Tying asserts to optimization is bad design. Asserts are for debugging; you write an assert when you are not 100% confident that it holds, or there is a risk of it not holding in the future even if it holds now.
The "assert" word is not a great choice, but it's entrenched that way. In a debate, you'd not want to assert something (present it as a fact) if you're not sure.
Optimization promises have the same form; they are also assertions, but with a different purpose. They can use the same syntax, but some different word.
The word assert is attractive for this purpose: I know for sure something is true, and I'm communicating it to the compiler; thus, I am asserting something.
The designer must not be thereby seduced into conflating debugging assertions with optimization promises.
I understand that in this case they have a safety switch for assertions, but conflating debug assertions with optimization promises by means of a switch is still a poor design.
You want that switch on a case-by-case basis for each individual assertion, and the easiest way for that is to just have a different identifier.
To add on to this, the C++ proposal for [[assume]] [0] came to a similar conclusion and cites MSVC's experience with an assert-in-debug-assume-in-release construction (i.e., removing the assumptions resulted in a 1-2% speedup and increased reliability) as a reason to not tie assertions to assumptions.
Asserts in my Rust code use a custom runtime macro "assert_eq_as_result" which does what the article is describing, by returning Rust Ok or Rust Error.
I recently worked for a company with a large C++ codebase. For many years they relied entirely on asserts in QA builds to catch errors, no unit testing or production asserts. When I joined about three years ago they had just started to pick up unit testing, and would be enabling asserts in production “soon” (once all the most common ones had been fixed).
I left a little over a month ago, and the status on that was still “soon”. Some bad practices early on are just really hard to dig yourself out of.
And I'm pretty happy with its design considering its age.
Notably this is not a function call and indeed things are not called unless you enable it. Contrast with Zig. So I guess you will only suffer from code bloat if you never enable them.
The tutorial mentions the dangers of side effects. But it also mentions how to use them for more complicated assertions. That's natural since you'll want something like that when you need to check a post-condition.
Programming assertions get joked on because of, ahem.
- Step one: Turn on extra checks in test environments where the stakes are low
- Step two: Turn them off in production (with realistic data because prod eq. reality) to save cycles
And that seems to be partially accurate. However the truly interesting assertions can test complex conditions that might break complexity (Big O) contracts. Like a private mininum function that is advertised as O(1) on account of taking a sorted list. But there is no type guarantee that the list is sorted. So you assert that it is sorted. But that breaks the complexity contract.
Overall I have not used assertions in Java for trivial conditions in like five years. They're better deployd for something more complex than that.
Then there is the whole thing about -- and more topics to be sure -- crashing the whole application or not. That's not necessarily great for us regular Java programmers. However we can (though discoraged) catch AssertionError if we want to.
"And I'm pretty happy with its design considering its age."
Java did the right thing for assertions but then completely failed for the analoguous issue when it comes to logging.
I admit that logging is more complex because you often want it configurable dynamically at runtime. But I'd argue that the language should not be in your way if all you need is a compile time decision and the contortions we made for logging to stay low cost when nothing is logged are crazy in Java.
> Imagine a big codebase with this somewhere in it:
fn processThing(thing: Thing) void {
// this function must always be invoked on
// a thing that has already been started
assert(thing.is_started);
// ...
}
I know you mentioned fuzzing earlier in the article but seriously, fuzzing deserves an extra mention after asking me to imagine that.
Tying asserts to optimization is bad design. Asserts are for debugging; you write an assert when you are not 100% confident that it holds, or there is a risk of it not holding in the future even if it holds now.
The "assert" word is not a great choice, but it's entrenched that way. In a debate, you'd not want to assert something (present it as a fact) if you're not sure.
Optimization promises have the same form; they are also assertions, but with a different purpose. They can use the same syntax, but some different word.
The word assert is attractive for this purpose: I know for sure something is true, and I'm communicating it to the compiler; thus, I am asserting something.
The designer must not be thereby seduced into conflating debugging assertions with optimization promises.
I understand that in this case they have a safety switch for assertions, but conflating debug assertions with optimization promises by means of a switch is still a poor design.
You want that switch on a case-by-case basis for each individual assertion, and the easiest way for that is to just have a different identifier.
To add on to this, the C++ proposal for [[assume]] [0] came to a similar conclusion and cites MSVC's experience with an assert-in-debug-assume-in-release construction (i.e., removing the assumptions resulted in a 1-2% speedup and increased reliability) as a reason to not tie assertions to assumptions.
[0]: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p17...
[1]: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p20...
Asserts in my Rust code use a custom runtime macro "assert_eq_as_result" which does what the article is describing, by returning Rust Ok or Rust Error.
The Rust crate: https://crates.io/crates/assertables
All the macros have forms for different outcomes:
If anyone here wants to help me port it to Zig, I'm happy to do it.I recently worked for a company with a large C++ codebase. For many years they relied entirely on asserts in QA builds to catch errors, no unit testing or production asserts. When I joined about three years ago they had just started to pick up unit testing, and would be enabling asserts in production “soon” (once all the most common ones had been fixed).
I left a little over a month ago, and the status on that was still “soon”. Some bad practices early on are just really hard to dig yourself out of.
I believe this is just about the behavior of std.debug.assert.
You can pretty easily have a different kind of assert that disappears in release builds (if you want).
I really like the Oracle tutorial on Java asserts.
https://docs.oracle.com/javase/8/docs/technotes/guides/langu...
And I'm pretty happy with its design considering its age.
Notably this is not a function call and indeed things are not called unless you enable it. Contrast with Zig. So I guess you will only suffer from code bloat if you never enable them.
The tutorial mentions the dangers of side effects. But it also mentions how to use them for more complicated assertions. That's natural since you'll want something like that when you need to check a post-condition.
Programming assertions get joked on because of, ahem.
- Step one: Turn on extra checks in test environments where the stakes are low
- Step two: Turn them off in production (with realistic data because prod eq. reality) to save cycles
And that seems to be partially accurate. However the truly interesting assertions can test complex conditions that might break complexity (Big O) contracts. Like a private mininum function that is advertised as O(1) on account of taking a sorted list. But there is no type guarantee that the list is sorted. So you assert that it is sorted. But that breaks the complexity contract.
Overall I have not used assertions in Java for trivial conditions in like five years. They're better deployd for something more complex than that.
Then there is the whole thing about -- and more topics to be sure -- crashing the whole application or not. That's not necessarily great for us regular Java programmers. However we can (though discoraged) catch AssertionError if we want to.
"And I'm pretty happy with its design considering its age."
Java did the right thing for assertions but then completely failed for the analoguous issue when it comes to logging.
I admit that logging is more complex because you often want it configurable dynamically at runtime. But I'd argue that the language should not be in your way if all you need is a compile time decision and the contortions we made for logging to stay low cost when nothing is logged are crazy in Java.
> Imagine a big codebase with this somewhere in it:
I know you mentioned fuzzing earlier in the article but seriously, fuzzing deserves an extra mention after asking me to imagine that.Asserts in any language must not be added willy-nilly but following a specific methodology viz. Design By Contract (DbC) - https://en.wikipedia.org/wiki/Design_by_contract
DbC was first introduced in Eiffel but the ideas can be used in any language. See the following;
1) Design by Contract and Assertions - https://www.eiffel.org/doc/solutions/Design_by_Contract_and_...
2) Applying "Design by Contract" (pdf) - https://se.inf.ethz.ch/~meyer/publications/computer/contract...
3) Also see the book; Persuasive Programming by Jerud Mead and Anil Shinde which uses asserts systematically to write the proof along with code.
Finally, DbC using asserts is now even more important with AI generated code since it allows one to map the specification to generated code.