You have my empathy for this kind of sentiment. Personally this seems somewhat rare in practice. That being said I'm curious if anyone has anecdotes they can share about these kinds of things?
1. People add retries w/o backoff. Now we have a retry storm.
2. People don't add jitter so we get huge waves of highly correlated retries that cause self-perpetuating overload cycles and failure.
3. People add retries (even w/ backoff) at more than one layer (e.g. one in process, one in envoy), so now we have a deeply confusing multi-level of retries with super weird n^2 patterns.
4. People find ways to fetch from multiple data sources to make a composite object but don't cache/reuse data they fetched, so one data source being down causes DOS on _other_ data sources because of retries.
5. People add failover mechanisms where all failures happily pick the same failover target because, ya know, it has the most free capacity.
6. People underestimate connection setup cost so "failover" causes huge increase in overall load due to connection setup (often tls setup) causing huge influx of "new" cpu work in a loaded system. CPU spike causes unhealthy destinations, causing more failovers, causing metastable failure.
7. People realize purely local decisions aren't optimal so add a layer of global health metrics (e.g. global retry budgets) but these systems add a time delay to the metric (ie its the view of the world 5ms ago). This delay makes their naive control algorithm go into oscillations or divergence spraying stuff everywhere.
All of these things have a similar pattern that when things are almost all good, they will make the system more robust. You'll get an extra 9 of reliability on good days. But they make everything far worse when things are bad.
Configuring postgres to automatically failover instead of doing it manually. The automated system caused more downtime in a few months than manually doing it did for years before. All in the name of more automations and less downtime
For resolution of a problem you can’t prevent categorically, you have to spend a good bit of time pretending to be the computer before you can trust the computer to do it. A run book is a draft of a requirements doc for a program to do the same thing. And you can’t afford to test that program in production without loads of simulations first. And that’s running the process manually and checking aggressively for any signs of problems.
Some coworkers snaked the autoscaling work out from under me, and then ignored the advice I offered on low water marks and weeks of testing and rehearsals. All I can figure about their irrational exhuberence is something to do with claiming something for annual reviews? All I do know is we didn’t make it 40 hours from when they flipped it on until we had our worst outage in two years. Classic FAFO.
They were hoping to eventually get to a low water mark of just over 30% of our static cluster size, and they decided to use that for their initial go instead of my estimate of 40% being the low, and a recommendation that they start with 50-60% for the initial weeks and then ramp it down. When I refused to let it go in the status meeting, the team decided we should vote on it. Two guesses how that went.
A bug in the deployment logic the Ops team had for doing things like upgrading VMs caused the next update during daylight hours to spin up the new instances at the minimum cluster size, instead of the current cluster size. It was done outside peek hours but that still had us cut over with just under half of the hardware we needed at that moment. Because we were in such a fucking hurry to be something like third in line to use the new autoscaling support, nobody else had run into this problem yet (or system did a pretty good job of buffering everyone else).
A couple weeks later they’d looked at the stats and decided that we were spending less than a half hour a day running at the low water mark, and the additional shutdowns were causing churn that made it harder for us to detect problems like memory leaks. Surprise, surprise, they increased the min cluster size to exactly what I’d told them two weeks before.
Since “listen to me” isn’t a lesson that transfers to other teams, I will instead say, don’t transfer initiatives to a new team when there is less than 10% of the project left. The lack of friction you encounter may result in a lack of respect for the danger, and attention to potential problems. And if you must transfer, if your ideas about rollout timeline end up being less conservative than theirs, be patient and do it the slow way. They probably have encountered problems you haven’t seen yet.
I think the "happy path" might be a slightly wrong classification in GP, since the post is in reply to a retry-storm issue and explicitly talks about retry storms and thundering herds.
I've seen many cases where engineers optimize the sad path, but pessimize the wretched path. Or in less flowery language, they cut the occurrence rate of common non-critical failures, but by doing that they introduce code that can make rare failures much worse.
The cases I've seen generally boil down to naive retry logic or poorly tested and poorly maintained fallback paths (such as killswitches that break their environment[1], graceful degradation turned graceless, dormant feature flags that get reactivated).
The case you see with a retry storm here is the most classic one and the one that annoys me the most. I've seen engineers adding aggressive retries even into places where the impact is minor (you could show an error and let the user manually retry instead). Retries that improve user experience can be great if done correctly, but I've never seen the authors of such pull request addressing the risk and mitigation techniques for retry storm or retry amplification.
I've seen cases which had:
1. Retries on the client side (browser or mobile app).
2. Retries on the BFF.
3. Retries on Microservice A used by the BFF.
4. Retries on Microservice B used by Microservice A.
5. Retries on Critical Service C used by Microservice B.
Most of these retries had very short timeouts (e.g. 100ms), in order to keep latency SLOs during normal operations (not a good idea on retries). Every time QA saw a layer without retries, that would be a bug, and adding retries is easy, so we'd get a new retry without much thought. But the first time Critical Service C became overloaded, Microservice B started timing out a couple of times and retrying. This was too much too much for Microservice A that had a short timeout that couldn't hold the 3 retries done by Microservice B, so it making doing its own retries, all of them dropped in the middle of the way. Eventually you'll get a full-blown retry storm where every request from the client side got amplified with 3^5 retries, easily bringing down Critical Service C.
We'd usually introduce a circuit breaker for the particular path that caused the issue, but a variation of this kept happening several times because designing safe retries across a vast collection of microservices takes a lot of effort, and it's always easier to just add a quick-and-dirty retry at any point where you think you might need one and call it a day.
A proper solution (which I've never seen implemented) would be an mandating a corporate-wide inventory of retry-paths, and monitoring it for any path that is at risk of triggering a retry storm, or adding mandatory headers that cross microservices and track the amount of retries done up the chain and the time spent in total waiting for previous retries. You could have a budget for both and automatically stop performing more retries. Both solution require extra effort and a large degree of coordination.
> I've seen many cases where engineers optimize the sad path, but pessimize the wretched path.
As said GP, yes this is exactly what I meant and "happy-path" was an unclear choice of words. They optimize for the unhappy path on the good days, and make the bad days much worse.
