head change coalescer

[vyzo]

No description provided.

[arajasek]

Logic looks good to me!

[arajasek]

Nit: Can these be changed to

if !cancel {
    newRevert = append(newRevert, ts)
}

[whyrusleeping]

@vyzo so the code here appears to work pretty well, but I'm still seeing the mempool triggering recent state calls. Maybe we can just bump the timeout number? Will try now...

[whyrusleeping]

I bumped the delay up to 3 seconds, and ran this for a bit. Its better, but i'm still seeing state computations being triggered by the mempool (and if this were fully fixed, i'd expect to see them only be triggered by the FundManager for now)

[whyrusleeping]

To check this, I just grabbed a random 120 second profile of the lotus node while its fully caught up

[vyzo]

We can try to add some more smarts -- for instance, when it's time to trigger the callback check to see if the last coalesce was within some bound (say 100ms) -- and if that's the case wait some more.
This might succeed in coalescing all the incoming blocks in the beginning of the epoch.

[vyzo]

Also, some state computations are unavoidable, we do have to compute at some point, we would just like to avoid the unnecessary ones :)

[vyzo]

Added some more smarts to the coalescing algorithm, controlled by 3 parameters: minDelay, maxDelay, and mergeInterval.

When the first head chance (since the last dispatch) is received, the coalescer schedules a timer for minDelay to coalesce more head changes.
When the timer fires, it checks the time of the last coalesce; if it is within the mergeInterval, then it extends the coalesce time to coalesce some more.
Finally, the delivery of head change is not delaye more than maxDelay.

For the mpool, I have set minDelay=2s, maxDelay=6s, and mergeInterval=1s.
Hopefully this will capture the common case of 6s or less for block delivery.

[arajasek]

New version looks good to me