A tricky bug was found (thanks Leonardo for giving me a test case showing the bug!) when copying data from the buffer back into a vector: the "insert" call from the STL does the copy in the range [start, end[ so we were always missing the last element. Most of the time, this was not an issue since we had such a large buffer, but with small buffers, one could see that the last data point was missing... So now, the "end" iterator is always "last_required+1".