autobatch: thread-safe, debounce, max delay and implement Batching

[MichaelMure]

Revamp the venerable autobatch in different ways:

• make it thread safe (test -race is happy)
• deduplicate writes (avoid writing the same key multiple time), which make it act as a debounce
• introduce a maximum delay before the write happen, to avoid keeping the pending writes in memory if the count trigger is not reached
• implement the Batching interface for compatibility with more usecases

[aschmahmann]

@whyrusleeping you might be interested in this.

[MichaelMure]

I can't comment on Query(), so will comment here. Is the intention of flushing in Query() before issuing the query, to make sure that there are no pending ops before sending the query? If so, then I think that's racy (it unlocks after flushing and re-locks before querying).

This is correct, but at the same time Query can be very long lived (I used some hours-long Query before) so I don't think blocking writes during a Query is reasonable.

[guseggert]

I can't comment on Query(), so will comment here. Is the intention of flushing in Query() before issuing the query, to make sure that there are no pending ops before sending the query? If so, then I think that's racy (it unlocks after flushing and re-locks before querying).

This is correct, but at the same time Query can be very long lived (I used some hours-long Query before) so I don't think blocking writes during a Query is reasonable.

Ah that makes sense, thanks :)

[BigLep]

@MichaelMure : I've converted this to a draft. If/once you've incorporated feedback, feel free to publish and we'll happily take a look.

[BigLep]

@MichaelMure : are you going to take this, or should we close this for now?

[MichaelMure]

@BigLep I'm sure there is value in there but I ended up not using any of this in my project as things took a different turn. I suppose we can close.