ban_limit parameter // Re: BAN - Memory consumption exploding under load

[Nils Goroll]

(discussion started on -misc, added -dev)

Hi,

On 03/11/16 11:04, Dridi Boukelmoune wrote:
> I was thinking (against the too-many-knobs trend) that maybe we should
> have some sort of ban_queue_limit parameter to avoid piling up too
> many bans.

Dridi and myself have briefly discussed this and at this point I think, yes, we
should have such a parameter.

It seems that the best option when exceeding the ban list maximum would probably
be to have (by the ban lurker) all objects removed which are untested at the
tail of the ban list.

This way, we would loose the long-tail bit of the cache, but not the hot objects
(because they will already have been tested up the ban list) while still
retaining the correct ban behavior (because all objects which are supposed to be
banned will be banned - implicitly).

The alternative to refuse additional bans appears problematic because
applications will rely on the ability to invalidate the cache for correctness.

Nils