Transform FC module internals into DAG

[jangko]

No more BranchRef, only BlockRef.
Simpler algorithm in many places.
Pruning leaves become more uniform among branches.

[jangko]

Tested on hoodi during syncing and chain following. No incident.

[arnetheduck]

this algorithm can be simplified based on the observation that block numbers are monotonically increasing - basically, when we've identified the finalized block, we walk the parent chain and compare block numbers - if the block number of the examined block is lower than that of the finalized block but is not the finalized block, we know that the branch is stale because the block where it meets the chain must be older than finalized.

proc isRelevantBranch(head, finalized): bool =
  var blk = head
  while blk != finalized:
    if blk.number < finalized.number:
      return false
    blk = blk.parent
  true

[arnetheduck]

notably, we only need to run isRelevantBranch for branches that are not the canonical head right now

[jangko]

just tested, it does not work for branches with head lower than the new finalized.

e.g. old fin = A3. stale chain B5-B8. chain A continue from A4 to A10. new fin = A9.
The simplified algorithm basically not aware of B.

[arnetheduck]

In your example, the starting point is the list of heads before fin was updated to A9, ie [A10, B8] .. then you do:

var toprune: seq[blockref]
for head in fc.heads:
  if head == canonical:
   continue
  if not isReachable(head, newfin):
    toprune.add(head)

b will hit the if blk.number < fin.number condition on first iteration and will be removed

[jangko]

Oops, sorry. I mean not about the head. But how to stop the pruning? When we move backward from head of a stale branch, there is no marker of a junction/forking point.

[jangko]

int eth1 we still need parent to calculate base movement. So there is no significant difference with color algorithm.

[arnetheduck]

well, the base movement computation happens first - you establish the new base, then you clear the parents then you prune the branches - the difference is that you don't need a color field and you don't need to "uncolor" things after you're done - that's the simplification.

[arnetheduck]

btw, clearing the parent has the added benefit that the GC can release the structure earlier as well

[jangko]

not all finalized blocks are written to database,

[arnetheduck]

aah, right - hm - I guess one could use a finalized bool to avoid the "uncoloring" in that case - while still a simplification, maybe that's not as attractive.