We have a scenario that someone might shed some light on.
We are sucessfully implementing Shared Development Server and Data Registries with shared cache and user groups in our deep learning development. It all works perfectly, shared cache, remotes, etc except one particular issue very specific for our use case.
Our current approach for data registries is that for any particular dataset we have a master copy of raw original data in that data registry repo on the server. Any data injections are done only from within that repo. So we always have a copy of raw datasets (not hashed) along with caches and a copy on the remote, which are hashed.
The issue with this scenario is that we need several users be able to add, commit and push from within a particular repo. And even when they are in the same group we have to manually set permissions to 2775 and 0664 for everythiong in .dvc folder (cache is not there as per shared cache scenario) so that dvc works. Git commands work without any permission issues. And from time to time permissions break on the dvc remote too and we have to reset them manually to 2775 and 0444.
Is it something that we might need to setup dvc some other way? Or is it just not supported but can be supported and possibly needs contribution?