On 18 Apr 2016 at 17:35, Klemens Morgenstern wrote:
At the current state, the tests pass on linux as well as windows (gcc-5 & MSVC-14). Requirements are C++, boost.fusion, boost.asio, boost.iostreams, boost.filesystem and boost.system.
I really could use some feedback and hope you're interested.
Firstly, well done for working on this. Last contract I got annoyed with Boost.Process and ended up replacing it with a hacked shoe in based on ASIO. More recently, I needed a child process spawner for my ACCU presentation I'm giving this week and ended up again reinventing Boost.Process, only this edition is a "bare metal" reinvention based on Outcomes [1]. So I'm all for this in principle. However I'm not sure if I'm for your specific formulation. Here are my top issues: 1. You should rip out all usage of Boost.Iostreams. It's been without maintainer for years now, and has a fundamentally broken design for async i/o. Nobody should be encouraged to use it in new code. 2. You should completely eliminate all synchronous i/o as that is also fundamentally broken for speaking to child processes. Everything needs to be async 100% of the time, it's the only sane design choice [2]. You can absolutely present publicly a device which appears to quack and waddle like a synchronous pipe for compatibility purposes, indeed AFIO v2 presents asynchronous i/o devices as synchronous ones if you use the synchronous APIs even though underneath the i/o service's run() loop gets pumped during i/o blocks. But underneath it needs to be 100% async, and therefore probably ASIO. 3. Instead of inventing your own i/o objects, I think you need to provide: (a) Async and sync objects extending ASIO's base objects with child_stdin, child_stdout and child_stderr - or whatever your preferred naming. (b) std::istream and std::ostream wrappers. These are not hard, ASIO helps you a lot with these once you have the native ASIO objects. 4. Child processes are not like threads and should not be represented as a first order object. They should instead be an opaque object represented by an abstract base class publicly and managed by a RAII managing class from whom the opaque object can be detached, assigned, transferred etc. 5. Replace all the on_exit() machinery with future continuations i.e. launching a process *always* returns a future. If someone wants to hook code onto when the process exits, a future continuation is the right tool. Similarly for fetching return codes, or detaching oneself from the child. Python's new subprocess.run() returns exactly the struct your future also needs to return. 6. Looking through your source code, I see references to boost::fusion and lots of other stuff. Great, but most people wanting a Process management library don't want to drag in a copy of Boost to get one. It's easier to just roll your own. So drop the Boost dependency. 7. Looking through your source code, I am struck about how much functionality is done elsewhere by other libraries, especially ASIO. I think less is more for Boost.Process, I always personally greatly preferred the pre-peer-review Boost.Process even with its warts over the post-peer-review one which had become too "flowery" and "ornate" if that makes sense. The latter became unintuitive to program against, I kept having to look up the documentation and that annoys me. This stuff should be blindingly obvious to use. It should "just work". I conclude my mini-review by suggesting "less is more" for Boost.Process. 99% of users want the absolute *minimum* featureset. Look at Python 3.5's new subprocess module, that is a very good API design and featureset to follow. It's intuitive, it gets the job done quickly, but it exposes enough depth if you really need it to write a really custom solution. I'd *strongly* recommend you copy that API design for Boost.Python and dispense with the current API design entirely. The absolute clincher in Python's subprocess is you can never, ever race nor deadlock stdout and stderr. That makes an underlying async i/o implementation unavoidable. I'd personally suggest save yourself a ton of hassle and use ASIO's pipe/unix socket support facilities, it's becoming the Networking TS anyway. Hope this is helpful. Niall [1]: https://github.com/ned14/boost.afio/blob/master/include/boost/afio/v2/ detail/child_process.hpp [2]: I refer to the stdout/stderr deadlock problem which is the biggest reason anyone reaches for a process management library instead of just using the syscalls directy. The internals of the child i/o needs to be 100% async to prevent deadlocking. You can absolutely present publicly a device which appears to quack and waddle like a synchronous pipe for compatibility purposes, indeed AFIO v2 presents asynchronous i/o devices as synchronous ones if you use the synchronous APIs even though underneath the i/o service's run() loop gets pumped during i/o blocks. -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/