Repository navigation
Epic: Duckdb projection expression pushdown #8310
Description
We want to push down scalar and aggregate functions that are part of SELECT into vortex.
Example: SELECT strlen(col) doesn't need to decompress the strings.
Subtasks:
- Scalar functions pushdown Scalar functions pushdown to Duckdb #8303
- Migrate cast pushdown to Vortex Cast pushdown for duckdb #8620 Don't push down CAST if there's a TRY_CAST duckdb/duckdb#23535
- Unrouped aggregate function pushdown Ungrouped aggregate function pushdown for duckdb #8645
- Grouped aggregate pushdown
Upstreaming:
- Cast pushdown (upstream) Push down type casts to file readers duckdb/duckdb#22788 Move type pushdown after filter pushdown, don't push it filter is present duckdb/duckdb#23734
- Upstream scalar functions pushdown Push down scalar functions to file readers duckdb/duckdb#23528
Additional work:
- Push down chains of functions f_1(...(f_n(col)) if leaf child is
BoundColumnRef. - Push down multiple expressions for column if we can push all of them i.e.
allow pushing down (strlen(col), col). This requires a change on projection
ids passed to vortex. In this example we need to pass projection_ids=[0, 0].
PR with upstreaming aggregation pushdown callback duckdb/duckdb#24697 is declined because Duckdb maintainers feel their plan-time pass is sufficient.
Reacted by Valerii Mironov
Activity
Pushdown some expressions to Dict layout reader (#8341)
Pushdown some expressions to Dict layout reader (#8341)
myrrc commentedon Jun 18, 2026
This change regressed vortex-compact significantly on the same query. We can't solve this issue now as ZSTD stores lengths with data, so you need to decompress. Implementing a kernel for byte_length which only used decompressed length without constructing a VarBinView didn't yield any performance improvements. We'll accept the regression as there's no way for reader to distinguish vortex vs. vortex-compact
[-]Tracking Issue: Duckdb projection expression pushdown[/-][+]Epic: Duckdb projection expression pushdown[/+]