Prysm: Implement Precise Ticker for Slot Interval

[rawfalafel]

Background
In the current spec, new blocks are allowed to advance the chain every slotDuration seconds. As a concrete example, if the genesis block is published at time T, the first block can be published at time T+slotDuration, the second block time T+2*slotDuration, etc. Prysm currently has a naive implementation that uses the standard library Ticker, but it has two issues.

1. The clock doesn't tick precisely at T+N*slotDuration, where N is a non-negative integer. This means that the ticker will be "off beat" from the correct beat, even if the interval length is correct.
2. The clock may drift over time. The ticker uses a monotonic clock, meaning that it won't adjust to miniscule inaccuracies in the interval length. This is an issue if we expect the beacon chain to run for days/weeks/months.

Requirements
Replace the current naive ticker with a ticker that addresses both issues above. It's probably best to confirm that clock drift is an issue, although this thread seems to confirm that it is. The thread also contains a sample solution by Russ Cox -- that may be helpful :smiley:

[danielschonfeld]

@rawfalafel are you referring to SLOT_DURATION instead of cycleLength ?

[rawfalafel]

@rawfalafel are you referring to SLOT_DURATION instead of cycleLength ?

@danielschonfeld yeah, thanks!

_{Sent with GitHawk}

[danielschonfeld]

I am no expert on this, but from reading Russ Cox's comments and code examples, it seems to me that both him and the original OP come to agreement that there is no drift problem with Go or their ticker implementation. It seems that the problem is between the virtualized environment (Azure VM in this case) and the real world clock.

Russ also goes on to say that Go's implementation will generally stay loyal to the start time its aligned to.

Having that said, raises two questions:

1) Do you agree with my assertion? If not, please explain
2) If you do agree, the next question is whether or not we should go the extra length to account and adjust for virtualized environments unpredictability?

[rawfalafel]

From the thread

It does seem that for whatever reason your VM's monotonic timer is running (1+2/99)X faster than wall time, assuming wall time is correct

We're also making the assumption that the wall time is always correct, but the monotonic clock might be too fast or too slow.

Russ also goes on to say that Go's implementation will generally stay loyal to the start time its aligned to.

Right. Go's ticker stays loyal to the monotonic clock, even though the monotonic clock isn't aligned with the wall clock.

I realized I left out an important detail. We're assuming that the wall clock is always in sync with every other node's wall clock using NTP server. We can verify this by occasionally comparing time.Now() with NTP's current time and failing if the difference is over some threshold.

[rauljordan]

Could be useful perhaps to check NTP at each cycle transition as that is a sensitive point for all beacon clients

[rauljordan]

Closed by #635