I love that one:

It's not a bug, it's an optimization opportunity!

It would be nice to have a comparison on the same machine: pure replay vs replay with full validation vs sync.

I will I've added that to my long long TODO list.

Nah, I'll start it right away.

Starting replay from scratch (using existing, fresh block_log and block_log.artifacts):

~/bin/hived-1.27.6 -d /home/hive/datadir --set-benchmark-interval=100000 --exit-before-sync --replay 2>&1 | ts -s "%s" | tee -ai benchmark-hived-v1.27.6-replay.log

Once it's finished, I will repeat it with added --validate-during-replay.
At that time, I will be able to compare it to my old sync run (performed on the same machine up to block 79795165 with version 1.27.4). But I guess it would be great to take the opportunity to compare the exact same version up to the same block height, which would be around block 87300000.

Extrapolating from the previous sync run, which completed 79.7M in 33 hours, hopefully we can complete the sync on that machine in 36.5 hours.
Seems like I will be able to present results by the end of next week.

And just to be clear, for those tests I'm using minimalistic (except for the witness plugin) configuration:

backtrace = yes

plugin = witness

shared-file-dir = "/run/hive"
shared-file-size = 28G

flush-state-interval = 0

p2p-endpoint = 0.0.0.0:2001