Re: [Teas-ns-dt] Availability

[Eric Gray <[e-mail]>]

Greg,

              As mentioned in a separate thread, it is possible to look at availability as a parameter to be associated with each individual SLO.  I would need to see evidence that it is worth the extra complexity, however, before I would agree that we should do it that way.

              For example, it is conceivable that a service might be defined where the availability is different for packet loss and delay variation.

              There are simpler ways to express this sort of unequal weighting of SLO parameters, however, and my impression (from talking to provider representatives in relevant standards work over multiple decades) is that defining a single availability is the usual case.

              For example, packet loss and availability are both expressed as a percentage, then you can choose a value for packet loss percentage that – in combination with a given availability percentage – results in any desirable SLO loss percentage.

              In other words – if you’re looking to define an SLA that might allow for out-of-bounds delay variation 5 minutes a year, and want to define a packet loss SLO with a consistent availability, you have only to choose an appropriate packet loss percentage (where losing less than the acceptable percentage 99.999 % of a year and then 100% of all packets for 5 minutes is about equal to that percentage).

              As you should see from this example, if the availability is a very high percentage, this sort of math produces a difference that is way down in the statistical “noise” – hence very few people bother dealing with this level of complexity and simply set a single availability for the service.

              Note that another key reason for having a single availability number is that – in the event of a severe service outage (loss, impairment or degradation) – many (if not all) SLO parameters are likely to be out of bounds for approximately the same amount of time.

              Since the case Jeff refers to (where all SLO values are essentially ANDed) is the common case, it is simplest to treat failing to meet any SLO as service unavailability.

              Lastly, in any case where the provider detects the service failure, the way that they will likely have implemented violating their SLA with a subscriber will typically involve switching the service to an alternative – which will likely impact the service’s ability to meet other SLOs.

--
Eric

From: Teas-ns-dt <[e-mail]> On Behalf Of Jeff Tantsura
Sent: Monday, June 1, 2020 6:07 PM
To: Greg Mirsky <[e-mail]>
Cc: Kiran Makhijani <[e-mail]>om>; [e-mail]; Eric Gray <eric.gray=[e-mail]>
Subject: Re: [Teas-ns-dt] Availability

Hi Greg,

A service (SLA) could have 1 or more SLO’s associated with it.
A SLO is met (TRUE), when its objective is met (within boundaries specified).
Usually a SLA is composed of a set of SLO’s with logical AND, e.g if any of SLO’s is FALSE -> SLA (or else)

Example:
If  SLO (availability) is met but SLO (packet_loss) isn’t, availability becomes an irrelevant objective.

Cheers,
Jeff
On Jun 1, 2020, 2:55 PM -0700, Greg Mirsky <[e-mail]<mailto:[e-mail]>>, wrote:

Hi Jeff,
thank you for the clarification. Does measuring uptime considers whether all metrics included in SLO are within their respective acceptable limits? In other words, if the quality of the TS degraded, due to, for example, excessive packet loss, below the requested threshold, would that time period be attributed to the Service uptime period? In my experience, uptime of a node (router, server) is easy to express. Uptime of a service? Much appreciate it if you help with an example or a reference to the definition.

Regards,
Greg

On Mon, Jun 1, 2020 at 2:46 PM Jeff Tantsura <[e-mail]<mailto:[e-mail]>> wrote:
Hi Greg,

SLO - is an objective (as the name suggests), not a metric. A metric without a context is meaningless.
SLO makes use of the metrics gathered to derive whether the objective has been met.

Example:
SLO (availability) = uptime 90% over 10 hours
total_time=10h
uptime=8h

using the metrics above we can conclude that the total_availability = 80%, which is less than the service objective set (90%) ->  SLA(or else)

Hope this clarifies

Cheers,
Jeff
On Jun 1, 2020, 2:32 PM -0700, Greg Mirsky <[e-mail]<mailto:[e-mail]>>, wrote:

Hi Jeff,
in my reading of the definition, it is the intersection of metrics already listed in the SLO. If that is the case, how useful is another metric that is only a reflection of other metrics?

Regards,
Greg

On Mon, Jun 1, 2020 at 2:24 PM Jeff Tantsura <[e-mail]<mailto:[e-mail]>> wrote:
Greg,

I thought the definition provided was pretty clear and comprehendible, why do we need to rephrase it?

Cheers,
Jeff
On Jun 1, 2020, 2:00 PM -0700, Greg Mirsky <[e-mail]<mailto:[e-mail]>>, wrote:

Hi Jeff,
if we define availability as the ratio of the period all requested in the SLO metrics are within an acceptable range to the time since the service was handed to the customer (I propose to refer to this metric as "availability ratio"), then I think it can be expressed as
[\bigcap _{i=1}^{n}A_{i}], where Ai is the time period the particular metric remained within its acceptable boundary.

Regards,
Greg

On Mon, Jun 1, 2020 at 1:35 PM Jeff Tantsura <[e-mail]<mailto:[e-mail]>> wrote:
Mostly agree with Eric/Kiran

It should not be removed, but further clarified.
Network/service availability is a measurable metric, availability = uptime/total_time(uptime+downtime)
Rule of thumb - a service is deemed available when all the SLO’s associated with it are met(TRUE).
In a complex/multidimensional service, different objects might have different availability metrics .
For simplicity sake - total_availability(normalized metric) = Σ(subservice-1..subservice-n), so both, per SLO as well as composite metrics can be used.

Cheers,
Jeff
On Jun 1, 2020, 10:08 AM -0700, Kiran Makhijani <[e-mail]<mailto:[e-mail]>>, wrote:

Thanks! I support not removing it.
Sticking with individual SLO seems to be a right decision but can be deferred to NBI document. we need not state that here.
-Kiran

From: Teas-ns-dt <[e-mail]<mailto:[e-mail]>> On Behalf Of Eric Gray
Sent: Monday, June 1, 2020 7:35 AM
To: [e-mail]<mailto:[e-mail]>
Subject: [Teas-ns-dt] Availability

I agree that the definition needs to be cleaned up, but I disagree that it should be omitted.

A part of what probably should be cleaned up is the part that talks about service degradation.  In general, this is an important factor in determining availability, but it is a bit vague for the purpose of definition.

I also disagree that availability is not measurable.

As a proof of concept for measuring , if there are any mandatory measurable objectives, then failing to meet any of those objectives makes the service measurably unavailable.  That is, if you can determine if specific mandatory objectives are being met, then you can determine if they are not being met and therefore determine if the service is unavailable.

Availability is an important aspect of any service, because it is understood that the higher the required availability, the more difficult (and thus expensive) it is to provide that service.

Defining availability as a fraction as we have done in the draft, allows for services that may experience a certain amount of outages over a service period.  A service request may ask for as high an availability as the provider and requester have agreed to (under the terms they agreed to) in advance.

Note that this elevates the importance of having (at least mostly) measurable objectives, simply because you cannot determine if a non-measurable objective is being met – hence you cannot (necessarily) determine the availability of any service that depends on that objective.

It is further interesting to note that the notion of a service depending on objectives that it cannot determine are not being met is a non-sequitur.

Measuring availability in terms of mandatory objectives – as a whole – is the simplest approach; one could group one or more mandatory objectives and define an availability separately for the group – thus allowing for a higher degree of acceptance for failing to meet one set of service objectives compared to others.

If we were going to do that, it would probably be better to define availability as a parameter that applies individually to service objectives.

In my opinion we should at least initially stick to the simple case, where availability is defined as a service objective, rather than as a parameter of every service objective – but I am willing to go either way.

--
Eric
--
Teas-ns-dt mailing list
[e-mail]<mailto:[e-mail]>
https://www.ietf.org/mailman/listinfo/teas-ns-dt
--
Teas-ns-dt mailing list
[e-mail]<mailto:[e-mail]>
https://www.ietf.org/mailman/listinfo/teas-ns-dt