What's more, I have no complaints about cgo. It hasn't been a performance problem at all (my game runs smoothly at 240hz). And the interface for binding to C is quite simple and nice to work with (I made my own bindings to SDL2, OpenGL, and Steamworks).
Cgo did make cross-compiling trickier to set up, but I managed to do it with a fair amount of elbow grease (Windows and MacOS builds on my Linux desktop).
Copying in a response to a similar question on Reddit:
I should really add some discussion around BLAS in particular, which has an good implementation[0] of the float32 dot product that outperforms any of the float32 implementations in the blog post. I'm getting ~1.9m vecs/s on my benchmarking rig.
However, that BLAS became unusable for us as soon as we switched to quantized vectors because there is no int8 implementation of the dot product in BLAS (though I'd love to be proven wrong)
I did consider Avo! I even went as far as to implement a version using Avo since it has a nice dot product example I could use as a starting point. But ultimately, for as small as these functions are, I felt that Avo was an unnecessary extra layer to grok. Additionally, it's x86-only, and I knew in advance I'd want to implement an ARM version as well since we also do some embeddings stuff locally.
If I were to ever take this further and add loop unrolling or something, I'd absolutely reach for Avo
> The default concurrency GOMAXPROCS is almost never what I want.
FWIW, the default concurrency has been changed to "unlimited" since the 0.1.0 release.
> Aggregated errors are almost never what I want.
Out of curiousity, what do you want? There is an option to only keep the first error, and it's possible to unwrap the error to an array of errors that compose it if you just want a slice of errors.
> Using it would place a burden on any reader that isn't familiar with the library
Using concurrency in general places a burden on the reader :) I personally find using patterns like this to significantly reduce the read/review burden.
> FWIW, the default concurrency has been changed to "unlimited" since the 0.1.0 release.
Nice! Will that end up on Github?
> Out of curiousity, what do you want
Most often I want to return just the first error. Some reasons: (1) smaller error messages passed across RPC boundaries (2) original errors can be inspected as intended (e.g. error codes) (3) when the semantics are to cancel after the first error, the errors that come after that are just noise.
A couple other thoughts
- I think the non-conc comparison code is especially hairy since it's using goroutine pools. There's nothing wrong with that and it's fast, just not the easiest to work with. Often goroutine overhead is negligible and I would bound concurrency in dumber ways e.g. by shoving a sync.Semaphore into whatever code I otherwise have
- I like that errgroup has .Go() block when concurrency limit is reached. Based on a quick skim of the code I think(?) conc does that too, but it could use documentation
It's already there! I just haven't cut a release since the change.
> Most often I want to return just the first error.
In many cases, I do too, which is why (*pool).WithFirstError() exists :)
> original errors can be inspected as intended
If you're using errors.Is() or errors.As(), error inspection should still work as expected.
> Often goroutine overhead is negligible and I would bound concurrency in dumber ways
Yes, definitely. And that's what I've always done too. However, I've found it's surprisingly easy to get subtly wrong, especially when modifying code I didn't write, and even more especially if I want to propagate panics (which I do, though that seems to be a somewhat controversial opinion in this thread). Conc is intended to be a well-known pattern that I don't have to think about too much when using it.
> I think(?) conc does that too, but it could use documentation
It does! I'll update the docs to make that more clear.
Fair criticism. The nice thing about the current API is it works with any input that's iterable (channels, slices, readers, etc.) and any output (callback, channel, append to slice, etc.). In most code I write, I avoid channels because I find them easy to misuse. I used channels in the examples because it's the easiest way to represent "some potentially unbounded stream of input."
I just wrote a bunch of separate short functions for different types (channels, slices, maps). There is more code duplication but code itself is simpler, and less boiler-platey and more embeddable.
For example
out := MapSlice(
func(i int) string { return fmt.Sprintf("-=0x%02x=-", i) },
MapSlice(
func(i int) int { return i + 1 },
MapSlice(
func(i int) int { return i * i },
GenSlice(10, func(idx int) int { return idx }),
),
),
)
I heavily used errgroup before creating conc, so the design is likely strongly influenced by that of errgroup even if not consciously. Conc was partially built to address the shortcomings of errgroup (from my perspective). Probably worth adding a "prior art" section to the README, but many of the ideas in conc are not unique.
> In go land, this seems desirable.
I mostly agree, which is why `Wait()` propagates the panic rather than returning it or logging it. This keeps panics scoped to the spawning goroutine and enables getting stacktraces from both the spawning goroutine and the spawned goroutine, which is quite useful for debugging.
That said, crashing the whole webserver because of one misbehaving request is not necessarily a good tradeoff. Conc moves panics into the spawning goroutine, which makes it possible to do things like catch panics at the top of a request and return a useful error to the caller, even if that error is just "nil pointer dereference" with a stacktrace. It's up to the user to decide what to do with propagated panics.
> That said, crashing the whole webserver because of one misbehaving request is not necessarily a good tradeoff. Conc moves panics into the spawning goroutine, which makes it possible to do things like catch panics at the top of a request and return a useful error to the caller, even if that error is just "nil pointer dereference" with a stacktrace. It's up to the user to decide what to do with propagated panics.
The problem is that panics aren't "goroutine scoped" in terms of their potential impact. So it really shouldn't be up to the user to decide how to handle a panic. Application code shouldn't handle panics at all! They're not just a different way to yield an error, they're critical bugs which shouldn't occur at all.
> panics aren't "goroutine scoped" in terms of their potential impact
I'm with ya there. However, there are also many classes of logic errors that are not goroutine-scoped. And there are many panics that do not have impact outside of the goroutine's scope. In my experience, this is true of most panics.
In practice, panics happen. They are (almost) always indicative of a bug, and almost always mean there is something that needs fixed. However, if a subsystem of my application is broken and panicking, there's a pretty good chance that reporting the panic without crashing the process will provide a better end user experience than just blowing up.
Yes, that means I'm accepting the risk that my application is left in an inconsistent state, but coupled with good observability/reporting, that's a tradeoff I'm willing to make.
(bonus: this is especially true when propagating panics allow me to capture more debugging information to fix the panics faster)
I guess this is the crux of the issue. I don't think this is true, or needs to be true. It certainly hasn't been my experience. I think assuming panics are normal will take you down some paths that make it basically impossible to write reliable software. But, to each their own.
> I'm accepting the risk that my application is left in an inconsistent state,
Inconsistent state makes it impossible to reason about your program's execution or outcomes. An account value that previously had balance = 0 may now have balance = 1000. Is this acceptable risk?
Since defers run during panics for exactly this reason, no. You can in fact guarantee that is not the case.
Runtime-safety "panics" in Go, like concurrently modifying and iterating a map that can lead to other memory being corrupted, tend to abort the whole process immediately and not be suppress-able panics.
> Runtime-safety "panics" in Go, like concurrently modifying and iterating a map that can lead to other memory being corrupted, tend to abort the whole process immediately and not be suppress-able panics.
> The usual way to report an error to a caller is to return an error as an extra return value. . . . But what if the error is unrecoverable? Sometimes the program simply cannot continue. For this purpose, there is a built-in function panic that in effect creates a run-time error that will stop the program
Panics express unrecoverable failures. This is plainly stated in the language documentation. There are exceptions to this rule, but they are exceptional.
It is absolutely a correctness issue. Panics do not provide safety guarantees that generalize enough that it is safe to arbitrary recover from them. The statement in the previous sentence is not a subjective opinion, it's a statement of fact. I'm not sure how else to convey this information.
When some bit of code invokes `panic` it is saying that there is an error which is unrecoverable, and the default expectation is that the process will terminate. There is no way to assert that panics do not violate runtime or memory model expectations. They can.
> An account value that previously had balance = 0 may now have balance = 1000. Is this acceptable risk?
Your entire web app process crashes due to a panic every time a request triggers an extremely rare edge case. A hacker discovers this and uses it to conduct a DoS attack. Is this acceptable risk?
This is equivalent to asking "Why the heck are you writing code with bugs?"
Sure, if we could write code without bugs, we wouldn't need to suppress panics.
But since we do tend to write code bugs and some of them are bugs that can be detected by the runtime - we get panics.
If you hate panics, you can do better than Go and go for a language with a stronger type system, where you won't get nil pointer panics or interface conversion panics, but even an almost onerously-tyepesafe language like Haskell still panics on some bogus operations such as division by zero or trying to read the head of an empty list. Perhaps Idris really have no runtime errors but they are quite niche.
Runtime assertions without defensive checks are programmer errors that are not difficult to spot in code review and should not be expected to make it to deployed code.
// RED FLAG
x := y.(type)
// good
x, ok := y.(type)
if !ok { return an error }
Joking aside, you could clearly plot the probability of running into a runtime error by programming language.
Of course, a language with less runtime errors is a far cry from being a panacea. Avoiding runtime errors is not the same as avoiding all categories of bugs. And while I personally prefer stronger type systems - they definitely come with increasing levels of cognitive cost.
But I still feel that the type-safety vs. runtime trade-off is more often ignored, underestimated or undersold than it is being hyped. Yes, certain languages (cough Rust cough) are being hyped, but not the conscious choice of balancing programmer learning curve with runtime type-safety.
And while on the topic of Rust, it's probably not the best choice for a language that sees less runtime panics. Especially since unwrapping an error is always the easiest way to handle an error, and thus quite common.
But lazy error unwrapping aside, Rust does avoid null dereference exceptions, type casting exceptions and most types of race conditions that can be quite prevalent with go[1].
it's pretty obvious that it could influence new developers into the wrong direction though. Saying things like "ha, let's not bother checking this, at worst it'll just panic and i'll simply abort the request".
Which would definitely impact the quality of the software overall in a bad way.
I'd not be so sure. Accepting that everything that can fail will fail shaped me as a young developer, and "Exceptional C++" had a huge influence on me. Now my approach for new code I review is this:
* Make sure you support properly unrolling the stack
* Keep a clean failure boundary, probably somewhere on top of your loop
Right, all of these are good points, but the problem is that the "failure boundary" of a panic is the entire process. You can't constrain it, or assume that it's scoped to a single goroutine. Errors do not have this property.
> the "failure boundary" of a panic is the entire process.
This is trivially falsifiable by panicking yourself and immediately recovering. Neither failure domain nor failure boundary need to align with the entire process.
The impact of a specific panic does not extrapolate to the impact of all panics, and my claim is not falsified by such an example. Panics are defined by the language to represent unrecoverable errors.
Recovering from a panic thrown by Method invalidates the state of the Thing which threw that panic. If that Thing is shared among concurrent actors, the entire program state is invalidated.
No, the state represented by a connected graph of variables accessible by the concurrent actors is tainted. This is hardly "the entire program state". Often it's just a few cache entries.
Also, see my first, most important, bullet point:
"* Make sure you support properly unrolling the stack."
Which means a request to Thing errored out, but it never enters an invalid state. If you fail at that, all bets are off. But then you're dealing with a mediocre codebase anyway.
And finally, let me rewrite your example to something, that I see much more often in real life code, which problematic _even without concurrent actors_ because somethingThatAssumesTheAboveDidntReturnAnError might do horrible things all by themself:
I'm not sure how to respond to this. You seem to believe that panics express problems which are constrained to the call stack which instantiated the panic. This isn't true. But I'm not sure how to express this to you in a way that will convince you. So I guess we're at a stalemate.
> You seem to believe that panics express problems which are constrained to the call stack which instantiated the panic.
Not inherently, but it's your job as a developer to make sure this is the case, that's what:
"* Make sure you support properly unrolling the stack."
means.
In case you're dealing with unknown code it's your job to find out what the connected graph of potentially tainted objects is and discard them. That's what "keep a clean failure boundary" means.
If you can't, because you don't want to (short lived process, prototypes) or are unable to (hairy ball of code), tearing down the process is indeed the only option and a sane fallback choice made by the language designers. But it's not necessarily a hallmark of robust software.
I hope I had a final shot to clear up what I meant, thanks for the discussion anyway.
In Go when some code writes `panic` it is expressing an error condition which should not be intercepted by callers and is expected to terminate the process. A panic is not an error, and panics should not be recovered as if they were errors.
> I guess this is the crux of the issue. I don't think this is true, or needs to be true. It certainly hasn't been my experience. I think assuming panics are normal will take you down some paths that make it basically impossible to write reliable software. But, to each their own.
I'd rather have the control to log the panic on a service rather than it forcibly dying and taking down any other connections with it. Kube will just spin up a new one anyway, which just introduces a downtime gap that doesn't need to exist.
I don't think I'm effectively communicating the impact of handling a panic and continuing program execution. A panic that comes from a memory model violation (as one example) can change the value of anything in the memory space of the program. If the program continues, that change will go undetected, and can have results that make the program completely nondeterministic. This isn't a doom and gloom, sky-is-falling prognostication, it's literally what is defined by the spec and memory model of the language.
> A panic that comes from a memory model violation (as one example) can change the value of anything in the memory space of the program ... This isn't a doom and gloom, sky-is-falling prognostication, it's literally what is defined by the spec and memory model of the language.
I do not think you are correct. Go has a class of unrecoverable panics for this specific reason. Go also runs deferred functions after a recoverable panic, so the notion that it's unsafe to handle it, or continue executiona after doesn't hold at all - it is literally a first-class feature of the language.
I have not seen an instance of a recoverable panic that is raised _after_ such a fatal operation. If you have an example of such, I would love to see it.
There seems to not be any standard list of unrecoverable panics/aborts, but this Stackoverflow post [1] has a list of a few.
As far as the user/developers are concerned, it doesn't matter too much, since you have no option to recover them, but it would be nice if it was explained if defers are still ran. I'm assuming they are not.
If there is no way for callers to reliably distinguish recoverable panics from unrecoverable panics, then this distinction doesn't really exist, does it? Panics are panics.
I'm not sure what point you are trying to make anymore.
Of course you cannot distinguish between unrecoverable and recoverable panics, because by definition an unrecoverable panic is not recoverable. There is no caller to distinguish between it - it is killed.
Oh. You're using the word panic to describe a superset of actual panics and other even more serious errors. Those things you call unrecoverable panics are not actually panics.
The point I'm trying to make is that panics are not errors by another name, and they are not safe to recover from in general.
I would agree if it weren’t super easy to cause a panic in go.
Index slice out of bounds? panic. Close a channel twice? Panic. Incorrect type assertion? Panic. Dereference nil pointer? Panic.
I would argue that all of these examples which are the most common in my experience are “goroutine scoped” because the goroutine was aborted before they potentially modified the application state in an unknown way.
It’s like not in C, or C++ where out of bounds access has now put the entire application into an unknown state.
> Index slice out of bounds? panic. Close a channel twice? Panic. Incorrect type assertion? Panic. Dereference nil pointer? Panic.
These are all really bad things which should never survive to production code. It is not difficult to detect and prevent them.
> I would argue that all of these examples which are the most common in my experience are “goroutine scoped” because the goroutine was aborted before they potentially modified the application state in an unknown way.
What makes you think that terminating the goroutine that triggered these panics prevents them from impacting the process state?
> It’s like not in C, or C++ where out of bounds access has now put the entire application into an unknown state.
What makes you think this is the case? Panics have unknowable impact, and many panics (e.g. data races) absolutely do put the program into an unknown state.
>These are all really bad things which should never survive to production code. It is not difficult to detect and prevent them.
This is equivalent to saying "out of bounds memory writes are not difficult to detect and prevent in C code". Like actually equivalent (possibly worse), not just "well if you squint they look similar".
Of course it's not hard most of the time. Being perfect is beyond hard though. And if you're not perfect, you might open the door to anything in C, or cluster-destroying rolling crashes in Go.
Sometimes shutting down every piece of your software if that happens is the correct choice, and sometimes it's so far beyond reasonable that it's ludicrous to argue in favor of "every panic is an abort".
> Sometimes shutting down every piece of your software if that happens is the correct choice, and sometimes it's so far beyond reasonable that it's ludicrous to argue in favor of "every panic is an abort".
Very much this. And even for the same project: in some cases, I'm a fan of employing a quite strict error handling policy in dev environments (crash and burn) and using a more lenient approach in prod (elevated log level). In my experience, this can result in a robust product. Most importantly, this means the decision is not even made by the application programmer, sometimes it's a config thing.
x == y can panic if interface values contain incomparable fields in unexported nested structs, how would I check for that? Should we let it become a query of death and bet thousands of peers’ jobs on it never happening?
The comparison is explicitly allowed in the language spec, there’s no warning for doing it, and it often works depending on the types. It’s a data-dependent runtime error, which is usually hard to guarantee test coverage for.
more specifically, it's really strange to hear of people doing equivalence checks on objects with structure. What are you expecting that comparison to do? I doubt it is doing what you think is happening, and is indeed risky of panics.
> (e.g. data races) absolutely do put the program into an unknown state.
Data races do not necessarily result in panics.
Many (perhaps even most?) data races would not result in panic but just in garbled, missing, duplicate, out-of-order or otherwise incorrect data.
Data-race induced panics are generally the side-effect of a data race, not a direct protection against. They can often be inconsistent: e.g. a data race in a slice that contains a binary data format could garble a variable-length string prefix and produce an index-out-of-bounds panic. Or it could prematurely consume a shared pointer and overwrite it with nil, only to have the nil pointer dereferenced by another goroutine. These kind of panics are unpredictable.
If your application has shared global state (in-memory or even a database), it may become inconsistent due to data races. But whether data-race induced indicate irreversibly corrupted global state that requires (and can be fixed with) application restart - that is case-by-case thing.
Let's say your application has some shared state that got corrupted and the corruption triggered a panic down the line.
If your shared state is persisted in a database or some other distributed mechanism and that state got corrupted: restarting the application won't help you.
If your shared state is scoped at the HTTP request level (or whichever boundary you choose for suppressing your panics): you don't need to restart the application. The request is already terminated, along with its shared state.
Which leaves us with in-memory global state. This kind of state is generally minimized in the type of microservice and network infrastructure applications that Go is often used for.
A very small percentage of your panics will indicate corruption of such state. Will you be willing to risk service downtime in order to protect against the small possibility that the service has run into a state where its shared in-memory data became corrupted?
in memory or some other distributed mechanism and that state got corrupted: restarting the application won't help you.
I tend to agree with you that these are relatively easy things to detect. I see no reason for the downvotes.
Production systems should have relatively robust testing whose coverage can be increased over time. When something panics, the cause of the panic should be fixed so that the panic never happens again. Over time panics shouldn't be happening.
Then again the systems that I have relied on I have written on my own without other hands in the pot so maybe I just don't have to deal with the reality of other programmers phoning things in.
If your programming language handles very common errors by crashing the entire application, and if preventing these crashes is actively discouraged, then that suggests a flaw in the language itself.
This would be fine for a low-level language like C where you need to allow SEGFAULTs, but designing it into a high-level language makes no sense.
A lot of very common operations can panic: division, dereferencing a pointer, invoking an interface method, indexing/slicing an array/slice/string, asserting the type of an interface, and converting a slice to pointer to array. It’s possible to check, but I’ve never seen a tool that verifies you never use any of these without checking. You also have to check for nil channels, though they block forever (maybe consuming a goroutine) rather than panicking.
And there are some operations where you cannot check in advance whether a panic will happen: comparing interfaces (underlying values might not be fully comparable), indexing a map (could blow up during any concurrent write), sending to a channel (might be closed), and closing a channel.
You can recover from a panic though, so if you are implementing something that may panic you should have some sensible defer/recover in there if you can't afford to have your process crash.
Division by zero, dereferencing a nil pointer, invoking methods on a nil interface, invalid indexing of an array, unchecked type assertions -- these are not common operations! These are always easily detectable programmer errors.
