Hi Lars,

On Mon, Sep 29, 2025 at 06:38:18AM +0300, Lars Eggert wrote:
> Hi,
> 
> pitch for a discussion at 124.
> 
> [1]https://radar.cloudflare.com/ and similar stats have had H3 around 30% for 
> a
> few years now, with little changes since the first quichbram up to that level.
> 
> Topic: why is that and is there anything the WG or IETF can do to change it
> (upwards, of course)?

FWIW I've always suspected that the initial network reachability figures
for UDP were overestimated, and this could count for a significant part.
When I look at the stats for haproxy.org, and only focus on CSS downloads
(to eliminate most bots and git clones), I'm seeing the following
percentages by day for H3 accesses:

  20250901 43
  20250902 45
  20250903 41
  20250904 39
  20250905 44
  20250906 45
  20250907 51 Sat
  20250908 45 Sun
  20250909 41
  20250910 40
  20250911 38
  20250912 41
  20250913 43
  20250914 51 Sat
  20250915 46 Sun
  20250916 46
  20250917 39
  20250918 42
  20250919 38
  20250920 45
  20250921 47 Sat
  20250922 45 Sun
  20250923 42
  20250924 47
  20250925 55
  20250926 42
  20250927 41
  20250928 47

The averages are counted over roughly 1000-1300 visits on week days
and 500-700 on week-ends. But when graphed, it's clear that there's
a marked increase of the ratio on week-ends. This makes me think that
it works better from home than from the work place, most likely due to
local security policies that mandate the use of a proxy, and possibly
also due to installed local agents (anti-malware etc) which can only
inspect H1/H2 and probably just block H3 until they implement it. And
that might even be true for end-user deployments of comparable software.
I'm observing the same principle with IPv6 vs IPv4 BTW.

Please note that at the beginning of WebSocket and H2 several of us
on the server side had to deal with bugs repors apparently coming
from the client which revealed to be caused by such locally hosted
intermediary software not coping well with the new protocol, so I
don't see why this shouldn't be similar here. And if we add to this
that QUIC is somewhat harder to implement, we can expect some delay
before such intermediary software works fine (if ever at all). That's
probably not the sole cause though.

Regards,
Willy

Reply via email to