On Mon, Apr 16, 2012 at 6:01 PM, Robert Raszuk <[email protected]> wrote:
> Nope.
>
> This is about recognizing duplicate update which good implementation should
> be able to do regardless.

I think 'good implementations' don't then exist, since ... as
shane/danny have presented a few times, a large percentage of update
traffic seen is repeats from the same speaker.

I'm not sure that being able to track repeats is really necessary
though, if one speaker is +10% bad (pick a percent) over the last 5
mins (or 500 update messages, perhaps this is a chance for a knob for
either/both/mix) shutdown the session and keep it down (or bounce and
re-establish and after X bounces in Y time, shutdown)

If the concern is, this peer keeps sending me an update that is
unparsable, ignoring it (and logging a note/alert/etc) seems ok. If
the problem is that the neighbor did something quite wrong on the
session as a whole more immediate action should be taken.

There's some sympathy to be had for situations like ATTR-128 though,
why bounce 50% of the speakers on the internet repeatedly if some
downstream of a downstream screws up?

-chris

> Nothing to do with error itself.
>
> Regards,
> R.
>
>
>> So you're going to _remember what error occurred_????
>>
>> Complexity++
>>
>> No thanks,
>> Tony
>>
>>
>> On Apr 16, 2012, at 12:43 PM, Robert Raszuk wrote:
>>
>>> Ilya,
>>>
>>> I meant to say that you would increment the counter only for unique
>>> updates. So duplicate updates with the same NLRIs would not be counted N
>>> times.
>>>
>>>> On the other hand, if a session is reset (due to errors per other
>>>> rules) there's chance that problems will occur again. Rather than
>>>> bouncing session for extended period, I'd prefer such troublesome
>>>> peers to stay down. We spoke about this behaviour few times, but I'm
>>>> not sure if we have reached consensus.
>>>
>>>
>>> I agree. We could have max-prefix like configurable restart as an option.
>>>
>>> Rgs,
>>> R.
>>>
>>>> Robert,
>>>>
>>>>> -----Original Message-----
>>>>> Algorithm:
>>>>> ----------
>>>>>
>>>>> If such counter get's exceeded session get's reset.
>>>>>
>>>>> There is tracking (counter++ operation) and one additional conditional
>>>>> if
>>>>> statement. There is no analysis what so ever.
>>>>>
>>>>
>>>> Let's consider following scenario: a router in no-longer-kingdom
>>>> far-away
>>>> sends malformed update that somehow manages to slip through multiple
>>>> intermediate routers and comes to me via a peer that sends me few dozen
>>>> thousands of other prefixes (everyone has seen this in the wild not too
>>>> long
>>>> ago). I increase "error tracking counter", the far-away router for some
>>>> reason sends update again, I increase counter again, at some point I
>>>> reach
>>>> threshold. Clearly, resetting my session with a peer that sends me large
>>>> number of "good" prefixes isn't to my advantage, especially that it
>>>> won't
>>>> repair distant router. My personal preference would be to ignore such
>>>> errors
>>>> until kingdom come.
>>>>
>>>> On the other hand, if a session is reset (due to errors per other rules)
>>>> there's chance that problems will occur again. Rather than bouncing
>>>> session
>>>> for extended period, I'd prefer such troublesome peers to stay down. We
>>>> spoke about this behaviour few times, but I'm not sure if we have
>>>> reached
>>>> consensus.
>>>>
>>>> /iLya
>>>>
>>>>
>>>>
>>>
>>> _______________________________________________
>>> GROW mailing list
>>> [email protected]
>>> https://www.ietf.org/mailman/listinfo/grow
>>
>>
>>
>>
>
> _______________________________________________
> GROW mailing list
> [email protected]
> https://www.ietf.org/mailman/listinfo/grow
_______________________________________________
GROW mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/grow

Reply via email to