Hinted handoff relies upon the VNode entering "inactive" state.

[emauton]

The riak_core_vnode carries an inactivity timeout reset each time it handles a message:
https://github.com/basho/riak_core/blob/develop/src/riak_core_vnode.erl#L277

The timeout is necessary for handoffs of the "hinted" type to occur, via coordination of the VNode:
https://github.com/basho/riak_core/blob/develop/src/riak_core_vnode.erl#L463

... and the VNode manager (using maybe_trigger_handoff):
https://github.com/basho/riak_core/blob/develop/src/riak_core_vnode_manager.erl#L468

The upshot is that if you are sending periodic messages to your application's VNode (with a period less than the inactivity_timeout), hinted handoff never happens.

Sending regular messages is a reasonable thing to do - in my case, I had been periodically exporting statistics that way - and it took a lot of tracing down to understand that the VNode's never becoming inactive was the reason hinted handoff was not happening.

Since riak_core_vnode_manager:maybe_trigger_handoff/4 appears not to be an expensive operation (at least by my reading of the code), I think there's no particularly good reason to rely on inactivity here - I believe this could just be tried regularly in the manager's management_tick loop.

What do you think?

[jonmeredith]

Using the vnode manager is my preferred way of handling it. Periodically the vnode mgr should ping the vnode (using a proxy process so it doesn't block) and ask how many requests it has handled - if the vnode mgr sees that it is inactive and a fallback, it can trigger handoff/shutdown.

Ideally we'd remove any circular dependencies between the vnode and the vnode mgr as it would be very helpful for constructing unit/integration tests if instances of vnodes could be spun up in tests, combined with the Pid form of preflists so that requests can be sent to the vnodes.

[DeadZen]

The underlying question I think is, could handoffs get stalled if a vnode is never considered idle, is a management timer alone enough to determine this reliably, or is perhaps a timer + max # of requests more appropriate or a better option.