On Sunday, 7 February 2016 at 10:50:24 UTC, Johannes Pfau wrote:
I saw this on code.dlang.org some time ago and had a quick
look. First of all this would have to go into phobos to make
sure it's used as some kind of a standard. Conflicting stream
libraries would only cause more trouble.
Then if you want to go for phobos inclusion I'd recommend
looking at
other stream implementations and learning from their mistakes
;-)
There's
https://github.com/schveiguy/phobos/tree/babe9fe338f03cafc0fb50fc0d37ea96505da3e3/std/io
which was supposed to be a stream replacement for phobos. Then
there
are also vibe.d streams*.
I saw Steven's stream implementation quite some time ago and I
had a look at vibe's stream implementation just now. I think it
is a mistake to use classes over structs for this sort of thing.
I briefly tried implementing it with classes, but ran into
problems. The non-deterministic destruction of classes is
probably the biggest issue. One has to be careful about calling
f.close() in order to avoid accumulating too many open file
descriptors in programs that open a lot of files. Reference
counting takes care of this problem nicely and has less overhead.
This is one area where classes relying on the GC is not ideal.
Rust's ownership system solves this problem quite well. Python
also solves this with "with" statements.
Your Stream interfaces looks like standard stream
implementations (which
is a good thing) which also work for unbuffered streams. I
think it's a
good idea to support partial reads and writes. For an
explanation why
partial reads, see the vibe.d rant below. Partial writes are
useful
as a write syscall can be interrupted by posix signals to stop
the
write. I'm not sure if the API should expose this feature (e.g.
by
returning a partial write on EINTR) but it can sometimes be
useful.
I don't want to assume what the user wants to do in the event of
an EINTR unless a certain behavior is desired 100% of the time. I
don't think that is the case here. Thus, that is probably
something the user should handle manually, if needed.
Still readExactly / writeAll helpers functions are useful. I
would try
to implement these as UFCS functions instead of as a struct
wrapper.
I agree. I went ahead and made that change.
For some streams you'll need a TimeoutException. An interesting
question is whether users should be able to recover from
TimeoutExceptions. This essentially means if a read/write
function
internally calls read/write posix calls more than once and only
the
last one timed out, we already processed some data and it's not
possible to recover from a TimeoutException if the amount of
already
processed data is unknown.
The simplest solution is using only one syscall internally. Then
TimeoutException => no data was processed. But this doesn't
work for
read/writeExcatly (Another reason why read/writeExactly
shouldn't be
the default. vibe.d...)
In the current implementation of readExactly/writeExactly, one
cannot assume how much was read or written in the event of an
exception anyway. The only way around this I can see is to return
the number of bytes read/written in the exception itself. In
fact, that might solve the TimeoutException problem, too. Hmm...
I'd like to keep the fundamental read/write functions at just one
system call each in order to guarantee that they are atomic in
relation to each other.
Regarding buffers / sliding windows I'd have a look at
https://github.com/schveiguy/phobos/blob/babe9fe338f03cafc0fb50fc0d37ea96505da3e3/std/io/buffer.d
Another design question is whether there should be an interface
for such buffered streams or whether it's OK to have only
unbuffered streams + one buffer struct / class. Basically the
question is whether there might be streams that can offer a
buffer interface but can't use the standard implementation.
I think it's OK to re-implement buffering for different types of
streams where it is more efficient to do so. For example, there
is no need to implement buffering for an in-memory stream
because, by definition, it is already buffered.
I'm not sure if having multiple buffering strategies would be
useful. Right now, there is only the fixed-sized sliding window.
If multiple buffering strategies are useful, then it makes sense
to have all streams unbuffered by default and have separate
buffering implementations.
There is an interesting buffering approach here that is mainly
geared towards parsing:
https://github.com/DmitryOlshansky/datapicked/blob/master/dpick/buffer/buffer.d
* vibe.d stream rant ahead:
vibe.d streams get some things right and some things very
wrong. For
example their leastSize/empty/read combo means you might
actually
have to implement reading data in any of these functions. Users
have to
handle timeouts or other errors for any of these as well.
Then the API requires a buffered stream, it simply won't work
for
unbuffered IO (leastSize, empty). And the fact that read reads
exactly
n bytes makes stream implementations more complicated
(re-reading until
enough data has been read should be done by a generic function,
not
reimplemented in every stream). It even makes some user code
more
complicated: I've implemented a serial port library for vibe-d.
If I don't know how many bytes will arrive with the next
packet, the
read posix function usually returns the expected/available
amount of
data. But now vibe.d requires me to specify a fixed length when
calling
the stream read method. This leads to ugly code using peak...
Then vibe.d also mixes the sliding window / buffer concept into
the stream class, but does so in a bad way. A sliding window
should expose the internal buffer so that it's possible to
consume bytes from the buffer, skip bytes, refill... In vibe.d
you can peak at the buffer. But you can't discard data. You'll
have to call read instead which copies from the internal buffer
to an external buffer, even if you only want to skip data. Even
worse, your external buffer size is limited. So you have to
implement some loop logic if you want to skip more data than
fits your buffer. And all you need is a discard(size_t n)
function which does _buffer = _buffer[n .. $] in the stream
class...
These are the golden nuggets of experience I was looking for when
making this post. They definitely help to guide an ergonomic API
design. Standing on the shoulders of giants and such. Thanks!
TLDR: API design is very important.
Completely agree.