Comment by stellalo

That is essentially an online subgradient descent, with fixed stepsize 1, for minimizing the L1 loss.