What about the other 15% of the time?
for these I would use x -> as before 
underscores also have a “15% ambiguity” for functions like map(_, x), filter(_, x)
and realistically, I think front/back pipes have a (small but nonzero) chance of eventual inclusion, underscores have near-zero, so part of it is just an attempt to be economical with my advocacy
To clarify, I was referring to the effect of #24990 on piping, not its effect on creating small anonymous functions like 2_ or _ + 1. If I have to use x -> f(a, x, b) in order to write a pipe, then I just won’t write a pipe at all.
Is there already a forward facing function composition operator?
So, one from left to right?
I have not seen one, and I would create an issue on Github, if there is none. 
Of possible operators, I personally probably prefer something available as a ligature, so among:
Most often for generic data manipulation you need anonymous functions, and that’s why there are single underscores in those examples: as in @p tbl |> map(_.field + 1).
For the common argument order and no anonymous functions, there’s no need for underscores: eg @p tbl |> map(sum) or @p 1:10 |> filter(isodd) |> map(exp).
There is one in CompositionsBase.jl: ⨟. The package has hundreds of dependents, but pretty sure not because they use this operator – ⨟ isn’t really used much.
I think that a valuable example, as this looks significantly more familiar.
