I fear that it’s not that simple. noserverunicast=1 is correct, but there also seem to be periods where the dhcpv6 server is legitimately down/unreachable/overloaded for a few hours, causing leases to tick down. At least, my tcpdumps are seeing renews and sometimes eventually rebinds echo into the void for a while every so often. Might be more instances of the T1=0 bug pounding the servers into mush. I am seeing odd bursts of replies come back when it recovers, which would lean towards an overloaded server of some kind.
However, on my end the last extended such outage was wednesday, things have been pretty solid since.
In theory, short issues shouldn’t cause issues though - the assignments are coming back with preferred time 1h, valid time 24h, so the network should keep functioning for a day even after losing all contact with the dhcpv6 server.
Which isn’t what we observed, though, which is strange. We’ve seen issues an hour after the last proper reply.
After a >1h outage, my logs say
2018-08-22T09:19:11+02:00 warning odhcpd[2110]: A default route is present but there is no public prefix on br-lan thus we don't announce a default route!
(beware: Timestamps in this log seem a tad schizophrenic, this is likely actually 11:19+0200)
around the same time, the DHCPv6 script hook I set up reports that dhcpv6 is faithfully reporting a prefix, with preferred time expired but valid time still going.
========== DHCPv6 eth1 ra-updated ==========
Wed Aug 22 11:19:10 CEST 2018
RA_ROUTES=::/0,fe80::523d:e5ff:fe16:55ff,1799,512 2a02:168:2000:xxxx::/64,,2591999,256
PREFIXES=2a02:168:xxxx::/48,0,82046
Very strange. Unfortunately, this is probably going to send us into the bowels of netifd and odhcpd.
Either netifd is prematurely deassigning the address, or odhcpd’s detection of whether there is a valid prefix is broken. Ultimately, there’s a bug where something is effectively enforcing preferred lifetimes as if they were validity lifetimes, which is just all sorts of wrong.