HN user

throwaway427

288 karma
Posts0
Comments110
View on HN
No posts found.

This is a great presentation.

Where I work we don't use any process level parallelism at the ruby level, we hoist that up to the kubernetes level and use signals (CPU load, job queue sizes, etc) to increase/decrease capacity. Workloads (replica sets) are segmented across multiple dimensions (different types of API traffic, worker queues) and are tuned for memory, cpu and thread counts according to their needs. Some heavy IO workloads can exceed a single cpu ever so slightly because db adapter isn't bound by the GVL, but practically speaking a pod/ruby process can only utilize 1 CPU, regardless of thread count.

One downside of this approach though is it takes a long time for our app to boot and this along with time to provision new nodes can cause pod autoscalers to flap/overprovision if we don't periodically tune our workloads.

In a perfect world we would be able to spawn processes/pods that are already warmed up/preloaded (similar to forking, but at the k8s level and the processes are detached from the root process) in a way that's not constrained by the CPU capacity of some underlying k8s node it is running on and instead is basically an infinite pool of CPUs that we only pay for what we use. Obviously serverless sort of offers this kind of solution if you squint but it is not a good fit for our architecture.

Since you seem to have some insight into this, I interviewed a former Heroku eng a while back and he said that they deploy Heroku infra on Heroku, essentially dogfooding their own product. I also see a lot of claims that Heroku has essentially stagnated on features. I'm curious how do internal teams reconcile this (do they just work around the stagnation?) and is there a layer that is a choke point for product improvement that also affects to internal teams (if so that would seem demoralizing)?

I listened to this interview and it's interesting the way you put it:

Steve Inskeep rephrases a statement that Goelz makes to make it seem like the FAA is short staffed and that airline compliance is on the honor system.

Goelz then says what he means is that there cannot be an independent FAA inspector in every single corner of a manufacturer or maintenance shop, but by then it's too late because the seed has been planted listeners mind.

So here you are saying the "FAA and NTSB just doesn't have the resources to provide adequate inspections" and that the overall compliance model here is "cooperative self regulation" which... I guess barring the FAA having a 1:1 staff for every employee in the airline industry we are indeed going to have "cooperative self regulation".

I agree and I don't think you deserve the downvotes. The FAA is an institution with a deep sense of safety first and independence from (and even dominion over) private industry.

They are also a body of experts, with more collective knowledge than any other similar body of domain experts in world.

If we believe in institutions of experts who study and make recommendations regarding complex systems, we owe it to the FAA to trust they are charting a prudent course based on the evidence they are gathering and analysis they are doing.

Concur, lived in Texas and NYC both, Austin is hotter for sure, but air conditioning is on blast everywhere and the driving helps. Just walking a couple of blocks on a 90 degree day in NYC is enough to get you soaked.

On top of that you can wear shorts most workplaces in Austin and not have a problem. Shorts in NYC is frowned upon in the cosmopolitan set and professional industries.

Yeah it's spacemacs. Any soup to nuts emacs distribution like this is going to be slow. I experience it all the time in doom. You don't have anything misconfigured.

Features like helm, ivy, projectile, etc cost. Linters and autocomplete can really make emacs jank. All this stuff is configured out the gate in these distributions.

I will often scope google searches for product reviews or food/recipe/diet/fitness knowledge to reddit. This is not the sole research I do but it does tend to be pretty high quality data.

I would not be surprised if Reddit is an ML/AI dataset gold mine.

Suicide isn't new or specific to humans

Self-destruction is found in non-humans, but mostly as a reaction to acute circumstances. I'm assume you are widening the scope of what is suicide to include these examples, but they seem different than relieve of existential suffering that is human suicide.

Theoretically I am sympathetic, but in reality I have 3 different podcast apps:

1. Overcast, great usability, terrible discoverability

2. Apple Podcast, bad usability, excellent discoverability

3. Spotify, ok usability, ok discoverability

The reason I have not unified on any one is these different trade offs and the low burden to having 3 different podcast apps. I only use Spotify for the Fantasy Footballers podcast and I can see myself unifying on Spotify if they became the sole distributor of a couple of casts I follow.

Which is to say, in practice, I do not share your conviction.

Current Tundras, though being very different than your gen, are still seen as maybe the best forever truck. Those engines are pretty much bug free but also they get terrible gas mileage.

I'm hoping my V8 F150 goes for 20 years, though I doubt I'll keep it that long. I like the safety bells and whistles in modern cars and my 2016 is just old enough not to include adaptive cruise or a 360 camera or a blind spot camera.