In examples of the present invention, delivery of a segment of content is done using an initial congestion window, preferably chosen to give a delivery rate that ensures timely delivery of that segment. However, under high levels of packet loss, the delivery rate may not be sufficient. Hence, at any stage during delivery of a segment, the actual delivery rate is measured, and used, together with the previous congestion window size, the amount of content segment data still to be delivered, and the time left for delivery, to calculate an updated congestion window. The remainder of the segment is then delivered using this revised congestion window.