SRE here, Strong disagree with do not fail readiness and liveness checks on upstream dependencies failing. There are several reason to do so and unless you have extreme start up time, what's the problem with restarting?
Maybe DNS has changed on you but you are stuck with bad local cache because you poorly respect TTLs (Looking at you Java), reseting the process will clear that cache away.
Maybe TCP connections are in stuck weird state, resetting the process generally helps with that.
Maybe someone gave you bad ENV VARs and you cannot connect to database, by refusing to progress the rollout, no outage generated.
So yea, if you are not ready to do work including critical upstream dependencies, don't lie to system and say you are.
1. was already mentioned in sibling - cascade failures
2. you'll have massive number of restarts for various flake reasons and missing things that got papered over with restarts until you hit 1 and everything is broken. another popular version of this is "just restart when memory leaks too much"
Thundering herd / cascading outages. You take out a large enough portion of your fleet, and the remaining load overloads your remaining nodes one by one as they restart, so you can never have enough healthy nodes.
My personal take-away is this: whatever you choose, make sure it's consistent across services (not serviceA behaves like X and serviceB like Y) and make sure eng teams know _how_ these are configured and what can go wrong. They'll figure out the rest.
That's a problem for circuitbreakers on these kinds of actions, not lying on health checks.
Something like healthcheck fails -> restart -> healthcheck fails -> restart -> healthcheck fails -> circuit breaker trip, alarm raised, give up until manual intervention or X minutes have passed
backoff is only applied to individual pods/containers not across pods. the point is at scale it's easy to get into a situation where it's not possible to recover without (usually manual) full service drain
SRE here, Strong disagree with do not fail readiness and liveness checks on upstream dependencies failing. There are several reason to do so and unless you have extreme start up time, what's the problem with restarting?
Maybe DNS has changed on you but you are stuck with bad local cache because you poorly respect TTLs (Looking at you Java), reseting the process will clear that cache away.
Maybe TCP connections are in stuck weird state, resetting the process generally helps with that.
Maybe someone gave you bad ENV VARs and you cannot connect to database, by refusing to progress the rollout, no outage generated.
So yea, if you are not ready to do work including critical upstream dependencies, don't lie to system and say you are.
1. was already mentioned in sibling - cascade failures
2. you'll have massive number of restarts for various flake reasons and missing things that got papered over with restarts until you hit 1 and everything is broken. another popular version of this is "just restart when memory leaks too much"
Thundering herd / cascading outages. You take out a large enough portion of your fleet, and the remaining load overloads your remaining nodes one by one as they restart, so you can never have enough healthy nodes.
SRE team debates correctness versus availability for the 540th time this year
You're both correct, of course!
What this guy said :point_up:
My personal take-away is this: whatever you choose, make sure it's consistent across services (not serviceA behaves like X and serviceB like Y) and make sure eng teams know _how_ these are configured and what can go wrong. They'll figure out the rest.
That's a problem for circuitbreakers on these kinds of actions, not lying on health checks.
Something like healthcheck fails -> restart -> healthcheck fails -> restart -> healthcheck fails -> circuit breaker trip, alarm raised, give up until manual intervention or X minutes have passed
That circuitbreaker exists, by default. It is "CrashloopBackoff", here, and TFA covers it. (& it's an "until X minutes have passed" kind, by default.)
backoff is only applied to individual pods/containers not across pods. the point is at scale it's easy to get into a situation where it's not possible to recover without (usually manual) full service drain
My favourite: misconfigured Linkerd setup that causes CA certs to rotate every month :) Definitely worth restarting on that
The better solution is to not have too many critical upstream services :)
What are your feelings about using initContainers and wait-for-it to skirt the thundering-herd problem?
> what's the problem with restarting?
Exponential backoff can delay recovery up to kubelet’s maxContainerRestartPeriod (default 5m).
cascading failures on upstream services. then you get 20 different services failing instead of the single one.
This does not state anything new, but explains it so much well than the kubernetes documentation.
Sam has a real way with words when it comes to educational content.
Though he did find a legit Kubernetes bug while writing the post, so technically there was at least one new thing :)
Sometimes the k8s docs remind me of google's documentation
many times i have said a company could be built to just make better google documentation.
I need to know how to animate things like this for internal documentation.