# Why doesn’t C++ have networking support?

Recently, in a discussion of the Networking TS on one of WG21’s internal “reflector” mailing lists, Bjarne Stroustrup wrote:

One of the most frequent questions that I get from C++ programmers and others alike is “why doesn’t C++ have networking support?” It is not a question that is easy to answer effectively.

Billy O’Neal wrote a very long and detailed answer to the stated question. (This was not a reply targeted at Bjarne specifically; this was a detailed answer suitable for people who ask questions like “why doesn’t C++ have networking support?”) I think Billy’s answer is great and deserves wider viewership, so I asked Billy and Bjarne for permission to share it here, and they said “Feel free.” So: Billy, why doesn’t C++ have networking support in the standard library?

(1) We don’t know how to fix mistakes we put into std::. The existing containers/algorithms library, when it was ‘fixed in the stone of the standard’, was standardizing behavior we as an industry knew how to do for decades before that shipped. Sure, different languages and systems have different interfaces to get at those things, but linked lists, dynamic arrays, and quicksort were comparatively solved problems. We do not have that equivalent for concurrency and parallelism, and networking is an inherently concurrent environment.

(1a) This is extreme hearsay, take with mountains of salt: I have been told that even the ‘BSD Sockets’ bits exposed by the TS are problematic in a lot of environments. For example, I have been told that you can’t write something like an HTTP client that is well behaved on iOS on top of such an interface, because as soon as the user switches to another app, any sockets you have open are closed. But if you use the platform HTTP library, the platform knows what HTTP is and can do your download or similar in the background while your app is suspended. (Apple folks, please correct me if I misunderstood this one.)

(2) We have implementation experience attempting to standardize and ABI-stabilize complicated machinery for which we as an industry don’t have a complete, accepted solution, in the form of std::regex. The result is that I can’t reasonably recommend anyone use std::regex, because the 3 major standard library implementations of that are atrocious (as in 2300% to 83000% slower than a quality implementation for some inputs). Examples:

(3) Different concurrency and parallelism models don’t usually compose very well. For example, a program that wants to use ASIO and libuv needs to burn a separate thread for each and pay at least one context switch whenever crossing “universes.” This means the traditional response “don’t like std::vector? write your own” is effectively unavailable here, and anyone who has a large body of code that doesn’t map into whatever ends up in std:: needs to rewrite all of their code if the std:: model gains wide adoption. This makes the (2) problem much worse, since even making up another std2 name might not let us fix such mistakes.

(4) We don’t have any entity that can behave as the ‘benevolent dictator’ and encode any particular model / solution to these problems and force everyone else to comply. C++ effectively has several de-facto networking models, and choosing one means we are making some of our customers unhappy.

(5) The advantage of putting things into std:: is that they are available everywhere; a corollary to that is anything we put into std:: must be able to go anywhere. A lot of environments that have good networking support also reduce the complexity quite a bit by refusing to cater to some use cases. For example, Node.JS is effectively a single threaded environment that doesn’t care about being a ‘guest in someone else’s process’ or otherwise being unloadable, but things we put in std:: might be used by a print driver or shell extension that must be. It seems unlikely that such simplifying assumptions would cater to the full set of customers we have.

(6) C++’s lack of memory safety makes it an expensive language to use for anything that crosses trust boundaries, and networking has that all over the place. If it’s impossible to use the standard networking interface to implement anything reasonable to put in production (which in the modern era means at least something like TLS) this would further cement the ‘C++ doesn’t care about security’ stigma.

My being weakly against standardizing networking doesn’t come from thinking networking is an unimportant problem. It comes from being unable to reconcile the “ABI stabilized forever, can’t fix mistakes ever” world we’ve created for ourselves with this relatively rapidly changing problem domain.

(Also please note that the above ‘weakly against’ is my personal position, not that of Microsoft.)

Disclaimer: I don’t fully understand point (6). As best I understand it (with some further help from Billy off-list), it’s basically the point that if programmers come to believe that “C/C++ networking” is equivalent to “Heartbleed vulnerabilities,” then they might reasonably avoid C++ for networking — use a “safer” language for all their networking needs — at which point it doesn’t matter whether C++ ever provides standard networking, because nobody would trust such a standard networking library in production.

Other than perhaps point (6), I think I agree wholeheartedly with everything Billy said.

Posted 2019-10-09