GHC hacking without tears?

[Manuel M T Chakravarty]

Tim Chevalier:
> Now, usually the work stoppage is just a matter of waiting for darcs
> to complete. But today the HEAD was broken. so since I didn't know how
> long it was going to take for a patch that fixed it to get checked in,
> I spent a couple hours trying to roll back patches to get into a state
> where I at least had a version of GHC that would build so I could
> continue with other work while waiting for the fix to get checked in.
> This didn't work, so now I'm waiting while a new tree checks out from
> scratch.
>
> Is there a way to insulate myself from GHC breakage (and I realize
> that in the HEAD, a certain amount of breakage is inevitable) while
> still checking in my changes on a regular basis? I see two
> alternatives:
> 1) commit my changes frequently, on a branch and merge it to the HEAD
> at some future time (but past experience has shown that merging it
> later is just too painful)
> 2) never commit my changes until I consider my work "done", then
> commit them all at once (but it generally seems better to commit small
> amounts of code often; plus, this has more or less the same problem
> with merging)

I have two repos, ghc-test and ghc.  I develop in ghc and never edit  
in ghc-test.  (This is important as otherwise, local, non-recorded  
edits can make your local validate succeed, although your patches do  
break the tree.)

First thing, each morning, I pull into ghc-test and run validate.   
When it breaks, I complain on cvs-ghc and unpull the patches I pulled  
this morning.  (If I am not too busy, I may try to fix the problem,  
but that's optional.)

I develop in ghc and record patches there.  Then, I push them to ghc- 
test and run validate on them.  If validates goes through, I push to  
darcs.haskell.org.

Once in a while, I pull from ghc-test into ghc.

------

This scheme has a fairly low overhead for me.  If ghc-test breaks,  
there are usually not that many patches to unpull, and I am somewhat  
isolated from head breakages.  However, the **most important thing**  
is that everybody validates every single patch before pushing.  Every  
head breakage that we can avoid safes the time of many people.

Manuel