Tracking dependencies

[sluongng]

It would be nice if we can have a way to tracking inter-components dependencies relationship.

Would be useful in the following cases:

1. Ensure dependents SLO never exceed the dependencies
   I.e. if A depends on (B and C), and B and C each have the SLO of 99%, then A should not have the SLO exceeding 99^2 / 100 = 98.01

2. Circular dependency between services: Optional opt-out circular dependency detection. 2(or more) Services are circular dependents would have the same SLO which is the min SLO among them.

[gplasky]

In principle I think this is a great idea -- e.g. one could add commit/deploy time integration tests for things like #2.

That said, I think in practice for #1 this will be very situational and not sufficiently generalizable. At the very least, there's significant nuance to the interpretation of the nature of dependency and the relative impact of A on B. For example, if A depends on B, but A caches results from B for a specified freshness duration, what is the mathematical nature of the dependence of A on B?

[sluongng]

Hi @gplasky,

For example, if A depends on B, but A caches results from B for a specified freshness duration, what is the mathematical nature of the dependence of A on B?

-- begin-thinking-out-loud

I think this is the common circuit breaker pattern yeah? In this example, it would often be wiser to think of A as 2 services: A1 and A2. A1 responsibility is to update the cache(C) from B while A2 serves the data from (C). The dependencies chain will looks something like:

B -> A1 -> C -> A2

A1 and A2 can be classified with different SLA/SLO and different error budgets.

What happen when B fail?

A1, C, and A2 are all operating under a partial degradation manner: A1 cannot extract real data to populate cache, C holding stale cache which could potentially evict, A2 could be serving stale data. You would not have the ability to scale A up/down, not having the ability to failover A to a new AZ or deploy a new version of A with fresh cache for canary purposes.

I think in term of SLO, this would be impactful, but in term of SLA, end-user might not notice any differences.

-- end-thinking-out-loud

---

But you are right though. The OpenSLO spec should be flexible enough to enable folks to customize the different weights of the dependency links up to the actual business / organizational requirements. 🤔