append mode in open() is not posix

[vtjnash]

f = open("filename.txt", "a+")
println(f, "super\ncali\nfragilistic")
close(f)
f = open("filename.txt", "a+")
a = readall(f)
assert(length(a) > 0) # bug 1 -- not posix compliant
seek(a,0)
a = readall(f)
println(f,"hi") # bug 2 -- file is now twice as long as it should be
flush(f)
close(f)

[JeffBezanson]

I say we just get rid of "a" and "a+" since they don't do the same thing as fopen or posix open. @StefanKarpinski

[StefanKarpinski]

I feel like there's a joke somewhere in here. How about we just fix "a" and "a+" to do something useful?

[JeffBezanson]

Well, right now they just mean to start with the file position at the end. The behavior of "a+" in fopen is to read from the beginning but write at the end. We don't have that implemented, so we shouldn't pretend to support "a+". I was wrong at first, "a" is fine and only "a+" is problematic.

[StefanKarpinski]

The "a+" mode of writing at the end and reading from the beginning seems strange. Do files support separate read and write cursors?

[JeffBezanson]

No, it's just kind of a magic thing, and I wasn't really aware of it, which is why it wasn't there.

[JeffBezanson]

The other purpose of "a" or "a+" mode is to avoid the race condition that occurs if you separately seek to the end of the file before writing. Append mode has the kernel do it for you atomically.

[rapus95]

I ran into this problem now aswell. Maybe one could copy a corrected version of the one at fopen into the API instead of linking to an incorrect description of the arguments in order to prevent others running into this case in the future. :D

[StefanKarpinski]

Julia does not support separate read and write cursors in files, nor do I think we should. We could, however, delete the non-compliant "a+" mode. "a" is fine since it is write-only. We would still have the option to do open("file.txt", append=true, read=true) to get what "a+" currently does.

[rapus95]

right now opening a file like that: open("tmp.txt", true, false, false, false, true) will result in an error because there is no write permission... what about making write=false, append=true always trying to place the cursor at the end? (alias no writing at current cursor position)
EDIT: I'm on windows
EDIT2: In that case it might be clever to make "write" and "append" some sort of mutually exclusive in order to prevent misleading interpretation since if you can write you can do it wherever you want even at the end. But if you append you may only write to the end.

[brenhinkeller]

From a few quick tests, it looks like the current behavior now matches fopen

For reference, fopen:

a+: append/update: Open a file for update (both for input and output) with all output operations writing data at the end of the file. Repositioning operations (fseek, fsetpos, rewind) affects the next input operations, but output operations move the position back to the end of file. The file is created if it does not exist.