(Stochastic) subgradient methods are used in practice to optimize non-differentiable convex functions. They have a slower convergence rate than (stochastic) gradient descent though.

See for example : https://www.stat.cmu.edu/~ryantibs/convexopt-F15/lectures/07...