[dns-operations] prefetching and thundering herds

[Tony Finch]

Mark Delany <b9w at charlie.emu.st> wrote:
> On 15Jul20, Philip Homburg allegedly wrote:
> > From the point of view of latency hiding, it may make sense to always
> > reduce the ttl so that a recursive resolver can reply to the stub
> > resolver immediately and simultaniously trigger a refresh

Yes, I was wondering if "always" will work well enough, or (if there is in
fact enough of a herd to be worth trying to spread it out) whether
something jittery might be a bit better.

Bert's post suggested that the load is spread across enough domain names
that TTL expiry herds are more like a whisper than a thunder, so as he
said, if there is any prefetch logic it should err on the side of
simplicity...

> That's a clever idea, tho it does presuppose that most of the stubs
> are caching. In the worst case where no stubs are caching, all that's
> been achieved is to increase the query rate to the recursive resolver.

If they aren't cacheing the stubs will be ignoring the TTL from the
recursive server so it doesn't matter how the TTL is adjusted :-)

Tony.
-- 
f.anthony.n.finch  <dot at dotat.at>  http://dotat.at/
Rockall, Malin, Hebrides: West 4 or 5, backing southwest 5 to 7. Moderate
becoming rough. Drizzle, fog patches. Moderate or good, occasionally very
poor.