Under Windows, this cannot happen. The rename will fail if the target file exists.
C:\Projects_Perl\testing>rename newfile.txt originalfile.txt
A duplicate file name exists, or the file cannot be found.

MOVE /Y newfile.txt originalfile.txt
[download]

perl -e "rename('newfile.txt', 'originalfile.txt')"
[download]

Thanks for the info!

Its been more than a decade since I worked on a serious Windows installation program for a huge 24/7 application. I am sure that my recollections are fuzzy from Perl 5.6, Win NT days! I remember spending a lot of time dealing with some of the more unusual cases, like what happens if we have a new fileA, but currently running processes are linked to an old .dll version or have open the old version of fileA or are using the info from the old version even though they have closed their filehandle to the old file version. That kind of system coordination stuff can be a mess and beyond the scope in this thread.

The general advice in this thread:

1. Get the new copy "ready to go" on the HD.
2. Make a backup of the original file and have a plan to revert to that if something fails.
3. Do the file replacement (newFile replaces oldFile) in the fewest, fastest steps possible.
Are fundamentally sound, excellent practical advice.

From the perl 5.24.0 C sources, file win32.c:

DllExport int
win32_rename(const char *oname, const char *newname)
{
    char szOldName[MAX_PATH+1];
    BOOL bResult;
    DWORD dwFlags = MOVEFILE_COPY_ALLOWED;
    dTHX;

    if (stricmp(newname, oname))
        dwFlags |= MOVEFILE_REPLACE_EXISTING;
    strcpy(szOldName, PerlDir_mapA(oname));

    bResult = MoveFileExA(szOldName,PerlDir_mapA(newname), dwFlags);
    if (!bResult) {
        DWORD err = GetLastError();
        switch (err) {
        case ERROR_BAD_NET_NAME:
        case ERROR_BAD_NETPATH:
        case ERROR_BAD_PATHNAME:
        case ERROR_FILE_NOT_FOUND:
        case ERROR_FILENAME_EXCED_RANGE:
        case ERROR_INVALID_DRIVE:
        case ERROR_NO_MORE_FILES:
        case ERROR_PATH_NOT_FOUND:
            errno = ENOENT;
            break;
        case ERROR_DISK_FULL:
            errno = ENOSPC;
            break;
        case ERROR_NOT_ENOUGH_QUOTA:
            errno = EDQUOT;
            break;
        default:
            errno = EACCES;
            break;
        }
        return -1;
    }
    return 0;
}
[download]

we see that the Perl rename function on Windows is implemented using the Win32 MoveFileEx function.

Here is a standalone C program using the Win32 MoveFileEx function:

// movetest.cpp
// compile with: CL /W3 /MD movetest.cpp
// example run: movetest fromfile tofile

#include <windows.h>
#include <stdio.h>

int main(int argc, char* argv[])
{
   if (argc != 3)
   {
      fprintf(stderr, "usage: movetest file1 file2\n");
      return 1;
   }
   char* from = argv[1];
   char* to   = argv[2];
   fprintf(stderr, "move '%s' to '%s'...", from, to);
   if (!MoveFileEx(from, to, MOVEFILE_REPLACE_EXISTING))
   {
      fprintf(stderr, "failed: error=%lu\n", (unsigned long)GetLastErr
+or());
      return 2;
   }
   fprintf(stderr, "done.\n");
   return 0;
}
[download]

After compiling, you can run with:

movetest newfile.txt originalfile.txt
[download]

which overwrites the original file even if it already exists (if you have permission to do so).

Well, it's pleasing to see that Perl is doing the rename via a single Win32 MoveFileEx function call in preference to a DeleteFile followed by a MoveFile (which would have no chance of being atomic). However, the jury seems to be out on how truly atomic Windows rename is. After reading this stackoverflow question I'm still confused. Alternatives to MoveFileEx appear to be ReplaceFile and MoveFileTransacted.

Via brute force search of the perl 5.24.0 source code, I further noticed that MoveFileEx is also available to Perl code via the Win32API::File module. In the latest Perl distribution, Win32API::File::MoveFileEx is called by ExtUtils::Install.

Most modern filesystems, like ext3, ext4, NTFS, and HFS+ support journaling, so "The vast majority of folks here" with their "standard versions of Windows or Unix variants" are already using them. Do you still think "There is no truly atomic operation on the file system"?

"Atomic rename()" by POSIX specifies a single rename operation.

Show some working code that demo's replacing a file using Perl that uses a single rename.

echo foo>foo; echo bar>bar; perl -e 'rename("foo","bar")' works perfectly fine on Linux. But I don't get why you wanted me to test it, when I know what I'm talking about and I knew it would work ahead of time (but I did test again anyway just to make sure I don't look like a fool), whereas if you had read up on the subject and tested before posting you might have learned something that it seems you didn't know before.

What? Works for me (it's just not atomic, as far as I can tell).

echo foo >foo
echo bar >bar
perl -e "rename(foo,bar)"
[download]