Repository navigation
Commit a53c4f4
committed
fix: do not infer semi/anti join coverage from range-bounded distinct counts
`max_distinct_count` falls back to the cardinality of `[min, max]` when a
column has no `distinct_count`. That number is an upper bound that only
applies when the column has more rows than possible values, so values repeat
and nothing in the statistics says how many of the possible values occur.
`estimate_semi_join_cardinality` used it as the inner side's NDV in
`min(outer_ndv, inner_ndv) / outer_ndv`, which asserts that every value in
the range occurs and that (almost) every outer key is covered. On TPC-H SF1,
`customer NOT EXISTS orders` was estimated at 1 row against 50,004, and the
anti join in Q22 at 0 rows against 6,384.
`max_distinct_count` now reports where its count came from (measured, row
count, value range, unknown), with unchanged numbers. The semi join path uses
an explicit 0.5 overlap, PostgreSQL's `eqjoinsel_semi` default for a side
without a measured distinct count, for an inner key whose count is
range-derived. Measured and row-count-derived counts are used as before, so
`orders EXISTS customer` and `orders EXISTS (filtered customer)` keep their
estimates. Inner joins are untouched (apache#25618).
Closes apache#256171 parent 0d79252 commit a53c4f4
2 files changed
Lines changed: 410 additions & 20 deletions
0 commit comments