[bind10-dev] minimal-responses, default behavior, & consistency (was some ideas to improve query performance of DB-based datasrc)

[JINMEI Tatuya / 神明達哉]

At Mon, 8 Oct 2012 09:40:13 +0200,
Peter Koch <pk at DENIC.DE> wrote:

> > It also affects caching server behaviors.  For example, BIND 9 caches
> > NS's AAAA and A records in the additional section in some specific
> > way.  If the authoritative server omits this, this might result in,
> > e.g,, more queries from caching servers.
> 
> if I understand the option correctly, it would also change - on a large
> scale - the relation between the NS RRSets in the delegating and the
> delegated zone as well as the relation between additional (address)
> data originating from glue vs authoritative sources.  This is
> important for RFC 2181 credibility levels.  I'd rather not see a change
> of this impact applied lightly.

I'm not sure about specifically what kind of change you
mean...consider the following two cases:

The "example" zone that delegates the authority of "child.example" as
follows:

child.example. NS ns.child.example.
ns.child.example. A 192.0.2.1
ns.child.example. AAAA 2001:db8::1

This server returns all of these (NS in the authority section and A
and AAAA in the additional section) for queries on or under
"child.example", regardless of the minimal-responses option.

The child.example zone also has these records.  When this server
responds to a query for that zone (with an answer), it includes all of
NS, A, AAAA if minimal-responses is "no" (default for BIND 9).  If
it's set to yes, it only fills in the answer section, and omits these
three records in the response.

What kind of relation change do you specifically see in the
child.example case if minimal-responses is "yes"?

BTW, in my understanding powerdns behaves like minimal-responses being
set to yes by default (I'm not even sure if it's configurable).  I'm
not saying this to defend the idea of setting it to yes by default for
BIND 10, but unless I misunderstand the powerdns behavior, we should
have already seen the "impact" at some scale.

In any case, I see arguments for keeping the BIND 9's behavior by
default in general.  If that's our conclusion, my revised suggestion
would be to introduce caching these records and save DB lookup
overhead for every query.

---
JINMEI, Tatuya
Internet Systems Consortium, Inc.