You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit f99f277
Browse filesBrowse the repository at this point in the historyBrowse files
perf: charge an in-batch key group its row share of the parent
get_sliced_size() allocates per column for every buffered key group,
which made filtered semi/anti joins with many tiny key groups about
1.2x slower than main. Charge a group ending inside its inner batch
parent_size.div_ceil(num_inner) per row instead: O(1), and it cannot
overflow usize on 32-bit targets. Groups reaching the batch end are
still charged the whole parent.
0 commit comments