http-netty: Add phantom reference version of watchdog filters #3084
Conversation
Motivation The current watchdog filters are not exact: they don't account for responses that get drained later than the cleaner expects which can result in false positives. Because of this they can't proactively drain the responses. Modifications Add a PhantomReference variant of the watchdog filters. This filter uses a PhantomReference based implementation to be certain the response has been abandoned. Because we can now be certain, we can also proactively drain the response.
| } | ||
| }; | ||
| // This implementation uses an atomic reference stored on the context to track when requests have been leaked. | ||
| // It is efficient and doesn't depend on GC, but may be prone to false positives since its possible that the |
There was a problem hiding this comment.
As we discussed offline:
HttpMessageDiscardWatchdogClientFilter logs only in 2 use-cases:
- When we detect a new retry but the previous response was not drained. We expect that previous response payload is always drained before the retry is initiated.
- When we detect
onErrorat the outer client filter (close to the user) after we received response meta-data in the inner connection filter (close to channel).
In both cases the previous response is not available anymore and can not be drained later. The false positives are unlikely.
The PhantomReference approach might be helpful to cover more use-cases, like when users use streaming responses, successfully got the response metadata from the client, but then forgot to drain it later in the users scope. However, having 2 different approaches will increase the runtime cost (watchdog already adds visible overhead). On the other hand, removing watchdog and replacing it with PhantomReference at NettyChannelPublisher level will not allow us distinguish (1) from (2). Idk... Let's chat more about it on the team sync.
Motivation
The current watchdog filters are not exact: they don't account for responses that get drained later than the cleaner expects which can result in false positives. Because of this they can't proactively drain the responses.
Modifications
Add a PhantomReference variant of the watchdog filters. This filter uses a PhantomReference based implementation to be certain the response has been abandoned. Because we can now be certain, we can also proactively drain the response.