{@testrun[:output]}"
     extract_durations(e)
+  rescue Runner::Error::OutOfMemory => e
+    send_and_store client_socket, {cmd: :status, status: :out_of_memory}
+    Rails.logger.debug { "Running a submission caused an out of memory error: #{e.message}" }
+    @testrun[:status] ||= :out_of_memory
+    @testrun[:exit_code] ||= 137
+    @testrun[:output] = "out_of_memory: #{@testrun[:output]}"
+    extract_durations(e)
   rescue Runner::Error => e
     # Regardless of the specific error cause, we send a `container_depleted` status to the client.
     send_and_store client_socket, {cmd: :status, status: :container_depleted}
-    close_client_connection(client_socket)
     @testrun[:status] ||= :container_depleted
     Rails.logger.debug { "Runner error while running a submission: #{e.message}" }
     extract_durations(e)
   ensure
+    close_client_connection(client_socket)
     save_testrun_output 'run'
   end
-  # rubocop:enable Metrics/CyclomaticComplexity:
 
   def score
     client_socket = nil
@@ -256,14 +252,13 @@ def score
     client_socket&.send_data(JSON.dump(@submission.calculate_score))
     # To enable hints when scoring a submission, uncomment the next line:
     # send_hints(client_socket, StructuredError.where(submission: @submission))
-    kill_client_socket(client_socket)
   rescue Runner::Error => e
     extract_durations(e)
     send_and_store client_socket, {cmd: :status, status: :container_depleted}
-    kill_client_socket(client_socket)
     Rails.logger.debug { "Runner error while scoring submission #{@submission.id}: #{e.message}" }
     @testrun[:passed] = false
   ensure
+    kill_client_socket(client_socket)
     save_testrun_output 'assess'
   end
 
@@ -290,14 +285,13 @@ def test
 
     # The score is stored separately, we can forward it to the client immediately
     client_socket&.send_data(JSON.dump(@submission.test(@file)))
-    kill_client_socket(client_socket)
   rescue Runner::Error => e
     extract_durations(e)
     send_and_store client_socket, {cmd: :status, status: :container_depleted}
-    kill_client_socket(client_socket)
     Rails.logger.debug { "Runner error while testing submission #{@submission.id}: #{e.message}" }
     @testrun[:passed] = false
   ensure
+    kill_client_socket(client_socket)
     save_testrun_output 'assess'
   end
 

app/errors/runner/error.rb

@@ -24,6 +24,8 @@ class UnexpectedResponse < Error; end
 
     class WorkspaceError < Error; end
 
+    class OutOfMemory < Error; end
+
     class Unknown < Error; end
   end
 end

app/models/runner.rb

@@ -137,6 +137,9 @@ def execute_command(command, privileged_execution: false, raise_exception: true)
     rescue Runner::Error::ExecutionTimeout => e
       Rails.logger.debug { "Running command `#{command}` timed out: #{e.message}" }
       output.merge!(status: :timeout, container_execution_time: e.execution_duration)
+    rescue Runner::Error::OutOfMemory => e
+      Rails.logger.debug { "Running command `#{command}` caused an out of memory error: #{e.message}" }
+      output.merge!(status: :out_of_memory, container_execution_time: e.execution_duration)
     rescue Runner::Error::RunnerNotFound => e
       Rails.logger.debug { "Running command `#{command}` failed for the first time: #{e.message}" }
       try += 1

lib/runner/connection.rb

@@ -162,8 +162,14 @@ def on_close(event, sentry_span)
         @strategy.destroy_at_management
         @error = Runner::Error::ExecutionTimeout.new('Execution exceeded its time limit')
       when :terminated_by_codeocean, :terminated_by_management
-        @exit_callback.call @exit_code
-        list_filesystem
+        # Poseidon (without gVisor) and DockerContainerPool do not handle memory limits explicitly.
+        # Instead, they signal that the program was terminated with exit code 137 (128 + 9).
+        if @exit_code == 137
+          @error = Runner::Error::OutOfMemory.new('Execution exceeded its memory limit')
+        else
+          @exit_callback.call @exit_code
+          list_filesystem
+        end
       when :terminated_by_client, :error
         @strategy.destroy_at_management
       else # :established
@@ -223,6 +229,14 @@ def handle_stderr(event)
   end
 
   def handle_error(event)
+    # Poseidon (with gVisor enabled!) sends an error message when the execution exceeds its memory limit.
+    # This is not an error in the sense of the runner management but rather a message.
+    # We handle it here to avoid the error handling in the default case.
+    if event[:data] == 'the allocation was OOM Killed'
+      @error = Runner::Error::OutOfMemory.new('Execution exceeded its memory limit')
+      return
+    end
+
     # In case of a (Nomad) error during execution, the runner management will notify us with an error message here.
     # This shouldn't happen too often and can be considered an internal server error by the runner management.
     # More information is available in the logs of the runner management or the orchestrator (e.g., Nomad).