MonadFail instance (feature request)

[dmwit]

It would be very handy to have a MonadFail instance that retrys. As a minimal motivating example, consider:

pop :: TVar [a] -> STM a
pop v = do
    a:as <- readTVar v
    writeTVar v as
    return a

With a MonadFail instance, this is a very convenient way of writing a transaction that waits until there's a value available in the stack.

More generally, such an instance would allow one to write a transaction as if all your favorite patterns matched, and the transaction would then efficiently block until it's so.

[samuel-gelineau-at-well-dot]

Note that historically, STM used to have a Monad.fail method which was defined as fail = error. Then, when MonadFail was introduced, STM was not given a MonadFail instance with fail = error, as that would go against the spirit of MonadFail.

I don't know why STM originally had fail = error, I agree that it should be fail = retry instead. If STM currently had fail = error, then changing it to fail = retry would be a breaking change. But luckily, MonadFail was introduced a long time ago, so nobody is relying on fail = error anymore, so I do think that it would be a good time to define fail = retry!

[simonmar]

The example is compelling, but I worry that it would turn accidental pattern match failure into deadlock, which would make debugging much harder. Happy to hear other opinions though.