kbumsik

I notice the spot cost has grown up for 3 months recently, especially in U.S. regions, while some other regions are not affected at all.

Why is it happening to the U.S. regions especially?

Doesn't Cloudflare KV have to be used with Cloudflare Workers?

The field it is used is very different. So why not?

Hmm I haven't have significant issues for almost 10 years. And a few years ago Dropbox introduced a folder-level rewind feature so you can easily restore any broken folder.

Anyway, YMMV.

I mean, why not?

I use both git and Dropbox, because I have multiple devices to sync: laptops, a desktop, and linux servers.

If you use a desktop and a laptop back and forth, git is quite inconvenient. You don't need to commit and push from the desktop to switch to the laptop with a folder level realtime sync solution.

Also, git only protects files that are staged, but Dropbox can restore any files to any point of file event (not periodic snapshot, but it keeps all file events). So combining both you are safe from any kinds of mistakes.

I use Dropbox because Google Drive sync wasn't stable enough for this kind of workflow but I am not sure if it has been improved lately.

Oh I am considering to us FSx for Lustre with S3 export/import for our ML workloads. Isn't it stable enough?

I just read this post, it looks like a demo for basic operations.

When I tried it at the end of the last year, ent works fine with Planetscale for basic reads. However, it fails to read when I use complex queries with Ent GraphQL integration.

ent added quite a lot of features last year so things might changed from the time when the blog post was written.

I really wanted to use Planetscale, but there are two blockers:

- No foreign key support

- No CTE support

I tried using Planetscale knowing that there is no foreign key support because it is fine to me. But I had to stop using Planetscale because of the second. My ORM (ent for golang) relies on CTE so I simply cannot run complicated queries.

So if you are considering Planetscale, test it enough especially if you are using an ORM.

Doesn't Consul have the similar storage limit btw?

I have seen very few strongly consistent distributed KV store that scales beyond 10GB+

Thanks for the answer!

AFAIK doesn't Consul also use Raft?

I have seen some issues around Consul these days.

As a person with no background in distributed systems, I am wondering why people choose Consul over alternatives. Are there features that etcd doesn't offer?

S3 works over HTTP, which means that it is designed to work over the internet.

Other protocols you mentioned, including NFS, does not work well over the internet.

Some of them are exclusively designed to work within the same network, or very sensitive to network latency.

And the original paper is published even before the first version of C was introduced.

Not in PyTorch. A model is Python dictionaries containing states and Python module/class objects. I don't know why the PyTorch team did this but that happened. Maybe it boils down to the point #2 I said.

There is nothing wrong with ONNX, but rather the limitations of PyTorch.

1. PyTorch model files are neither portable nor self-contained, because PyTorch model files are pickled Python classes containing weights so you need Python class code to run it.

Because it needs real Python code to run the models, PyTorch suffers from a numerous issues porting to other non-Python platforms such as ONNX.

PyTorch offers a way to export to ONNX but you will encounter various errors. [1]

Sure, you might be lucky enough to troubleshoot a specific model to export it to ONNX, but if your objective is to export arbitrary 964 models from a model zoo (TIMM) it is almost impossible.

2. There are organizational or cultural problems with it. Because of the above problem, PyTorch model needs to be designed with portability in mind from beginning. But porting & serving models are what engineers do, whereas researchers, who design models, don't care about it when writing papers. So it is often hard to use SOTA models that comes from an academic research.

[1] https://pytorch.org/docs/stable/onnx.html#limitations

> The thing most appealing about single server configs is the simplicity.

In my experience this ended up with more complicated.

Those systems are typically developed by people who already left and are undocumented, and they become extremely difficult to figure out the config (packages, etc files... oh, where even the service files are located?) and almost impossible to reproduce.

It might be okay to leave it there, but when we need to modify or troubleshoot the system a nightmare begins...

Maybe I was just unlucky, but at least k8s configs are more organized and simpler than dealing with a whole custom configured Linux system.

Very interesting. I have been using 4k 27" monitor but this makes me temping to switch to 32".

But I don't like that ugly and huge webcam. I would use my own webcam on top of it.

> via Mini DisplayPort 2.1

Why Mini Displayport? I haven't heard of Mini DP for years.

Why combining RAID 1 mirror and RAID 6 array in the same pool?

There is nothing wrong with committing generated code in Ent and Go. I don't know about Java but in Golang it does not create any platform-specific code in general, so committing code in the VCS would ensure reproducibility.

It is still OK to commit without generated code and things won't break as long as you pinned the package versions correctly. But some people don't like to put extra burden in the build phase. So it is totally organization's decision whether to commit generated code or not.

I started to like Go's code generation approach.

Code generation make you very productive since IDE's auto-complete and code search works perfectly and snappy with the generated code.

The only complaint is that it tends to make noise during code reviews because they include autogenerated code too.

I have no experience in Django but in Ent with GraphQL.

Ent is not a full-featured web framework so you need to implement many of features by your own or use other libraries (e.g. http server and session management).

If you are only looking for ORM + GraphQL then I highly recommend trying Entgql, an Ent extension for GraphQL with Gqlgen library [1]. Once you define an ORM schema, it will generate GraphQL Query for Relay server. Still you need to implement GraphQL Mutations by your own but at least it will create Input types for you (both for Create/Update).

[1]: https://github.com/99designs/gqlgen

Thinking about it, we also say “injured or dead" (사상자) when it's about the number.

I am Korean and we are getting "dead confirmed" number separately now, reported by hospitals. So yes, "cardiac arrest" is reported from the accident site and it is unlikely that they are going to survive.

I wonder how it is different in other countries.

Netmaker looks promising for k8s distros. It's shame that its license is SSPL.

Sounds like GPU uses system RAM like TPU as I can't find VRAM information in the page.

S3 Deep Archive is two time cheaper than B2. BUT you are not supposed to just rclone to S3 Deep Archive because it charges you based on the number of files you upload.

The upload fee (PUT request fee) is $0.065/1000files so it will charge you a lot when you have millions of files.

What is the maximum size of each .tar.zstd.ssl ?

We have a storage with 40TB with 100 million files, then what is the expected number of archive files?

Not sure in other countries, but in my country (Korea) they were called PDP and rarely called plasma TV even knowing the technology behind it. It's a bit funny thinking about "Plasma" but it should be fine.

BTW, it had been dead for 10+ years...time flies so fast.

Sounds very interesting! Do VMs work with a display session (Xorg) too?

Some GPU-enabled applications require an X server with a physical HDMI/Displayport attached.

You can use GraphQL for BFF btw.