Revert "Deprecate DevNull, STDIN, STDOUT, and STDERR to lowercase"

[ararslan]

This reverts PR #25959. Lowercasing these identifiers conflicts with the statistical meaning of stderr, i.e. the standard error. This is defined in StatsBase and widely used.

[iamed2]

Can we just not export them?

stderr search on Google waits until the 3rd page to show the statistical meaning. The next use was on page 7.

[StefanKarpinski]

IO.stderr, IO.devnull, etc. would be pretty good imo but IO is already a type so we either need a different name or some more name shuffling...

[KristofferC]

https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.stats.sem.html

[iamed2]

Python puts them in sys, maybe Base.Sys.stderr?

[StefanKarpinski]

Sure, that would be a good place for them.

[JeffBezanson]

These seem much more important than most of the Sys module --- you need the standard error stream far more often than you need to know e.g. which model of CPU you have.

[StefanKarpinski]

The idea behind putting them in Sys is not because they're less important but to avoid taking a name that is commonly used in statistics.

[JeffBezanson]

I really don't think they fit in that module though. Everything else in Sys is basically hardware and OS reflection, nothing like critical I/O functionality.

[iamed2]

Has anyone felt the desire for an IO-related module for other reasons? Maybe it would be useful for other things.

[mbauman]

julia> Base.getproperty(::Type{IO}, s::Symbol) = s == :stdout ? Base.stdout : s == :stderr ? Base.stderr : throw(ArgumentError())

julia> IO.stderr
Base.TTY(RawFD(0x0000000f) open, 0 bytes waiting)

julia> show(IO.stderr, "HAHAHA")
"HAHAHA"

[JeffBezanson]

😅

[nalimilan]

https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.stats.sem.html

Note that the SciPy stats.stderr function was deprecated in favor of stats.sem. sem is pretty short, but it doesn't seem terrible to do some sort of renaming in StatsBase. I suppose it makes the deprecation process more painful, however. (From #25959 (comment))

@KristofferC @stevengj That's a different thing. stderr in StatsBase returns the standard error of the model coefficients, not of the mean. So it's more general (and could be used in other contexts than modeling). A possible name would be se, but it's quite short. ste would be consistent with std, but it's less standard. EDIT: we could also rename std to sd (which is the name used e.g. by R) and stderr to se.

[KristofferC]

ste sounds good.

[stevengj]

The stdio streams are used a lot more than the standard error function in programming, so it doesn't seem crazy that the latter should be the one to be renamed. There are lots of reasonable options, e.g. you could just use stderror — lengthening it seems preferable to shortening it to an inscrutable abbreviation like se.

[ararslan]

Triage says rename the stats function.