Repository navigation

fix: Joom 1.1-3 — COUNT(DISTINCT) under memory pressure, plan task binaries, first/last aggregates, join-condition cost - #4

Merged

msafonov merged 13 commits into

branch-1.1from

joom/1.1-fix3

Oct 7, 2026

Commits on Oct 6, 2026
- fix: price the join condition of a sort-merge join in the cost-based engine choice
  
  msafonov
  and
  claude
  committed
  bef0b9f
- fix: leave validity-interval conditions of sort-merge joins out of smjCondition
  
  msafonov
  and
  claude
  committed
  250bfd2
- fix: keep native operator protobufs out of Java-serialized plans
  
  msafonov
  and
  claude
  committed
  61590a8
- perf: native groups accumulator for Spark first/last and boolean min/max
  
  msafonov
  and
  claude
  committed
  2fa930b
- test: differential fuzz suite for first/last and boolean min/max aggregates
  
  msafonov
  and
  claude
  committed
  25f52bf
- fix: run a PartialMerge aggregate so that it spills instead of emitting a group twice
  
  msafonov
  and
  claude
  committed
  6c8193e
- fix: exempt a validity interval only when both bounds read one row of one input
  
  msafonov
  and
  claude
  committed
  88c3edf
- fix: let a Spark producer feed a root Comet shuffle through the columnar format
  
  msafonov
  and
  claude
  committed
  b5fdf86
Commits on Oct 7, 2026
- fix: recognise a constant time shift without referencing TimeAdd, which Spark 4.1 renamed
  
  msafonov
  and
  claude
  committed
  c7df219
- perf: gather a wide native sort's payload from concatenated columns
  
  msafonov
  and
  claude
  committed
  6975eba
- perf: sort a wide native sort's row keys by an 8-byte prefix first
  
  msafonov
  and
  claude
  committed
  5e4ef4a
- feat: recalibrate the sort lines of the engine cost table on prod-sized tasks
  
  msafonov
  and
  claude
  committed
  3023e0b
- fix: keep the calib29 prices of nested sorts
  
  msafonov
  and
  claude
  committed
  c90d450

Navigation Menu

Repository navigation

fix: Joom 1.1-3 — COUNT(DISTINCT) under memory pressure, plan task binaries, first/last aggregates, join-condition cost - #4

fix: Joom 1.1-3 — COUNT(DISTINCT) under memory pressure, plan task binaries, first/last aggregates, join-condition cost#4
msafonov merged 13 commits into
branch-1.1from
joom/1.1-fix3

Commits on Oct 6, 2026

fix: price the join condition of a sort-merge join in the cost-based engine choice

fix: leave validity-interval conditions of sort-merge joins out of smjCondition

fix: keep native operator protobufs out of Java-serialized plans

perf: native groups accumulator for Spark first/last and boolean min/max

test: differential fuzz suite for first/last and boolean min/max aggregates

fix: run a PartialMerge aggregate so that it spills instead of emitting a group twice

fix: exempt a validity interval only when both bounds read one row of one input

fix: let a Spark producer feed a root Comet shuffle through the columnar format

Commits on Oct 7, 2026

fix: recognise a constant time shift without referencing TimeAdd, which Spark 4.1 renamed

perf: gather a wide native sort's payload from concatenated columns

perf: sort a wide native sort's row keys by an 8-byte prefix first

feat: recalibrate the sort lines of the engine cost table on prod-sized tasks

fix: keep the calib29 prices of nested sorts