@chrys is that what Microsoft referred to as ‘black-hole’ detect or have I misremembered? (Wondering also about a bad gateway [where alternatives are known] - if perhaps similar phrase was used in connection with something roughly like that as well)
I had assumed that IPv6 might be more clued up and I was aware of the 1280 byte PDU size. I needed to decide what to do about low MTU (1440) when my Firebrick switches to the stupid 3G dongle during failover. AA misled me about full 1500 byte packets on 3G/4G, and it isn’t the dongle either because my iPad also has reduced MTU on the AA / AQL / Three 4G service where the MTU is 1450 (Don’t know why it’s not 1440 there, guessing something about protocol stack alternatives, but how? And why?).
So I decided to keep the MTU for IPv4 at 1500 and do nothing about the failover case, so that when it switches over, traffic belonging to existing flows will just get fragmented and hopefully new TCP connections will use a reduced MSS. This is a dubious plan. It favours the normal non-failover case, which is 99.99% of the time. On the other hand, IPv6 now uses a reduced 1408 byte MTU all the time. I thought that this is safe because IPv6 systems have clue and so why not. When MTU is suddenly reduced because of failover, IPv6 cannot get fragmented at intermediate node anyway so it would all get dropped which would not be good. I don’t want to just hope that systems adapt to new IPv6 MTU halfway through, so I thought keep MTU low all the time for IPv6.
AA’s example config for failover suggests permanently reduced MTU for IPv4 and IPv6 iirc. But I went for a 1500 byte normal-condition IPv4 MTU because of two reasons (i) very slightly better efficiency - almost nothing in it as 1500 happens to be a very good number since 1500+32 bytes = my overhead is almost optimal, nearly a multiple of 48 bytes for ATM, and aside from ATM, given a free choice, the maximum possible size is of course always the most L3+L4-header efficient, and (ii) no risk of these ancient legendary reduced-MTU problems in the normal case (and nothing I can do about the failover case).