Making things seem simpler by finding an abstraction over them is great!
Making things seem simpler by finding an abstraction that leaks some details so you sometimes need to handle them at a high level is... less great. But sometimes still a win.
Making things seem simpler but neither handling the complexity nor allowing your users to handle the complexity is technical debt at best.
(0) Memory is infinite, and memory allocation will always succeed. If it doesn't, aborting the entire process is the right course of action.
(1) Communication over a network is reliable and data will arrive in the same order in which it was sent.
And here are some examples of unacceptably leaky abstractions:
(0) A few basic data structures [in particular, growable arrays and hashtables] suffice to implement every algorithm you might ever want to implement.
(1) Strings are arrays of characters, and thus you can efficiently index them by character.
(2) Every [nonempty] string is a valid file name or URL.
(3) Every external resource can be manipulated using an API that presents it as a sequence of bytes.
(4) An event loop is the universal solution to all asynchronous I/O problems.
I think in general, building on top of leaky abstractions is a form of technical debt, and you have to be aware of how much you're accumulating and what you're getting in exchange. Also, complexities stack super-linearly. I'm tempted to say, additionally, that building shallowly atop a leaky abstraction is less problematic than pushing leaks further down in a tall stack... but on reflection I have no idea how I'd support that.
So, what I meant to say in my previous post is that, as far as I can tell, good leaky abstractions only arise when doing systems programming, or, more precisely, at the boundary between low-level system services (e.g. runtime systems for managed languages) and the applications that consume them. Abstraction leaks in high-level application code are likely to be bad.
You need to worry about running out of memory precisely because the abstraction is leaky. If you need to micromanage, it's perhaps not leaky, but it's also not much of an abstraction.
Exploring this example a little deeper, if our model is "we never run out of memory, but our program may stochastically die at any point in time", then that's not as leaky as "we never run out of memory", and our program may die for reasons outside our control so it's a good idea to be robust against those. But now let's say we need to send a request to a service, and the service disconnects when we've sent the request. With a stochastic model of failures, we clearly should just re-send. But if the problem was that our payload demanded more memory than could be provided, retrying repeatedly will just keep killing servers.
We probably want a model of "we never run out of memory, our program may die at any point in time, possibly correlated with inputs", which I would contend is actually not very leaky. We still spill over into performance, but most things spill over into performance...
I would say that leaky abstractions constitute debt relative to non-leaky equivalently powerful abstractions. They may still pay their way.
In practice, I don't worry. I'm fine with the leaky abstraction, because the benefit of fixing the leak (reducing the probability that OOM crashes my program from almost zero to exactly zero) isn't worth the cost (increased complexity). But see below.
> If you need to micromanage, it's perhaps not leaky, but it's also not much of an abstraction.
I'm not sure about this. There are levels of abstraction and degrees of micromanagement. For instance, `malloc` lets you micromanage the memory layout of your data structures, and `mmap` and `mprotect` also let you micromanage what you can do with each memory page, but they're still abstractions over even lower-level concerns, like the mapping between physical and virtual memory addresses.
> We probably want a model of "we never run out of memory, our program may die at any point in time, possibly correlated with inputs", which I would contend is actually not very leaky.
Depends on what kind of program you're writing. If your web crawler crashes, you just restart it. If your web server crashes, your users will be annoyed, but at least your database won't be compromised. (You're properly using transactions, right?) But if your DBMS crashes, your data might even be in an irrecoverably inconsistent state, unless you properly anticipated all (software) causes for it to crash.
In practice, you're being sloppy or what you've described is the relevant worrying - considering whether your program has appreciable odds of OOM and the odds of things being particularly bad when that happens, and then deciding that additional worrying is unnecessary. But this can need re-evaluating as circumstances change.
> There are levels of abstraction and degrees of micromanagement.
No disagreement there.
> Depends on what kind of program you're writing.
A bit, but mostly in terms of performance.
> But if your DBMS crashes, your data might even be in an irrecoverably inconsistent state, unless you properly anticipated all (software) causes for it to crash.
I disagree. If you have designed your program with an assumption that it might fail at any point regardless of reason, then you don't need to have "properly anticipated all (software) causes for it to crash", and further you should be robust against some portion of non-software causes.
I'm unhappy with this line of reasoning. For me, an abstraction leak is a failure to behave as advertised. Since memory allocation can always fail, an abstraction that advertises “allocations always succeed” is necessarily leaky. Even if the leak isn't important in practice, the leak exists.
> If you have designed your program with an assumption that it might fail at any point regardless of reason
I'm not sure this is even possible (if I take nothing for granted, how do I do anything?), but I'd like to be proven wrong.
To my mind, it must fail to behave as advertised in a way that is meaningful. Otherwise, the differences are precisely what is being abstracted away.
> I'm not sure this is even possible (if I take nothing for granted, how do I do anything?), but I'd like to be proven wrong.
I don't mean "might act in arbitrary ways", but specifically "might be unexpectedly be terminated at any point".