HTTP tests have a very high-volume response compared to the request size. For example response bodies for censored webpages may weight up to 4MB. If the request is repeated hundreds of times due to traceroutes, TCP Segmentation, various HTTP camouflage and other tricks it can significantantly increase the bill for traffic without a good reason. OONI Probe should terminate the connection as soon as it is reasonably sure that the HTTP body is not going to differ from the one that it has already recorded.
A limited TCP initcwnd (initial congestion window) at servers ensures that OONI Probe has one RTT to decide if it wants to continue fetching the data.
OONI Probe may also set low SO_RCVBUF or TCP_WINDOW_CLAMP to do TCP backpressure.
TBD: is it possible to drain the kernel buffer without sending an ACK with some non-zero window? The data that is already in the buffer should probably make its way into the measurement data.