Regarding the comments around reproducibility, what parts of the CI are people having trouble reproducing now? I'm in favour of making our AMIs public for transparency reasons (and so that people can provide suggestions on how to improve them), but I'm not sure it would help in terms of reproducibility for any system other than Windows. When I have an error in CI I generally just do a `make clean` in my mxnet root source dire, then copy the failing command from CI, i.e. `tests/ci_build/ci_build.sh cpu --dockerbinary docker make DEV=1 USE_PROFILER=1 USE_CPP_PACKAGE=1 USE_BLAS=openblas -j$(nproc)`. Are there CI tasks (other than Windows) that don't work for people? If so maybe we can help fix those?
On Sat, Jan 6, 2018 at 11:50 AM, kellen sunderland < kellen.sunderl...@gmail.com> wrote: > +1, thanks for the work Marco. > > On Sat, Jan 6, 2018 at 12:24 AM, Naveen Swamy <mnnav...@gmail.com> wrote: > >> this sounds fine to me as long as there is at least one MXNet committer >> who >> is also an admin. >> >> thanks Marco for making this happen :) >> >> - Naveen >> >> On Fri, Jan 5, 2018 at 2:54 PM, Marco de Abreu < >> marco.g.ab...@googlemail.com >> > wrote: >> >> > I'm proposing following permissions: https://i.imgur.com/uiFBtuW.png. >> The >> > meaning of every permission is explained at https://wiki.jenkins.io/ >> > display/JENKINS/Matrix-based+security. >> > >> > Any objections? >> > >> > On Fri, Jan 5, 2018 at 11:03 PM, Marco de Abreu < >> > marco.g.ab...@googlemail.com> wrote: >> > >> > > I'm currently working on a prototype of SSO based on GitHub and a few >> > > issues arose: >> > > >> > > We are not able to use the permission strategy which determines the >> > access >> > > rights based on the read/write permission to a project as the >> > > Jenkins-plugin is not able to resolve the link between Jenkins-jobs >> and >> > > GitHub-repositories. Instead I would propose to use a role-based >> approach >> > > using https://wiki.jenkins.io/display/JENKINS/Role+Strategy+Plugin. >> In >> > > this case we would have three roles: Anonymous, Administrator and >> > > Committer. While everybody would authenticate using their regular >> GitHub >> > > account, the role assignment would have to happen manually. >> Considering >> > > that the amount of administrators and committers doesn't change that >> > > frequently, this shouldn't be too much of an issue - auto populating >> the >> > > status is not possible unfortunately. >> > > >> > > Reason for splitting Administrators and Committers into two separate >> > roles >> > > has a security reason. At the moment, we're using Chris Oliviers >> GitHub >> > > credentials to populate the commit status. If all committers would >> gain >> > > full admin rights, they would have access to these credentials. Chris >> is >> > > not fine with this approach and would like to limit the amount of >> people >> > > with access to his credentials as much as possible. >> > > >> > > In order to address his concerns, I propose to add Chris to the >> committer >> > > as well as to the admin role, while all other committers will only >> > receive >> > > the committer role without read access to the credentials. In a later >> > > email, I will make a proposal for the detailed committer role rights. >> You >> > > can check all available options at https://wiki.jenkins.io/ >> > > display/JENKINS/Matrix-based+security. >> > > >> > > All people who have access to the underlying AWS account would be >> granted >> > > the Administrator role with full access. At the moment, this would be >> > > Meghna Baijal, Gautam Kumar and myself. >> > > >> > > An alternative solution would be to create a bot account specifically >> for >> > > MXNet CI and use its credentials instead of Chris'. This account >> requires >> > > write permission to the repository, but would give us the advantage >> that >> > > these credentials would be shared within the committers and thus >> making >> > the >> > > restrictions regarding credentials obsolete (and Chris would be happy >> not >> > > the see his face within every single PR :P ). I've asked around and >> > > received the feedback from multiple people that Apache Infra does not >> > want >> > > to grant bot accounts write permission to a repository, but I would >> like >> > to >> > > confirm back considering that AppVeyor, for example, has a bot account >> > with >> > > write permission. I would like to check back with a mentor and create >> an >> > > Apache Infra ticket to request details and permission. >> > > >> > > I would propose to take both approaches at the same time, meaning we >> can >> > > start with Chris in the committer AND admin role while trying to get >> > > permission for a bot account in the meantime. >> > > >> > > wdyt? >> > > >> > > On Fri, Jan 5, 2018 at 8:21 PM, Chris Olivier <cjolivie...@gmail.com> >> > > wrote: >> > > >> > >> I am fine without a vote unless a vote is required? Any objections, >> > >> anyone? You're sort of adding functionality here, not changing or >> > >> restricting... We can always change to Apache later. >> > >> >> > >> On Fri, Jan 5, 2018 at 11:18 AM, Marco de Abreu < >> > >> marco.g.ab...@googlemail.com> wrote: >> > >> >> > >> > I'd be in favour of GitHub. Shall we open a vote or would you like >> me >> > to >> > >> > create a POC with GitHub first and afterwards we can check if >> that's >> > >> > enough? >> > >> > >> > >> > -Marco >> > >> > >> > >> > On Fri, Jan 5, 2018 at 8:13 PM, Chris Olivier < >> cjolivie...@gmail.com> >> > >> > wrote: >> > >> > >> > >> > > Apparently Apache supports OATH, so I am open to either. >> > >> > > Good idea for the docker thing. >> > >> > > >> > >> > > On Fri, Jan 5, 2018 at 11:02 AM, Marco de Abreu < >> > >> > > marco.g.ab...@googlemail.com> wrote: >> > >> > > >> > >> > > > GitHub SSO allows the neat feature that login and permission >> can >> > be >> > >> > > > selected depending on the access rights a user has to a >> project. >> > >> > Somebody >> > >> > > > with write access (committers) would be get different >> permissions >> > >> than >> > >> > > > somebody with only read access. >> > >> > > > >> > >> > > > We could check back with Apache for SSO, but this would involve >> > >> Apache >> > >> > > > infra. We could put it up to a vote whether to use GitHub or >> > Apache >> > >> > SSO. >> > >> > > > >> > >> > > > In order to reproduce a build failure we have been thinking >> about >> > >> > > changing >> > >> > > > the ci_build.sh in such a way that it can be run manually >> without >> > >> > > Jenkins. >> > >> > > > The setup I took over binds the Jenkins work directory into the >> > >> docker >> > >> > > > containers and uses a few hacks which are hard to reproduce >> > >> locally. We >> > >> > > > plan to reengineer this script to make it easier to run >> manually. >> > >> > > > But making the AMI public is a good idea! We plan to make the >> > whole >> > >> > > > infrastructure code (based on Terraform) completely public - at >> > the >> > >> > > moment >> > >> > > > it's in a private repository as it contains credentials, but >> they >> > >> will >> > >> > be >> > >> > > > moved to KMS soon. It would definitely be a good approach to >> just >> > >> > supply >> > >> > > > the AMI so everybody could recreate the environment in their >> own >> > >> > account. >> > >> > > > >> > >> > > > -Marco >> > >> > > > >> > >> > > > Am 05.01.2018 7:51 nachm. schrieb "Chris Olivier" < >> > >> > cjolivie...@gmail.com >> > >> > > >: >> > >> > > > >> > >> > > > Well, login to the Jenkins server, I would imagine. >> > >> > > > >> > >> > > > github or Apache SSO (does Apache support OAUTH?) seems like a >> > good >> > >> > idea >> > >> > > as >> > >> > > > long as there's a way to not let everyone with a github account >> > log >> > >> in. >> > >> > > > >> > >> > > > Access to actual slave machines could be more restricted, I >> > imagine. >> > >> > > > >> > >> > > > Eventually, a public current AMI for a build slave would be >> good >> > in >> > >> > order >> > >> > > > to reproduce build or test problems that can't be reproduced >> > >> locally. >> > >> > > > >> > >> > > > wdyt? >> > >> > > > >> > >> > > > >> > >> > > > >> > >> > > > On Fri, Jan 5, 2018 at 10:41 AM, Marco de Abreu < >> > >> > > > marco.g.ab...@googlemail.com> wrote: >> > >> > > > >> > >> > > > > Would it be an acceptable solution if we add SSO or do you >> also >> > >> want >> > >> > > > access >> > >> > > > > to the actual AWS account and all machines? >> > >> > > > > >> > >> > > > > Yes, the build jobs are automatically getting created for new >> > >> > branches. >> > >> > > > > >> > >> > > > > -Marco >> > >> > > > > >> > >> > > > > Am 05.01.2018 7:35 nachm. schrieb "Marco de Abreu" < >> > >> > > > > marco.g.ab...@googlemail.com>: >> > >> > > > > >> > >> > > > > I totally agree, this is not the way it should work in an >> Apache >> > >> > > Project. >> > >> > > > > It's running on an isengard account, meaning it is only >> > accessible >> > >> > for >> > >> > > > > Amazon employees. The problem is that a compromised account >> > could >> > >> > cause >> > >> > > > > damage up to 170,000$ per day. There are alarms in place to >> > notice >> > >> > > those >> > >> > > > > cases, but we still have to be very careful. These high >> limits >> > >> have >> > >> > > been >> > >> > > > > chosen due to auto scaling being added within the next >> week's. >> > >> > > > > >> > >> > > > > I'd be happy to introduce a committer into the CI process and >> > all >> > >> the >> > >> > > > > necessary steps as well as granting them permission. The only >> > >> > > restriction >> > >> > > > > being that it has to be and Amazon employee and access to >> > console, >> > >> > > master >> > >> > > > > and slave only being possible from the Corp network. >> > >> > > > > >> > >> > > > > There is no open ticket. What would you like to request? >> > >> > > > > >> > >> > > > > -Marco >> > >> > > > > >> > >> > > > > >> > >> > > > > Am 05.01.2018 7:22 nachm. schrieb "Chris Olivier" < >> > >> > > cjolivie...@gmail.com >> > >> > > > >: >> > >> > > > > >> > >> > > > > Like John and other mentors were saying, it's not proper for >> CI >> > to >> > >> > be a >> > >> > > > > closed/inaccessible environment. Is it running on an >> Isengard >> > >> > account >> > >> > > or >> > >> > > > > in PROD or CORP or just generic EC2? I think that we should >> > >> remedy >> > >> > > this. >> > >> > > > > It's very strange that no committers have access at all. Is >> > >> there a >> > >> > > > ticket >> > >> > > > > open to IPSEC? >> > >> > > > > >> > >> > > > > On Fri, Jan 5, 2018 at 10:17 AM, Marco de Abreu < >> > >> > > > > marco.g.ab...@googlemail.com> wrote: >> > >> > > > > >> > >> > > > > > Hello Chris, >> > >> > > > > > >> > >> > > > > > At the moment this is not possible due Amazon AppSec >> > >> (Application >> > >> > > > > security) >> > >> > > > > > restrictions which does not permit user data and >> credentials >> > on >> > >> > these >> > >> > > > > > machines. >> > >> > > > > > >> > >> > > > > > I have been thinking about adding single sign on bound to >> > >> GitHub, >> > >> > but >> > >> > > > we >> > >> > > > > > would have to check back with AppSec. >> > >> > > > > > >> > >> > > > > > Is the reason for your request still the ability to start >> and >> > >> stop >> > >> > > > > running >> > >> > > > > > builds? >> > >> > > > > > >> > >> > > > > > Best regards, >> > >> > > > > > Marco >> > >> > > > > > >> > >> > > > > > Am 05.01.2018 7:11 nachm. schrieb "Chris Olivier" < >> > >> > > > cjolivie...@gmail.com >> > >> > > > > >: >> > >> > > > > > >> > >> > > > > > Marco, >> > >> > > > > > >> > >> > > > > > Are all committers able to get login access to the Jenkins >> > >> Server? >> > >> > > If >> > >> > > > > not, >> > >> > > > > > why? >> > >> > > > > > >> > >> > > > > > -Chris >> > >> > > > > > >> > >> > > > > >> > >> > > > >> > >> > > >> > >> > >> > >> >> > > >> > > >> > >> > >
