Uhh, a statement like this somewhat hurts the ability to trust the other claims of the benchmark. Sure, Linode has the lowest clock speed, but you're comparing Epyc Milan which is almost a decade newer than the Sandy-Bridge EP of Cloudfanatic for example. Unsurprisingly, this is even reflected in the one cpu benchmark run, geekbench, which Linode comes first in.
> Uhh, a statement like this somewhat hurts the ability to trust the other claims of the benchmark.
Why?
> Sure, Linode has the lowest clock speed, but you're comparing Epyc Milan which is almost a decade newer than the Sandy-Bridge EP of Cloudfanatic for example.
Sure, but also take into consideration that the Linode instance costs $48/month while the Cloudfanatic one costs $18/month. Both of them have the same amount of CPU cores too.
Seems like this is a Linode problem of offering old hardware, instead of a problem with the benchmark itself.
Clock speed alone is a spectacularly bad metric to use for a summary result regarding relative performance of different CPUs.
If one wants to assess performance on a workload of integer addition/subtraction with a linear chain of data dependencies on a working set that fits in the L1, then sure, clock speed will correlate strongly with performance (in practically every metric other than maybe power).
On the other hand the performance of a lot of real-world workloads will be influenced by things like the number of execution units per core, the types of operations they perform, the throughput of various instructions, the specifics of the front end (e.g. assuming x86/CISC, the types of uops; the size of the BTB; the branch-prediction algorithm), the size of the various caches (TLBs in addition to I$ and D$), the specifics of the coherency protocol, the memory bandwidth, etc.
All that to say that CPUs should be compared by benchmark results, not blind comparisons of clock speed, and so to quote clock speed in a summary might lead one to question the rest of the results.
> Seems like this is a Linode problem of offering old hardware, instead of a problem with the benchmark itself.
Linode has the newest hardware, and the highest single- and multi-core speeds despite the lowest clocks. Mentioning the higher clock speeds instead of the benchmarks is pointless and very misleading.
To add on to what denotational wrote, an assertive and wrong statement on the cpu performance also implies they didn't research too in depth before running the benchmark, so it casts doubt on if they controlled for other factors in their test too. Like for the disk test, did they take care to ensure all the VPSes are running the same filesystem? Do any of the providers make you choose between directly attached storage and network block storage? etc.
Hetzner was likely excluded because they have a habit of preventing signups, randomly banning accounts, or blackholing traffic exactly when your app is taking off because their shitty tech thinks your system was compromised or under a DDOS attack.
Is this anecdotal or based on widespread publicly available evidence? Not meant as an accusation, but it’s the first time I’m hearing this and it’s slightly alarming
Tbh, many comments here are FUD and downvotes by big tech employees. And somehow DO is HN’s little baby or something. They did have great ads during their launch period.
Why in the world do you think “big tech” would single out one shitty German hosting company not even covered by the post in question, and I can't just be one pissed off ex-customer warning others to stay far away?
I should do a write-up but basically it was a combination of running tailscale (which sends a packet to an unroutable UDP destination nonstop [1]) and my game taking off at the same time. It triggered their false-positive-laden DDOS protection. I responded to the email straight away explaining it was fully under my control and not being attacked. I filled out their attestation form of the same, which is supposed to grant you 24 hrs before they block all traffic. I had responded within a couple hours.
They blackholed my server right when things were starting to hockey stick and took until the following day to unlock it. I lost a ton of users to a full outage while I had an unplanned migration to AWS.
When they finally unlocked my server their tech responded to my ticket begging them to unlock my server with an extremely condescending “It’s unlocked but I’m certain you’re not in control of this server”
Anyway, never again. I’ll pay a premium for stability and intelligent support staff and systems that don’t automatically blackhole your server.
The pricing here always confused me. I thought VPS's were supposed to be cheaper than hardware. At $50+/month you're into dedicated pricing in my mind.
What? Hetzner will sell you a 1c / 2gb instance for $5/mo and you can do an incredible amount with that. 99% of people’s side projects can run on that. Hell, plenty of startups could run on that.
YMMV, but this doesn’t pass the smell test for me. modest sized projects will freeze a $5 droplet during an install and build for me. I’ve tried across multiple providers. DigitalOcean, Linode, and Vultr. Tried all of them, and a $5 droplet can’t handle anything but a static site IMO. I don’t disagree a startup can go far on a small droplet but inflation has made $5 droplets too slow to be useful IME.
I use Vultr, Digital Ocean, and Scaleway quite extensively, and my personal anecdote is the Scaleway, while gives a great bang for buck, comes with disk speed issues in their low priced servers.
None of them are. There’s no way you could make money doing that in the VPS market especially with the ludicrous bandwidth costs at AWS.
Most VPS providers rent rack space at large colocation facilities in multiple cities. In some cases they locate right next to major peering points for very good peering and very cheap bandwidth.
Vultr in particular is massively underrated and is certainly good enough to use in production.
One caveat though: most of these providers don’t handle sanitization of storage or encryption at rest as thoroughly as the big clouds do. I’d recommend handling your own encryption at rest with these if you have sensitive data. See: Vault, LUKS, etc.
Digital Ocean does offer managed Postgres with encryption in transit and at rest. I had a nice experience with their Postgres offering. They also have an S3 equivalent (spaces) and managed Redis.
I think it’s fair to not call it equivalent because of that, but it’s also fair to acknowledge what is unsupported is pretty power user level stuff for the most part. I used spaces in s3 compatibility for a few use cases and never ran into the problem that what I was trying to do was unsupported. Granted I was not trying to do fancy things, but still.
Some examples of things that I used the S3 compat for that worked just fine:
Yep, for basic object storage via boto3 etc. it mostly works fine. hence why I called it a nit pick.
There are additional interop things like Apache Spark <--> Spaces that don't work like Spark <--> AWS S3 (endpoint URL config issue which I lost the will to sort out as it was quicker to just use AWS S3 :shrug:).
It doesn’t have full feature parity, but as the post you shared says, they “aim for full interoperability with S3”. If you have an S3 client doing basic object storage stuff, it will work fine with Spaces.
Eh, not Apache Spark <--> Spaces. Doesn't work out the box unlike Spark <--> AWS S3.
See other child comment above. Faster to use AWS S3 instead of trying to fix it (IIRC can't set the endpoint url or something, which is required for Spaces). :shrug:
I appreciate I'm making a "No True Scotsman" point, but Spark isn't really "a typical S3 client doing basic object storage stuff". I'm thinking of your standard web app persisting uploads and static files.
I pay far less for AWS than I ever did for the providers in this article. Clearly YMMV depending on your workload and also how well you optimise for a cloud providers strengths. Hence making broad statements is unhelpful. There's a place for AWS/GC/Azure, and also for the likes of Vultr/Do and also for rolling your own. It all depends on what you need to achieve. There's no one size fits all.
There is also a large number of providers well below the prices of DO and Vultr as well. That are just as big, more profitable, and provide you with a better product and service.
Its crazy to me how many smart people don't realize they are only looking at the most expensive options out there.
It's like thinking the Ferraris and Porsches are your only two options for getting from point a to b
No. You're paying so much because (a) they make a lot of money ($20bn yearly profit) and (b) they have an extremely large headcount which are building the large number of services they provide.
Never heard of them running on other providers though.
Surely that would a problem for all of the compliance e.g. HIPAA, FIPS ?
Both A and B are true. But that doesn't preclude them from still charging a heavy premium over other providers in the market. Which is exactly what is happening, the raw COGS of operating a provider are dramatically lower than the prices they charge.
Yes, of course they have their own infra. But there are also reasons to partner. Not having a presence in a region. Not wanting to do the initial capital outlay. Offloading risk. etc.
I've seen many other services do this type of partnering and sometimes expect them to depending on the size, but this is definitely new information to me about Amazon doing this. Thanks.
The machine on one provider might be experiencing more load than a similar machine at a different provider, biasing the results.
That being said, Vultr has been good for our purposes.