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
Summary:
Pull Request resolved: #992
The first real per-chunk statistic. Chunk stats are still purely positional;
this records how many nulls each chunk holds, so a later reader can skip chunks
on null-related predicates and we can eventually retire the read-time
byte-peeking row estimates in ChunkedDecoder.
- The writer counts nulls per chunk during encode, only when chunk stats are
enabled. The fleet default is off, so no writer pays for this yet.
- Stored as a new field appended to the existing chunk-stats flatbuffer. Files
written before this simply lack it, and the reader reports "unknown" rather
than a wrong value.
- Reader accessor StreamIndex::chunkNullCount returns the count, or nullopt for
pre-stats files.
Rationale: https://docs.google.com/document/d/1cTCXje3RkPdhwFGK-FTY5muy4Cky2qnfnP-e_VFkjQA/edit?tab=t.0#heading=h.rpljwjiiglku
Differential Revision: D111913795
0 commit comments