> On May 30, 2018, 8:17 p.m., Benjamin Mahler wrote: > > src/common/resources.cpp > > Lines 866-868 (original), 875-883 (patched) > > <https://reviews.apache.org/r/67274/diff/4/?file=2032028#file2032028line875> > > > > Can we reject 0? We were previously allowing it? > > Jie Yu wrote: > Yes, we debated that. I feel that specifing the resource to be zero is > not a valid use case anyway (simply not specify the resource in that case). > We plan to send an email to user list. > > Benjamin Mahler wrote: > I think this breaks things like quota: > > limit: cpus:10;mem:10;gpu:0 > vs > limit: cpus:10;mem:10 > > These meean two different things, in the latter there is no gpu limit > whereas the former has a limit of 0.
I reverted the patch to revision 2 where we only rejected non-zero scalar resource values which rounded to zero (resulting in empty resources). FYI, I checked the behavior for quota and see no difference before and after the patch in this revision; the allocator never sees `gpus:0`. AFACIT this cannot be modelled with the current `Resources` class which always drops empty resources in operations (e.g., invoked when parsing above resource string). - Benjamin ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/67274/#review204060 ----------------------------------------------------------- On May 30, 2018, 11:50 p.m., Benjamin Bannier wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/67274/ > ----------------------------------------------------------- > > (Updated May 30, 2018, 11:50 p.m.) > > > Review request for mesos, Jie Yu, Joris Van Remoortere, Michael Park, and > Neil Conway. > > > Bugs: MESOS-8945 > https://issues.apache.org/jira/browse/MESOS-8945 > > > Repository: mesos > > > Description > ------- > > Internally values of scalar resources are stored in a fixed point > representation. This can lead to (expected) precision loss. > > This patch updates the resource validation ensuring that values of > scalar resources are either zero or large enough so that they are > still represented as non-zero values after conversion to the internal > fixed point format. With that we can still parse resources containing > resource values explictly set to zero to be able to distinguish empty > and abset resources, but prevent propagation of values accidentially > rounded to an effectively empty result. > > While this patch should not change valid, intended use cases, it might > change the behavior for erroneous, already broken framework workflows. > > > Diffs > ----- > > src/common/resources.cpp 192d086e60b81c0ae8e26d4120beee2034f63d3e > src/tests/master_tests.cpp 6efcbd12902de32428cb7f23ed0d3967a18f749d > src/tests/resources_tests.cpp 1201efec3ff0abc268f896c1fe2330571390b3fd > > > Diff: https://reviews.apache.org/r/67274/diff/5/ > > > Testing > ------- > > `make check` > > > Thanks, > > Benjamin Bannier > >
