How do you calculate the compound Service Level Agreement (SLA) for cloud services?

Question

Cloud services hosted by Amazon Web Services, Azure, Google and most others publish the Service Level Agreement, or SLA, for the individual services they provide. Architects, Platform Engineers and Developers are then responsible for putting these together to create an architecture that provides the hosting for an application.

Taken in isolation, these services usually provide something in the range of three to four nine's of availability:

• Azure Traffic Manager: 99.99% or 'four nines'.
• SQL Azure: 99.99% or 'four nines'.
• Azure App Service: 99.95% or 'three nine five'.

However when combined together in architectures there is the possibility that any one component could suffer an outage resulting in an overall availability that is not equal to the the component services.

Serial Compound Availability

In this example there are three possible failure modes:

• SQL Azure is down
• App Service is down
• Both are down

Therefore the overall availability of this "system" must lower than 99.95%. My rationale for thinking this is if the SLA for both services was:

The service will be available 23 hours out of 24

Then:

• The App Service could be out between 0100 and 0200
• The Database out between 0500 and 0600

Both component parts are within their SLA but the total system was unavailable for 2 hours out of 24.

Serial and Parallel Availability

In this architecture there are a large number of failure modes however principally:

• SQL Server in RegionA is down
• SQL Server in RegionB is down
• App Service in RegionA is down
• App Service in RegionB is down
• Traffic Manager is down
• Combinations of Above

Because Traffic Manager is a circuit breaker it is capable of detecting an outage in either region and routing traffic to the working region, however there is still a single point of failure in the form of Traffic Manager so the total availability of the "system" cannot be higher than 99.99%.

How can the compound availability of the two systems above be calculated and documented for the business, potentially requiring rearchitecting if the business desires a higher service level than the architecture is capable of providing?

If you want to annotate the diagrams, I have built them in Lucid Chart and created a multi-use link, bear in mind that anyone can edit this so you might want to create a copy of the pages to annotate.

Answer 1

After reading Tensibai's excellent answer, I realised I used to be able to calculate this for network analysis purposes. I dug out my copy of High Availability Network Fundamentals by Chris Oggerino and had a crack at working this out from, not quite first principals.

Taking my serial example directly out of Tensibai's answer is simply a case of multiplying the probability of each component being available by the other:

So

99.95% * 99.95% = 99.9%

Calculating it in parallel is a little more complicated as we do need to consider what the percentage unavailability will be:

The calculation is done as follows:

1. Multiply the unavailability of the two regions together.

   0.1% * 0.1% = 0.0001%

2. Convert that back to availability

   100% - 0.0001% = 99.9999%

3. Multiply the Traffic Manager availability by the availability of the two regions.

   99.99% * 99.9999% = 99.9899%

4. The result is the whole system availability.

   99.9899% is close to 99.99%

I ended up using Excel to perform the calculations, here is the values:

... and the formulas ...

Answer 2

I'd take that as a math problem with the SLA being the probability of being OK.

In this case we can rely on probability rules to get an overall.

For your first case the probability that App Service (A) and Sql Service (B) are down at the same time is the product of their probability:

P(A)*P(B) = 0.0005 * 0.0005 = 0,00000025

The probability that one of them is down is the sum of their probability:

P(A)+P(B) = 0.001

When two events are independents the resulting formula to take in account the probability of both being down is:

P(A,B) = P(A) + P(B) - P(A)*P(B) = 0.001 - 0,00000025 = 0,00099975

So the overall SLA would be 1 - 0,00099975 = 0,99900025 wich in percent is 99.900025 %

A simplification is the product of the first probability: 0.9995 * 0.9995 = 0,99900025.

Applied to your 1h/24h outage (4,166666% of a day) this gives (decimals are abbreviated):

0.0416 + 0.0416 - (0.0416 * 0.0416) = 0,081597222

So the probability of being OK is 1 - 0.0816 = 0.9184 in percent: 91,84%

24 * 0.0816 = 1.95 h

This is less than the worst case of 2 hours because there's a chance both are down at the same time.

Keeping that in mind, you may notice the availability for each is 95,84% and 0,958333333 * 0,958333333 = 0,918402778 which is our 91.84% from above (sorry for the full decimals here, but they are needed for the demonstration)

Now for your second case, we'll start gain from our compound probability for each region (Sorry I dismissed the change for SQL to keep it reasonable), assuming there's no independent probability for the region itself and that each region is isolated and as such a DB failure take only its region down.

We have the traffic manager OK probability P(T) = 0.9999 and each app+DB couple with a OK probability P(G) = 0,99900025 from

How much region we have play a role as we have to apply the product of failure probability only to get the probability both region are down as the same time:
0,00099975 * 0,00099975 = 0,0000009995000625 which means an overall availability of at least one region of 99,049375 %

Now we have the overall regions availability, the product with the traffic manager one give us the overall availability of the system:

0.9999 * 0,9999990004999375 = 0,99989900059988750625

The overall availability is 99.989900 %

Another source as explanation is available on Azure's docs (link courtesy of Raj Rao)

Answer 3

In the second diagram you multiply the two regions to calculate a full system availability. In all cases you are calculating a probability that the system shall be down (or up).

There is however, no reference to the worst case scenario where each component fails in a serial manner, and the duplicated components fail together. (i.e. traffic manager fails, comes back as both app services fail (or app-A and SQL-B), they are in turn restored as both SQL database fail)

So a dual site solution can't provide any more guarantee than a single site solution in the worst case. But can only reduce the probability of an issue occurring.