Special handling is needed to ensure we have the transaction ID and Rack ENV. Otherwise the error cannot be matched to a transaction in Sentry. This special behavior is needed because of the underlying EventMachine, which executes code in a different thread/context.
Previously, we were always fetching files, even if not required (e.g., for score runs). Now, we reduce the number of file listings and use a dedicated callback.
* The Faye::WebSocket library will "buffer" some output of the connection and emit the `on :message` events in the order of the messages. However, when a ping is sent while the connection has already been closed, it will emit the `on :close` event immediately and drop all other messages (in that "buffer"). This is problematic for very short running executions that generate a long output (as this will be cut off without a proper exit message sent by Poseidon).
Raising the errors would crash the current thread. As this thread
contains the Eventmachine, that would influence other connections
as well. Attaching the errors to the connection and reading them
after the connection was closed ensures that the thread stays
alive while handling the errors in the main thread of the current
request.
In order to provide an alternative to Poseidon, a strategy for the
DockerContainerPool is added that is used by the runner model.
Co-authored-by: Sebastian Serth <Sebastian.Serth@hpi.de>
Timeouts are now handled correctly and the Runner automatically
creates the execution environment if it could not be found in
Poseidon. The runner is deleted locally if Poseidon returns
a bad request error.
The errors are raised in the runner model and in the runner connection
class. In the submission controller the errors are rescued and,
depending on the error, the status timeout / container depleted is
sent to the client.
The old approach was to require the runner connection. This did
not work anymore with Zeitwerk in Rails 6. @sebastian.serth and I
moved the Connection class in `lib` into the ActiveRecord class
`Runner`. This will also work with future changes like specific
error classes. Furthermore the config was fixed and simplified.
Co-authored-by: Sebastian Serth <Sebastian.Serth@hpi.de>
