Just to add here that the columns and not just the rows are also sampled at each node so the trees are deliberately prevented from learning too much. This helps improve diversity and reduce overfitting. This is typically around 1/3rd of the number of columns and controlled by the mtry parameter.