The problem wouldn’t be the developers but the reverse engineers.

Though there are of course ways to RE without looking at what the original system does.

That’s typical for plain-text email which this is.

It was being compared to another implementation.

I’m quite certain it was being compared to mainline WINE, so no esync or fsync which themselves usually double FPS in CPU-bound scenarios.

Hers is actually better

[citation needed]

From what I gather from the ntsync feedback thread where some users have tested the WIP patches, it’s not clearly better than esync/fsync but rather slightly worse. Though that isn’t very clear data as it’s still in development. Still, if it was very clearly better than the status quo, we should have already seen that.

can be fully implemented in Wine

It cannot, hence the kernel patch.

It’ll be better but no one really knows the full concrete extend of improvement until it lands

I see no reason to believe it should be “better”. If anything, I’d expect slightly worse performance than esync/fsync because upstream WINE primarily wants a correct solution while the out-of-tree esync/fsync patches trade some correctness for performance in games.

Ideally, I’d like to be proven wrong; that ntsync is both correct and performant but that’s not what you should expect going into this.

That’s just an ACK and Elizabeth replied that she’ll resend again with further changes.

Nothing is in any tree that is going to Linus yet AFAICT.

We could be reasonably sure that it’ll go to Linus if it’s in char-misc but that hasn’t happened yet. I’m also actually not sure whether Greg’s or Arnd’s tree is the canonical one there.

Fake news. It’s merely a re-send of the patches; nothing landed yet.

The Phoronix article and their title makes that clear; you editorialised it to state differently. (Also, that’s…cringe.)

Pricing for teams is the same as regular monthly pricing (per user ofc.). It’s only charged if the user actually used Kagi in the billing period too which, honestly, is the only acceptable way to do this. Charging despite not actually providing a service would be quite dishonest IMHO.

It’s pretty cool that the app could soon be shown on the DMA search choice screen. It’s bound to cause some confusion though given that, unlike all other search apps, it’s paid.

What the docs are quite a bit unclear though is whether the LLM stuff is actually private and does not exfiltrate your input. It says Kagi itself doesn’t store them or use them for any nefarious purposes but it makes no comment on what the LLM providers do with your inputs. That’s… worrying. Please clarify @kagihq@mastodon.social.

Read closely and you’ll notice they used a thumb drive.

People usually refer to the act of copying the data directly onto the device as something other than “copying” to differentiate from copying the ISO as a file to a filesystem on the drive.

How many millions did this utterly useless rebranding cost?

Why haven’t the people who decided to waste money on this rather than retaining talent gotten fired yet?

Note that the clients being FOSS is of little relevance because all they do is forward a recording to a blackbox proprietary service run by a for-profit company.

The code that has access to your audio and does the actual task at hand is not FOSS in the slightest.

I have no idea what that is.

Well, it’s an engineer who said it, not a sales rep.

Not if it’s 1999-12-31 ;)

Sure :)

I knew about bit rot but thought the only solution was something like a zfs pool.

Right. There are other ways of doing this but a checksumming filesystem such as ZFS, btrfs (or bcachefs if you’re feeling adventurous) are the best way to do that generically and can also be used in combination with other methods.

What you generally need in order to detect corruption on ab abstract level is some sort of “integrity record” which can determine whether some set of data is in an expected state or an unexpected state. The difficulty here is to keep that record up to date with the actually expected changes to the data.
The filesystem sits at a very good place to implement this because it handles all such “expected changes” as executing those on behalf of the running processes is its purpose.

Filesystems like ZFS and btrfs implement this integrity record in the form of hashes of smaller portions of each file’s data (“extents”). The hash for each extent is stored in the filesystem metadata. When any part of a file is read, the extents that make up that part of the file are each hashed and the results are compared with the hashes stored in the metadata. If the hash is the same, all is good and the read succeeds but if it doesn’t match, the read fails and the application reading that portion of the file gets an IO error that it needs to handle.

Note how there was never any second disk involved in this. You can do all of this on a single disk.

Now to your next question:

How do I go about manually detecting bit rot?

In order to detect whether any given file is corrupted, you simply read back that file’s content. If you get an error due to a hash mismatch, it’s bad, if you don’t, it’s good. It’s quite simple really.

You can then simply expand that process to all the files in your filesystem to see whether any of them have gotten corrupted. You could do this manually by just reading every file in your filesystem once and reporting errors but those filesystems usually provide a ready-made tool for that with tighter integrations in the filesystem code. The conventional name for this process is to “scrub”.

How do I go about manually detecting bit rot? Assuming I had perfect backups to replace the rotted files.

You let the filesystem-specific scrub run and it will report every file that contains corrupted data.

Now that you know which files are corrupted, you simply replace those files from your backup.

Done; no more corrupted files.

Is a zfs pool really that inefficient space wise?

Not a ZFS pool per-se but redundant RAID in general. And by “incredibly costly” I mean costly for the purpose of immediately restoring data rather than doing it manually.

There actually are use-cases for automatic immediate repair but, in a home lab setting, it’s usually totally acceptable for e.g. a service to be down for a few hours until you e.g. get back from work to restore some file from backup.

It should also be noted that corruption is exceedingly rare. You will encounter it at some point which is why you should protect yourself against it but it’s not like this will happen every few months; this will happen closer to on the order of every few decades.

To answer your original question directly: No, ZFS pools themselves are not inefficient as they can also be used on a single disk or in a non-redundant striping manner (similar to RAID0). They’re just the abstraction layer at which you have the choice of whether to make use of redundancy or not and it’s redundancy that can be wasteful depending on your purpose.

if it’s a 1:1 full disk image, then there’s almost no difference with the costs of raid1

The problem with that statement is that you’re likening a redundant but dependant copy to a backup which is a redundant independent copy. RAID is not a backup.

As an easy example to illustrate this point: if you delete all of your files, they will still be present in a backup while RAID will happily delete the data on all drives at the same time.

Additionally, backup tools such as restic offer compression and deduplication which saves quite a bit of space; allowing you to store multiple revisions of your data while requiring less space than the original data in most cases.

In this case he’s talking about restic, which can restore data but very hard to do a full bootable linux system - stuff needs to be reinstalled

It’s totally possible to make a backup of the root filesystem tree and restore a full system from that if you know what you’re doing. It’s not even that hard: Format disks, extract backup, adjust fstab, reinstall bootloader, kernels and initrd into the boot/ESP partition(s).

There’s also the wasteful but dead simple method to backing up your whole system with all its configuration which is full-disk backups. The only thing this will not back up are EFI vars but those are easy to simply set again or would just remain set as long as you don’t switch motherboards.

I’m used to Borgbackup which fulfils a very similar purpose to restic, so I didn’t know this but restic doesn’t appear to have first-class support for backing up whole block devices but it appears this can be made to work too: https://github.com/restic/restic/issues/949

I must admit that I also didn’t think of this as a huge issue because declarative system configuration is a thing. If you’re used to it, you have a very different view on the importance of system configuration state.
If my server died, it’d be a few minutes of setting up the disk format and then waiting for a ~3.5GiB download after which everything would work exactly as it did before modulo user data. (The disk format step could also be automatic but I didn’t bother implementing that yet because of https://xkcd.com/1205/.)

I was thinking whether I should elaborate on this when I wrote the previous reply.

At the scale of most home users (~dozens of TiBs), corruption is actually quite unlikely to happen. It’ll happen maybe a handful of times in your lifetime if you’re unlucky.

Disk failure is actually also not all that likely (maybe once every decade or so, maybe) but still quite a bit more likely than corruption.

Just because it’s rare doesn’t mean it never happens or that you shouldn’t protect yourself against it though. You don’t want to be caught with your pants down when it does actually happen.

My primary point is however that backups are sufficient to protect against this hazard and also protect you against quite a few other hazards. There are many other such hazards and a hard drive failing isn’t even the most likely among them (that’d be user error).
If you care about data security first and foremost, you should therefore prioritise more backups over downtime mitigation technologies such as RAID.

ZFS and BTRFS’ integrity checks are entirely independent of whether you have redundancy or not. You don’t need any sort of RAID to get that; it also works on a single disk.
The only thing that redundancy provides you here is immediate automatic repair if corruption is found. I’ve written about why that isn’t as great as it sounds in another reply already.

Most other software RAID can not and does not protect integrity. It couldn’t; there’s no hashing. Data verification is extremely annoying to implement on the block level and has massive performance gotchas, so you wouldn’t want that even if you could have it.

RAID 1 can absolutely be faster than a single disk for read perf, and on Linux it is tuned to be faster.

You’re missing the point entirely. I never said to use a single disk, I explicitly compared it to RAID0.

As far as data security is concerned, JBOD/linear combination and RAID0 are the same, so you’d obviously use RAID0 if you didn’t need redundancy.

You should probably say “NVK users” as most Nvidia GPU users will not be using NVK quite yet.

staging rebuild cycles only happen every two weeks or so.

The reason is always that something changed and causes all dependent packages to change, requiring a rebuild of those too.

It depends on your uptime requirements.

According to Backblaze stats on similarly modern drives, you can expect about a 9% probability that at least one of those drives has died after 6 years. Assuming 1 week recovery time if any one of them dies, that’d be a 99.997% uptime.

If that’s too high of a probability for needing to run a (in case of AWS potentially very costly) restore, you should invest in RAID. Otherwise, that money is better spent on more backups.