If Modal's customers' workloads are mainly GPU-bound, then the performance hit of gvisor isn't as big as it might be for other workloads. GPU activity does have to go through the fairly heavyweight nvproxy to be executed on the host, but most gpu activity is longer-lived async calls like running kernels so a bit of overhead in starting / retrieving the results from those calls can be tolerated.

So I can agree that perhaps Modal might make sense for LLM's but they position themselves as sandbox including something like running python code etc. and some of this may be more intensive in workflows than others so I just wanted to point it out

Fly.io uses firecracker so I kinda like firecracker related applications (I tried to run firecracker myself its way too hard to build your own firecracker based provider or anything) and they recently released https://sprites.dev/

E2B is another well known solution out there. I have talked to their developers once and they mentioned that they run it on top of gcp

I am really interested in kata containers as well because I think kata runs on top of firecracker and can hook with docker rather quickly.

Kata runs atop many things, but is a little awkward because it creates a "pod" (VM) inside which it creates 1+ containers (runc/gVisor). Firecracker is also awkward because GPU support is pretty hard / impossible.

[1] https://fly.io/blog/wrong-about-gpu/

For my use cases usually, its Firecracker but I can now see why company like Modal would use gvisor because they focus a lot (and I mean a lot) on providing gpu access. I think that its one of their largest selling points or one of them, for them compute is secondary customer and gvisor's compute performance hit is a well worth trade off for them

Thanks for trying to explain the situation!

He discovered that on some airlines (I think American?), they use an advanced fortinet firewall that doesn't just look at the SNI -- it also checks that the certificate presented by the server has the correct hostname and is issued by a legit certificate authority.

My friend got around that restriction by making the tunnel give the aa.com SNI, and then forward a real server hello and certificate from aa.com (in fact I think he forwards the entire TLS 1.2 handshake to/from aa.com). But then as soon as the protocol typically would turn into encrypted application data, he ignores whatever he sent in the handshake and just uses it as an encrypted tunnel.

(The modern solution is just to use TLS 1.3, which encrypts the server certificate and hence prevents the firewall from inspecting the cert, reducing the problem back to just spoofing the SNI).

It's meant to get around the great firewall in China, so it has to avoid the GFW's active probers that check to make sure the external website is a (legit) host. However a friend was able to get it to work American's in-flight firewall if the proxy SNI is set to Google Analytics.

[1] https://github.com/XTLS/Xray-core

Thankfully for my blood pressure, whoever had set it up had left some kind of management portal accessible on a random high port number and it contained some strings which led me back to the Xray project.

Hah I was just about to say the same thing! I just got home from a ~3 week cruise. Internet on the ship was absurdly expensive ($50/day). And its weird - they have wifi and a phone app that works over the internet even if you don't pay. Google maps seemed to work. And my phone could receive notifications from apple just fine. But that was about it.

I spend some time staring at wireshark traces. It looks like every TCP connection is allowed to send and receive a couple packets normally. Then they take a close look at those packets to see if the connection should be allowed or blocked & reset. I'm not sure about other protocols, but for TLS, they look for a ClientHello. If preset, the domain is checked to see if its on a whitelist. Anything on their whitelist is allowed even if you aren't paying for internet. Whitelisted domains include the website of the cruise company and a few countries' visa offices. The cruise app works by whitelisting the company's own domain name. (Though I'm still not sure how my phone was getting notifications.)

They clearly know about the problem. There's some tools that make it easy to work around a block like this. But the websites for those tools are themselves blocked, even if you pay for internet. :)

If you figure out how to take advantage of this loophole, please don't abuse it too much or advertise the workaround. If it gets too well known or widely abused, they'll need to plug this little hole. And that would be a great pity indeed.

Don't forget to price in the costs of installing and maintaining a WiFi network that works consistently in a metal ship whose interior is composed from prefab metal modules. (Hint: every cabin, every space, has one or more APs).

I haven't done the math, and I'm sure they profit on the offering, but I doubt it's as egregious as these replies make it sound.

(I thought about this a bit when I was on a cruise that offered Starlink this past summer.)

Edit: also don't forget that everyone gets free WiFi, it's just that internet access is restricted for guests who don't pay. So it does need to support the ship's full complement and passengers.

I’m sure servicing thousands of people via starlink is expensive. But the cost is amortised over the number of people using it. Thousands of users should make internet access cheaper, not make it more expensive.

They also don’t provide “normal” internet speeds. I was usually getting about 20kBps - which is painfully slow. I tried to have a zoom call on the one day I paid for internet, and every minute or two we would get a latency spike of 10+ seconds. Those latency spikes went away on other days, but the speed never improved much.

The ship I was on is apparently quite old by modern standards. Maybe they don’t have enough starlink satellites installed or something. (It was definitely starlink). But if that’s the case, it makes the price they’re asking all the more outrageous. For $50/day I could probably bring my own starlink satellite on board and it would come out cheaper.

I have never used Starlink otherwise and, frankly, expected much worse service - especially on a cruise ship.

I'd definitely be unhappy paying $50/day for what you described. But I paid less (there was a discount for buying a package ahead of time for my family's devices) and got better service it sounds like.

Starlink hardware (aka community hub) is $1.25M. Actual bandwidth cost is 75k per gbps per month.

I never figured out a way to route internet on my phone through my laptop. But it was probably for the best. It was lovely spending a few weeks with no internet connection on my phone, in arms reach away at all times.

Almost all of these special pricing/zero-rating schemes will include platform push in the zero rated traffic. Can't use anything without it, and most of the platforms have public pages describing how to identify their traffic, because there's lots of networks that want to allow it.

But basically you get to see a bunch of destinations while all your travel is organized for you, you never have to switch rooms and constantly pack/unpack, and the actual travel part is infinitely more comfortable.

A room and sundeck and pool beats a plane seat or train seat any day.

I'm not into cruises myself, but the appeal seems pretty understandable in terms of convenience.

They’re also far less expensive than many other vacations, especially if you have kids and are considering Disney stuff.

Still a human Petri dish.

Go price out hotels and food in any major destination for one week. Now go price out a cruise for one week which also includes entertainment and a travel component. Somehow, the cruise is CHEAPER and offers more.

That's it. That's the whole answer.

Long hours and low pay - Some workers face shifts of more than 12 hours a day, seven days a week, often without overtime pay. Wages can be very low, sometimes below $20 per day, though tips can supplement income. Workers often live in small, shared cabins with limited personal space. Ships often registered in countries with lax regulations. No pay between workers contracts

These are ONLY some of the reasons ....

There is a level of convenience that is hard to get elsewhere.

I went on a Disney cruise 2 summers ago. All restaurants were in walking distance. All of deck 5 was dedicated to child care. They took you straight to excursions. Family was close, but not too close.

There were some downsides, too, but let's not focus on those. I think the "king" reason we went is because the grandparents were paying and they wanted everyone to be "there" and not leaving. I think the main reason we aren't going again is cost.

You're overlooking how incentives and motivations work. The gp (and their employer) wants to integrate the advertisement snippet -- even with broken XHTML -- because they receive money for it.

The semantic data ("advertiser's message") is more important than the format ("purity of perfect XHTML").

Same incentives would happen with a jobs listing website like Monster.com. Consider that it currently has lots of red errors with incorrect HTML: https://validator.w3.org/nu/?doc=https%3A%2F%2Fwww.monster.c...

If there was a hypothetical browser that refused to load that Monster.com webpage full of errors because it's for the users' own good and the "good of the ecosystem"... the websurfers would perceive that web browser as user-hostile and would choose another browser that would be forgiving of those errors and just load the page. Job hunters care more about the raw data of the actual job listings so they can get a paycheck rather than invalid <style> tags nested inside <div> tags.

Those situations above are a different category (semantic_content-overrides-fileformatsyntax) than a developer trying to import a Python library with invalid syntax (fileformatsyntax-Is-The-Semantic_Content).

EDIT reply to: >Make the advertisement block an iframe [...] If the advertiser delivers invalid XHTML code, only the advertisement won't render.

You're proposing a "technical solution" to avoid errors instead of a "business solution" to achieve a desired monetary objective. To re-iterate, they want to render the invalid XHTML code so your idea to just not render it is the opposite of the goal.

In other words, if rendering imperfect-HTML helps the business goal more than blanking out invalid XHTML in an iframe, that means HTML "wins" in the marketplace of ideas.

The real problem is the benefits of xhtml are largely imaginary so there isn't really a motivation to do that work.

Make the advertisement block an iframe with the src attribute set to the advertiser's URL. If the advertiser delivers invalid XHTML code, only the advertisement won't render.

I know which I, as a user, would prefer. I want to use a browser which lets me see the website, not just a parse error. I don’t care if the code is correct.

I don't know much about Ada. Is its type system any better than Rust's?

But a noteworthy excerpt: ```

Ada programs tend to define types of the problem to be solved. The compiler then adapts the low-level type to match what is requested. Rust programs tend to rely on low-level types.

That may not be clear, so two examples may help:

    Ada programmers prefer to specify integer types in terms of the ranges of values they may take and/or the precision of floating-point types in terms of digits. I ended up doing this at least once, where on Day 23 I specified a floating-point type in terms of the number of digits it should reproduce accurately: Digits 18. The compiler automatically chose the most appropriate machine type for that.

    Ada arrays don't have to start from 0, nor even do they have to be indexed by integers. An example of this appears below.

```

This seems to be an artifact of the domain that Rust is currently being used in. I don't think it's anathema to Rust to evolve to be able to add some of these features. char indexed arrays are something I've used a lot (most via `char c - 'a'`\, but native support for it would be nice).

F'90 added arbitrary lower bounds on assumed-shape dummy arrays, as well as on allocatables and pointers. Still pretty portable, though more confusing cases were added. F'2003 then added automatic (re)allocation of allocatables, and the results continue to astonish users. And only two compilers get them right, so they're not portable, either.

Ada's array indexing is part of its type system. Fortran's is not (for variables).

It's just much, much easier and more ergonomic in Ada.

There was a header only library on the Windows SDK that would wrap those HANDLEs into more specific types, that would still be compatible, while providing a more high level API to use them from C.

Unfortunely there is not much left on the Internet about it, but this article provides some insight,

https://www.codeguru.com/windows/using-message-crackers-in-t...

Naturally it was saner just to use TP/C++ alongside OWL, C++ with MFC back then, or VB.

  type Temperature_K is digits 6 range 0 .. <whatever is reasonable upper bound in your domain>;

C++: That, but with a billion operator overloads and conversion operators so they feel just like native integers.

Of course you can use a unit type that handles conversions AND mathematical operations. Feet to meter cubed and you get m³, and the library will throw a compile error if you try to assign it to anything it doesn't work with (liters would be fine, for example)

> Feet to meter cubed and you get m³, and the library will throw a compile error if you try to assign it to anything it doesn't work with (liters would be fine, for example)

Liters would not be fine if you are using standard floating point values since you lose precision moving decimal points in some cases. Maybe for your application the values are such that this doesn't matter, but without understanding your problem in depth you cannot make the generic statement.

I could write books (I won't but I could) on all the compromises and trade offs in building a unit type library.

News flash, C++ has every conceivable feature, it's the reason why it is so unwieldy. But you can even plug in a fucking GC if you so desire. Let alone stuff like basic meta programming.

https://mpusz.github.io/mp-units/latest/

The 'explicit' would force you to use static_cast

And as usual, it mostly comes with zero overhead, beyond optional runtime range checking and unit conversions.

But C++ is a meta-programming language. Making up your own types with full operator overloading and implicit and explicit conversions is rather easy.

And the ADA feature of automatically selecting a suitable type under the hood isn't actually that useful, since computers don't really handle that many basic types on a hardware level. (And just to be clear, C++ templates can do the same either way)

In C++ operator overloading can easily mess with fundamental mechanisms, often intentionally; in Python it is usually no more dangerous than defining regular functions and usually employed purposefully for types that form nice algebraic structures.

    class MyNum(int):
        def __add__ (self, other):
            return super().__add__(other) * 10
        
        
    n = MyNum(12)
    a = 45
    print(n + a) # oops

    class MyNum(int):
        def ten_times_sum (self, other):
            return (self+other)* 10

    n = MyNum(12)
    a = 45
    print(n.ten_times_sum(a)) 

Operator overload is indeed syntactic sugar for function calls, regardless of the language.

By the way, you can overload >> in Python via rshift(), or __rshift__() methods.

You mean accumulated prejudices, myths, and superstitions that most in any given community (programming language related or not) won't challenge for fear of being cast out of the group for heresy.

Scott in the video makes the argument that sharp knives are safer, because you don't have to use as much force. But the only time I've ever cut myself with a knife in the kitchen have been with very sharp knives, eg one time handling it while washing the blade.

It’s usually only worth it after ~tens of megabytes, but vast majority of npm packages are much smaller than that. So if you can skip it, it’s better.