A user reported a cluster outage due to an errant client which was updating the same row in a hot loop. This led to an out-of-disk situation and after restart tickled #24214.

The user is wondering whether we can limit the number of old versions per row or the number of updates per row somehow. This is not possible with the current MVCC layout (some metadata would have to be added and kept current for keys which may be monitored), and fine tuning the constants would be a difficult task to get right.

Still we should take a look at how these problems are solved in other products.

There may be some connection to #17229.