[yacovm]

A non-validator can receive a finalized Telock of the old epoch at seq k+2, where k is the sealing block, before it indexes k+1. The old epoch is still known at that point, so the Telock is verified and kept in the replication state.

Indexing block k+1, the first block of the new epoch, removes the old epoch. The non-validator then picks the stale Telock as the next seq to commit and tries to verify its finalization QC. The old epoch is gone, so this fails with
"expected epoch to have been validated", and the non-validator halts.

This commit tracks the epoch of the last committed block, restored from storage on startup. When a block of a newer epoch is indexed, every block of an older epoch is purged from the replication state. A stale Telock can then no longer take the place of the new epoch's block at the same seq.

[samliok]

I think you mentioned this to me, but maybe a short summary of the issue would help so I don't need to re-read mythos slop? could we also update the other prs as well

[yacovm]

I think you mentioned this to me, but maybe a short summary of the issue would help so I don't need to re-read mythos slop? could we also update the other prs as well

Done

[samliok]

Indexing block k+1, the first block of the new epoch, removes the old epoch.

I thought indexing K removes the old epoch?

[yacovm]

Indexing block k+1, the first block of the new epoch, removes the old epoch.

I thought indexing K removes the old epoch?

n.removeOldSequencesAndEpochs(md.Seq, md.Epoch) is what removes the old epoch and it can only cause this by being called with the new epoch.

[samliok]

I think there is a simpler way to do this than add to the replication struct + nonvalidator.

What if we just do

epoch, ok := n.epochs[block.BlockHeader().Epoch]
	if !ok {
		n.Logger.Debug("Next sequence to commit is from an unknown epoch", zap.Uint64("epoch", block.BlockHeader().Epoch))
		n.sequenceReplicator.ResendFinalizationRequest(block.BlockHeader().Seq, n.Comm.Validators().NodeIDs())
		return nil
	}

I think this would be the only necessary prod change