Nacks are treated as exceptions / failures in the subscriber

[mbrancato]

In the pubsub subscribe() function, there is no clean way to negatively acknowledge a message. The only way to do it inside the subscriber is to either run your own nacker() process, or to raise an exception, which will go to the default nacker(). The problem is that the library treats exceptions as failures, even though when processing a message, there are a number of possible reasons to nack a message due to transient / retryable conditions.

        except Exception as e:
            if nack_queue:
                await nack_queue.put(message.ack_id)
            log.exception('application callback raised an exception: %s', e)
            metrics_client.increment('pubsub.consumer.failed')
            metrics.CONSUME.labels(outcome='failed').inc()

It seems like it might be good to have either specific exception types, or the ability to configure how the exceptions are handled. Explicit ack() and nack() methods would be great on the SubscriberMessage class when used with subscribe(). That would decouple acking / nacking from potential exceptions, which could then all be treated as errors.

Right now I'm raising an exception on failures (for various reasons), but that then creates excessive exception logging that I don't need / want.

Related to #516

[igor-kantorski-spts]

+1, recently written small wrapper around pubsub and encountered WARNING severity log messages when raising an exception - even though raising an exception is the only explicit way to nack() a message.

For me personally, it is more than enough to have a specific exception which is not treated as an error in application (do not trigger warning) - instead is perceived as a nack request by a library (class NackException(Exception): ...).

I can submit a PR that implements this, but first it would be nice to have some feedback from maintainers.

Please note that although we can always call SubscriberClient.modify_ack_deadline() - this differs from the behavior of handling an exception - which collects nacks in a window defined by a nack_window parameter. It would be great if we could nack, put it into nack queue and still do not trigger warnings.