Suggested by @Xover at https://commons.wikimedia.org/w/index.php?title=Commons_talk%3AMaximum_file_size&type=revision&diff=606497138&oldid=605931293
T275752 was noticed because it caused database errors and timeouts, however a regression that was not so bad might not be picked up by humans in logs. We could record the rough throughput (size / time) for all files over 100MB in statsd to get us an average speed. Probably it should be split by DC as well.
If the resulting average isn't too noisy, we could have alerting based on it.