The new staged package installation shells out to the 'file' utility on Unix systems: https://github.com/wch/r-source/blob/31ee14c620eb1b939acd322f3b5617f998aab8e8/src/library/tools/R/install.R#L578 Although 'file' is usually present on most Unix systems, it might be missing from small Docker containers, where the aim is to make the container as small possible. The magic file of 'file' is about 5MB, so that is significant in this case. R uses 'file' to decide if a .so, .dll, etc. file is indeed a shared library, and (as I understand) if it is, it then goes on to try to fix the hardcoded installation path in it, using an os specific tool. As the second part needs to handle errors anyway, I wonder if it would make sense to skip the 'file' call completely, after all it is quite unlikely that a .dll or .so, etc. file is _not_ a shared library, and even if it is not, the errors will be caught later. Thanks, Gabor
package installation needs the file utility on Unix
3 messages · Tomas Kalibera, Gábor Csárdi
Well, this only applies to source installs of packages that have some files with the special extension, so on systems where a compiler toolchain needs to be installed, so the image cannot be really tiny, anyway. But ok, I've made stage install use "file" only when it is available. When it isn't and some file with extension sl, so, dylib or dll in the package installation is not in fact a shared object, one will get a number of error messages from the respective tools, but the installation passes. Currently on Linux this is the case of RcppParallel and FastRWeb. Tomas
On 3/7/19 11:57 PM, G?bor Cs?rdi wrote:
The new staged package installation shells out to the 'file' utility on Unix systems: https://github.com/wch/r-source/blob/31ee14c620eb1b939acd322f3b5617f998aab8e8/src/library/tools/R/install.R#L578 Although 'file' is usually present on most Unix systems, it might be missing from small Docker containers, where the aim is to make the container as small possible. The magic file of 'file' is about 5MB, so that is significant in this case. R uses 'file' to decide if a .so, .dll, etc. file is indeed a shared library, and (as I understand) if it is, it then goes on to try to fix the hardcoded installation path in it, using an os specific tool. As the second part needs to handle errors anyway, I wonder if it would make sense to skip the 'file' call completely, after all it is quite unlikely that a .dll or .so, etc. file is _not_ a shared library, and even if it is not, the errors will be caught later. Thanks, Gabor
______________________________________________ R-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Thanks! Good point about the compiler toolchain.....on Alpine Linux, which tends to be small, gcc can be less than 100MB, so still might matter a bit.... Gabor
On Fri, Mar 8, 2019 at 1:44 PM Tomas Kalibera <tomas.kalibera at gmail.com> wrote:
Well, this only applies to source installs of packages that have some files with the special extension, so on systems where a compiler toolchain needs to be installed, so the image cannot be really tiny, anyway. But ok, I've made stage install use "file" only when it is available. When it isn't and some file with extension sl, so, dylib or dll in the package installation is not in fact a shared object, one will get a number of error messages from the respective tools, but the installation passes. Currently on Linux this is the case of RcppParallel and FastRWeb. Tomas On 3/7/19 11:57 PM, G?bor Cs?rdi wrote:
The new staged package installation shells out to the 'file' utility on Unix systems: https://github.com/wch/r-source/blob/31ee14c620eb1b939acd322f3b5617f998aab8e8/src/library/tools/R/install.R#L578 Although 'file' is usually present on most Unix systems, it might be missing from small Docker containers, where the aim is to make the container as small possible. The magic file of 'file' is about 5MB, so that is significant in this case. R uses 'file' to decide if a .so, .dll, etc. file is indeed a shared library, and (as I understand) if it is, it then goes on to try to fix the hardcoded installation path in it, using an os specific tool. As the second part needs to handle errors anyway, I wonder if it would make sense to skip the 'file' call completely, after all it is quite unlikely that a .dll or .so, etc. file is _not_ a shared library, and even if it is not, the errors will be caught later. Thanks, Gabor
______________________________________________ R-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel