[Python-Dev] PEP 471 -- os.scandir() function -- a better and faster directory iterator

[Ben Hoyt]

Re is_dir etc being properties rather than methods:

>> I find this behaviour a bit misleading: using methods and have them
>> return cached results. How much (implementation and/or performance
>> and/or memory) overhead would incur by using property-like access here?
>> I think this would underline the static nature of the data.
>>
>> This would break the semantics with respect to pathlib, but they're only
>> marginally equal anyways -- and as far as I understand it, pathlib won't
>> cache, so I think this has a fair point here.
>
> Indeed - using properties rather than methods may help emphasise the
> deliberate *difference* from pathlib in this case (i.e. value when the
> result was retrieved from the OS, rather than the value right now). The main
> benefit is that switching from using the DirEntry object to a pathlib Path
> will require touching all the places where the performance characteristics
> switch from "memory access" to "system call". This benefit is also the main
> downside, so I'd actually be OK with either decision on this one.

The problem with this is that properties "look free", they look just
like attribute access, so you wouldn't normally handle exceptions when
accessing them. But .lstat() and .is_dir() etc may do an OS call, so
if you're needing to be careful with error handling, you may want to
handle errors on them. Hence I think it's best practice to make them
functions().

Some of us discussed this on python-dev or python-ideas a while back,
and I think there was general agreement with what I've stated above
and therefore they should be methods. But I'll dig up the links and
add to a Rejected ideas section.

> * +1 on a new section in the PEP covering rejected design options (calling
> it iterdir, returning a 2-tuple instead of a dedicated DirEntry type)

Great idea. I'll add a bunch of stuff, including the above, to a new
section, Rejected Design Options.

> * regarding "why not a 2-tuple", we know from experience that operating
> systems evolve and we end up wanting to add additional info to this kind of
> API. A dedicated DirEntry type lets us adjust the information returned over
> time, without breaking backwards compatibility and without resorting to ugly
> hacks like those in some of the time and stat APIs (or even our own codec
> info APIs)

Fully agreed.

> * it would be nice to see some relative performance numbers for NFS and CIFS
> network shares - the additional network round trips can make excessive stat
> calls absolutely brutal from a speed perspective when using a network drive
> (that's why the stat caching added to the import system in 3.3 dramatically
> sped up the case of having network drives on sys.path, and why I thought AJ
> had a point when he was complaining about the fact we didn't expose the
> dirent data from os.listdir)

Don't know if you saw, but there are actually some benchmarks,
including one over NFS, on the scandir GitHub page:

https://github.com/benhoyt/scandir#benchmarks

os.walk() was 23 times faster with scandir() than the current
listdir() + stat() implementation on the Windows NFS file system I
tried. Pretty good speedup!

-Ben