On March 9, 2019 5:48, Kevin Daudt wrote: > On Sat, Mar 09, 2019 at 10:19:03AM +0000, Dimitri Joukoff wrote: > > Hi, > > > > As a relatively novice user of git, there have been far too many times > > that I have lost data, sometimes quite a lot. So this proposal is > > about catering for the less experienced users and averting fits of > > anger and frustration. The only reason my computer still works is > > because my sub-conscious mind stops me from smashing it or throwing it > > against a wall. It seems my sub-conscious mind has a pragmatic view > > of the world and understands that whilst I may receive instantaneous > > satisfaction at the time, in the long term, the pain will be far > > worse, and thus prevents me from doing something rash. > > > > Yes, it can be very frustrating to lose things you did not intend to lose, so > making sure your tooling limits the chances of that happing is certainly a > worthwile goal. > > > > > Below is the detail of my proposal: > Whenever a command is issued in > > git that will cause git to overwrite or delete *ANY* files whose > > current state isn't already recorded in the repository, git should > > prompt the user to confirm the operation. This includes untracked > > files as well as files that are in the 'not staged' > > and 'staged' lists. > > > > To make the consequences of the command transparent, the confirmation > > should include a list of files that will be affected (perhaps in a > > similar way to how git status works). The scope of the files listed > > must match the scope of the command to be executed. No hidden > > changes, no side-effects. > > > > Saying no to the confirmation should abort the command. > > > > It may be useful to allow confirmation of individual files, but as a > > novice user, I can't argue this point objectively, nor reason about > > its implications and complexity. > > > > This feature should be enabled by default whenever a clone or init > > operation are performed. > > > > The user should be able to progressively reduce the range of commands > > and amount of confirmation interactions that take place. The > > configuration technique could follow the already established procedure > > for other configurable data in git. So this could be done globally > > for the user, or locally within each repository. > > > > > > As a novice user, there may be further useful extensions of this idea, > > about which I'm unable to reason. So I welcome further elaboration of > > the idea discussed above. > > A lot of confirmations only result in people automatically dismissing them > (confirmation saturation), missing the goal of what you intend. > > Instead of asking for confirmation, it's much better to allow people to undo > these mistakes. You see the same pattern in gmail for example, where they > hardly ask you for any confirmation, but instead show an undo button that > allows you to undo the last operation. In my opinion this is a better way to > go then to add confirmations everywhere. > > I know this has come up on the git mailing list more often, but I cannot find a > relevant thread at this moment. First, I really do not like the idea of confirmations. This could complicate scripting and would drive much of the work I do with git in Jenkins up a wall. You would need access to stdin for almost anything. Second, I think an automatic undo has merit and could further differentiate git from other DVCS and VCS systems. My thought is along the lines of starting with the stash concept for each undo - almost like an auto-stash. Basically, any time you perform a working-directory modifying operation, a stash-like commit is added to the repository at HEAD (possibly ignoring .gitignore or precious files, like --include-untracked but in a config like undo.untracked=on, to avoid needing to remember to do this). I envision it being a stash without modifying the working-directory or changing the repository state other than the "undo" unlike what stash does. Considering the performance hit this *will* cause, I would want an option to not do this (say, undo.enable=on/off, off by default unless there was some newbie metric <j/k>, or maybe undo.fearful=high <j/k>), and a limit to the number of undoes (undo.limit=n), and an auto-drop capability so that when you finally commit, you have the option to drop the undoes of the previous parent commit (undo.autoclean=on/off), or limit it to cleaning after more than one commit is done beyond the commit where the undo exists (undo.autoclean=n). Deriving an "undo" off of a specific parent commit (HEAD), instead of deriving "undoes" on each other, might be helpful in resolving the question of how to you roll off (get rid of) undoes over time - making it just based on the time of the snapshot and how many you want to keep. The reason I would hang it off of a HEAD commit is that a checkout/switch would preserve the undo stack so that when you returned to a branch, its undo stack would be available, like stash. I would also see an impact on gc, potentially, to clean up old undoes beyond a specific date. This might need to start as a modification to stash, like --keep-index, but more like taking picture, for example, --snapshot-only. Once you had that, building an undo stack should be straight-forward, and undoing would be virtually the same as a stash apply (might even *be* stash apply if the "undo" and stash were somehow the same thing conceptually). We're probably also talking about a new command and subcommands, very similar to the stash structure but querying either from HEAD or a specified commitish. If I only had the time... ;) Just my musings. Randall -- Brief whoami: NonStop developer since approximately 211288444200000000 UNIX developer since approximately 421664400 -- In my real life, I talk too much.
