harpratap

Dental and vanity surgeries aren't happening in a vacuum. There are baseline costs eg. anesthesia, recovery medications, medical machinery etc which are all bloated due to the rest of industry not being under price pressure (rising tide lift all boats)

It's similar to how AI data center buildout race is raising the prices for consumer electronics in 2026 and beyond. The suppliers have no incentive to sell lower cost products to tiny niche

Because insurance companies incentivize upward price momentum. The ones who innovate and bring the prices down are not rewarded for their efforts. Health inflation is higher than headline inflation because of this absence of price pressure

This is very good use case of micro-transactions. If AWS makes $100 off Redis, they should be pay back X% to Redis project, from which the money is distributed to contributors based on how important their contributions were. Also Redis project is also supposed to pay back to the software components and 3rd party libraries it uses, so C project gets a fair share of the pie contributed back to them as well.

Wrote something on similar lines[1]. To understand a song completely I needed to dig deep in the artist's life & philosophy, the same with lyrics. I stopped trying to relate and match my existing life experience with the artist and explicitly tried understanding what the artist was trying to convey, dissolving my own understanding of what I thought it should sound like. Never went back to listening music the old way.

[1] - https://harpratap.com/2023/05/31/on-how-to-listen-Nusrat-fat...

Is that really a problem in Cloud environments where you would typically use a Cluster Autoscaler? GKE has "optimize-utilization" profile or you could use a descheduler to binpack your nodes better

Wouldn't the network cost be absurd in such case? Not only the pod-to-pod communication cost skyrocket, all the heartbeats, health checks, metrics, daemonsets pinging each other will probably end up costing more than the CPU and Memory

GKE Autopilot is pretty much useless, very few cases where it actually turns out cheaper than simply using Cluster Autoscaler + Node autoprovisioning. Not only is the pricing absolutely absurd, they don't even allow normal K8s bursting behavior (requests need to be equal to limits) which means you not only end up paying more than regular K8s cluster but now also need to highly overprovision your pods

Datadog allows you to export all of this but I don't see how that's any useful. You can't really port Datadog dashboard to let's say Grafana easily. The query languages they use do not have 1:1 mapping, the way dashboards are organized and the different visualization tools you get are not same either

> unless your a whale they give zero fucks about giving you any flexibility on price.

Even the flexible pricing they offer ends up being a sham. It just seems better on paper but you end up paying nearly the same because they have a really complicated billing model where they give you free stuff with Infra hosts, once you switch away from this model you stop getting those freebies, so your Infra hosts might cost less now but everything else is more expensive now. The house always wins!

> The company also noted that it'll offer 50% better multithreaded performance than an Apple M2, but didn't speak on single-core CPU performance.

Single threaded performance - https://twitter.com/IanCutress/status/1716904403259834529

VS Apple M2 Max: 14% faster at 30% less power

VS Intel i9-13980HK: 1% faster at 70% less power

How do you know the feeling(s) you are optimizing for are the right ones? A murderer is also optimizing for a feeling.

This is not a bare-metal vs Cloud comparison, the Cuber PaaS is supposed to run on anything including cloud based on their docs. So when they say "save 80% on cloud costs" I am assuming they mean for eg. save 80% on using Cuber on GCP compared to deploying same workload on GKE.

It would be very absurd to claim something like - I was using i9-13900KS but realized I could run the same workload on my raspberry pi, but hey I also used this packaging tool in the process therefore I saved 80% costs because of the packaging tool.

> We had 100% uptime and we saved 80% on cloud costs.

Is there a source on such bold claims?

GKE in general seems to be doing better than EKS though. They have a lot of things right and EKS seems to be playing catch-up. For example if you look at any Multi-cluster Kubernetes setup by AWS it's just a giant duct-tape rather than a ground-up solution. GKE worked on fundamentals first like multi-cluster endpoints, multi-cluster services, multi-cluster ingress, multi-cluster config-sync and now bringing it all together under GKE Enterprise.

It takes very long time for going from preview to actual production usage for anyone. We had T2D preview access more than a year ago, it took several months to get enough stock in Tokyo region (US always gets preference in such cases whenever a new machine type comes out). GCP already has Zen4 in preview for some US customers. Also, us being one of the largest GCP customers in Japan made things even slower

Yes, I point this out in the article too. Which CPU will perform better is heavily dependent on your workloads, so I refrain from relying 100% on synthetic benchmarks and directly ran canaries in production instead. It's definitely possible Ice Lake is superior for your workload than Milan

The dead thing is Intel (x86) and the successor is AMD (x86) as opposed to ARM in our case. Isn't it correct?

Cloud world is really slow. Imagine writing about Zen3 Milan in Aug 2023 when AMD has already announced Zen4c Bergamo. Actually we were "ahead of the curve" since we got access to T2D before it was made publicly available in Tokyo region, and even then it took several months to get enough capacity in Tokyo to fully migrate our production Kubernetes cluster.

I really wish we could test out RISC-V SoCs from the likes of tenstorrent, but it's a long journey

Thanks, that's a really insightful article!

> Since you are using Go and targeting a specific modern CPU, you may also get a measurable benefit from setting GOAMD64=v3

That's actually a long pending open issue in our backlog :)

That is true. But in cloud our hands our tied, we cannot really switch from one one generation to another so easily. GCP has so far never launched a new chip while keeping the price same or lower. We did same cost/perf analysis on newer generation chips like Ice Lake from Intel and even Milan from AMD in the form of N2D, but both are quiet expensive for the performance uplift they provide. The unique thing about T2D is that the price is competitive with 10 year old E2, which has been the case only for ARM based CPUs like Graviton and Ampere Altra (T2A from GCP)

We did a migration from GCP's Intel based E2 instances to AMD's T2D instances and saw huge 30% savings in overall compute! It is similar amount of savings folks got from switching to AWS Graviton instances, so looks like AMD might keep the x86 ISA still alive

Would be better to use OpenTelemetry vocabulary. This seems to be RUM product - https://github.com/open-telemetry/oteps/issues/169

Were people using it as an actual ES replacement for logs? fluentd -> GCP log sink -> BQ can get awful if you need to work with it on daily basis

If someone is making profit does not mean you are making a loss. Both consumer and producer could end up making profits depending on the situation

Wouldn't the cost of cooking be higher depending on who you are? If one could spend few hours doing something that has higher ROI than eating pre-cooked food, then you are actually losing money cooking your own food.

Tofu, Tempeh, beans, peas, Fu [1] (found in Japan but you might see local variants of this). If you are living a sedentary lifestyle you are bound to overeat if you want to reach anywhere near recommended protein levels. So it's just inevitable you will want to supplement your food with Protein powders. Soy Protein is a complete protein and much cheaper than Whey or Casein.

[1] - https://ja.wikipedia.org/wiki/%E9%BA%A9

There's nothing within the 5G spec itself that commands that every single cell site needs to be a massive MIMO supporting 1000s of users at once. Making a micro cell supporting just a family home is enough, then something a bit higher for small businesses supporting 10-100 users, then the likes of Starbucks can deploy bigger ones and finally the Sports stadiums and Corporate offices gets the really high end stuff.

Companies like Rakuten have already shown it's cheap enough to distribute 4G femtocells for free [1] to your users while increasing your coverage

[1] - https://network.mobile.rakuten.co.jp/guide/rakuten-casa/conn...

> If someone was willing to pay for the hundreds of tiny cell sites required.

We are already doing that, it's called WiFi. 5G SA is nothing but a souped up WiFi signal, and it should be treated as such and not how we have been treating 4G up until this point. Telcos need to change their business model otherwise the tech giants might end up eating their lunch

Because of cost?