Autostep weird behavior

[fihuer]

Hi,
I've just discovered a weird behavior in the autostep heuristics:

In python:

>>> l=['s20pmc0', 's20pmc1', 's20pmc2', 's22pmc0', 's22pmc1', 's22pmc2', 's24pmc0', 's24pmc1', 's24pmc2', 's26pmc0', 's26pmc1', 's26pmc2', 's28pmc0', 's28pmc1', 's28pmc2', 's30pmc0', 's30pmc1', 's30pmc2', 's32pmc0', 's32pmc1', 's32pmc2', 's34pmc0', 's34pmc1', 's34pmc2', 's36pmc0', 's36pmc1', 's36pmc2']
>>> print NodeSet.fromlist(l, autostep=3)
s[20,22,24,26,28,30,32,34,36]pmc[0-2]
>>> l=['s20pmc0', 's20pmc1', 's22pmc0', 's22pmc1', 's24pmc0', 's24pmc1', 's26pmc0', 's26pmc1', 's28pmc0', 's28pmc1', 's30pmc0', 's30pmc1', 's32pmc0', 's32pmc1', 's34pmc0', 's34pmc1', 's36pmc0', 's36pmc1']
>>> print NodeSet.fromlist(l, autostep=3)
s[20-36/2]pmc[0-1]

When using the nodeset command:

$ nodeset --autostep=3 -f s[20,22,24,26,28,30,32,34,36]pmc[0-2]
s[20-36/2]pmc[0-2]
$ nodeset --autostep=3 -f s20pmc0 s20pmc1 s20pmc2 s22pmc0 s22pmc1 s22pmc2 s24pmc0 s24pmc1 s24pmc2 s26pmc0 s26pmc1 s26pmc2 s28pmc0 s28pmc1 s28pmc2 s30pmc0 s30pmc1 s30pmc2 s32pmc0 s32pmc1 s32pmc2 s34pmc0 s34pmc1 s34pmc2 s36pmc0 s36pmc1 s36pmc2
s[20,22,24,26,28,30,32,34,36]pmc[0-2]
$ nodeset --autostep=3 -f s[20,22,24,26,28,30,32,34,36]pmc[0-1]
s[20-36/2]pmc[0-1]
$ nodeset --autostep=3 -f s20pmc0 s20pmc1 s22pmc0 s22pmc1  s24pmc0 s24pmc1  s26pmc0 s26pmc1  s28pmc0 s28pmc1  s30pmc0 s30pmc1  s32pmc0 s32pmc1  s34pmc0 s34pmc1  s36pmc0 s36pmc1
s[20-36/2]pmc[0-1]

This does not looks like a consistent behavior, is it expected ?

Cheers,

Romain Fihue

[martinetd]

Fun, it's actually more interesting that the different width, just having one less "row" in the first example makes it work e.g.

$ nodeset --autostep=3 -f s20pmc0 s20pmc1 s20pmc2 s22pmc0 s22pmc1 s22pmc2 s24pmc0 s24pmc1 s24pmc2 s26pmc0 s26pmc1 s26pmc2 s28pmc0 s28pmc1 s28pmc2 s30pmc0 s30pmc1 s30pmc2 s32pmc0 s32pmc1 s32pmc2 s34pmc0 s34pmc1 s34pmc2 s36pmc0 s36pmc1 s36pmc2
s[20,22,24,26,28,30,32,34,36]pmc[0-2]
$ nodeset --autostep=3 -f s20pmc0 s20pmc1 s20pmc2 s22pmc0 s22pmc1 s22pmc2 s24pmc0 s24pmc1 s24pmc2 s26pmc0 s26pmc1 s26pmc2 s28pmc0 s28pmc1 s28pmc2 s30pmc0 s30pmc1 s30pmc2 s32pmc0 s32pmc1 s32pmc2 s34pmc0 s34pmc1 s34pmc2 
s[20-34/2]pmc[0-2]

What's even more fun is when I try with RangeSetND from python, I get the same kind of behaviour (changes depending on list length), but with one less element... (did I ever say I hate this double-list comprehension syntax? I do.)

from ClusterShell.RangeSet import RangeSetND
> l = [i for x in range(20, 34, 2) for i in [(x,0), (x,1), (x,2)]]
> r = RangeSetND(l, autostep=3)
> print(r)
20-32/2; 0-2

> l = [i for x in range(20, 36, 2) for i in [(x,0), (x,1), (x,2)]]
> r = RangeSetND(l, autostep=3)
> print(r)
20,22,24,26,28,30,32,34; 0-2

... debugging happens...

Ok, so the problem is that depending on the list length, _fold_multivariate_expand can create a new RangeSet but does not associate the autostep in this case.
The fix is as simple as this:

diff --git a/lib/ClusterShell/RangeSet.py b/lib/ClusterShell/RangeSet.py
index a3cd3de..a7b33ec 100644
--- a/lib/ClusterShell/RangeSet.py
+++ b/lib/ClusterShell/RangeSet.py
@@ -1131,7 +1131,8 @@ class RangeSetND(object):
         if len(self._veclist) * (len(self._veclist) - 1) / 2 > max_length * 10:
             # *** nD full expand is preferred ***
             pads = self.pads()
-            self._veclist = [[RangeSet.fromone(i, pad=pads[axis])
+            self._veclist = [[RangeSet.fromone(i, pad=pads[axis],
+                                               autostep=self.autostep)
                               for axis, i in enumerate(tvec)]
                              for tvec in set(self._iter())]
             return

I'm however way too lazy to add the new test I'd be asked to add if I were to submit that to gerrit, so feel free to push that fix :D

[degremont]

Patch here: https://review.gerrithub.io/c/cea-hpc/clustershell/+/424491