Open ExpandingMan opened 6 years ago
This is typically not an issue, but enough people complained that I caved and made the default false. I think we could take the same approach in Feather.
Actually mmapping also creates (more pernicious) problems on Linux, e.g. https://github.com/JuliaData/CSV.jl/issues/180.
In light of @nalimilan 's comment it definitely seems to me that it should be true
by default in Windows as there's really nothing wrong there: it's just the file system blocking you from deleting a file that is currently being referenced in memory. I don't have any stake in this issue since I never use Windows, but it appears from discourse that we do have a surprisingly large number of Windows users, and I thought it would be nice if they could benefit from the (rather significant) advantages of memory mapping by default.
The problem is that users are not necessarily aware that the file is open, and they can't delete it nor replace it, which is very common (at least for CSV) when you do some modifications that you want to save.
I was reading the CSV.jl documentation today and I noticed that they give a possible explanation for why
mmap=false
by default on Windows.It actually seems like a surprisingly reasonable behavior on the part of the windows file system (or is it? I'm assuming if you deleted the file before gc'ing in Linux you would not then get segfaults).
Anyway, since memory mapping is now a very important part of Feather.jl, is this really a good enough reason to disable memory mapping by default on Windows? Is there another issue that I'm missing?