Eventual-Inc
Daft
Blog
Docs
Changelog
Blog
Docs
Changelog
Overview
Branches
Benchmarks
Runs
Performance History
Latest Results
Merge origin/main into desmond/native-extensions-ray-support Resolve conflicts in flotilla.py (combine extension runtime_env with worker dashboard/task env vars) and examples/hello (keep daft_extension macro with main's arrow_array/arrow_schema imports). Co-authored-by: Cursor <cursoragent@cursor.com>
desmond/native-extensions-ray-support
6 hours ago
fix(distributed): extend bisect autoscale to track GPU and memory dimensions
DogerW666:add-bisect-autoscaling-strategy
6 hours ago
fix(inline-agg): gate packed-u64 path on MIN_AVG_STRING_BYTES_PER_ROW CodSpeed flagged a -11.22% regression on test_tpch_sql[1-in-memory-10] (TPC-H Q1, scale 1). Q1's CHAR(1) `l_returnflag` / `l_linestatus` are short enough that the symbolize pass costs more than the saved per-row hash/compare work, even with the packed-u64 typed hash map. Local microbench showed the opposite but failed to capture the cache / branch behavior on CodSpeed's runner and TPC-H's actual data shape. Mirror the avg-bytes-per-row gate that `agg_symbolized_path` already uses (`MIN_AVG_STRING_BYTES_PER_ROW = 16`): tally bytes across both Utf8/Binary cols up front and bail when total bytes per row falls below the threshold. Short-string shapes (Q1) now fall through to the existing generic-hash-on-raw-strings path (unchanged from main). Long-string shapes still get the packed-u64 win. Local Rust benchmarks confirm: - Q1 short strings (CHAR(1) × CHAR(1)): packed-u64 no longer fires; inline path matches main behavior (1.36x-1.68x vs Daft fallback, same as before this PR series). - Long-string two-col shapes (avg >= 16 bytes/row): packed-u64 still fires, still ~1.08x-1.19x faster than PR #6748 inline path. Also update doc comments on `test_inline_multi_col_short_strings_*` and `test_inline_packed_u64_short_strings_*` — both now exercise the gate's short-string skip branch and route through the generic path. Addresses #6924 CodSpeed regression.
BABTUNA:perf/packed-symbol-groupby
6 hours ago
feat(distributed): add bisect autoscaling strategy for faster cluster scale-up
DogerW666:add-bisect-autoscaling-strategy
6 hours ago
fix(show): keep "left" as hard-coded default to preserve doctests Reverts the previous "auto" default. Changing _SHOW_DEFAULT_ALIGN broke ~40 doctests across dataframe.py and expressions.py whose expected ASCII output assumes left-aligned numeric columns. Auto-alignment stays available via DAFT_SHOW_ALIGN=auto or an explicit align="auto" argument, matching the opt-in semantics documented in the updated PR body.
BABTUNA:feat/show-preview-defaults
7 hours ago
style(inline-agg): drop needless borrow on cols[i] in agg_packed_u64_path clippy::needless_borrow (denied under `-D clippy::style`) flagged the `&cols[i]` arguments to `symbolize_string_col`. `cols` is already a sequence of `&Series`, so indexing yields a reference directly. Fixes the `style` CI failure on #6924.
BABTUNA:perf/packed-symbol-groupby
8 hours ago
fix(show): make 'auto' the actual default and fix docstring drift Change _SHOW_DEFAULT_ALIGN from "left" to "auto" so the PR delivers what its description claims — numeric columns right-align by default without requiring an env var or explicit argument. Also drop the stale "off" token from the DAFT_SHOW_MAX_WIDTH docstring (only "none"/"null" disable the max-width clamp).
BABTUNA:feat/show-preview-defaults
9 hours ago
chore(inline-agg): apply formatting and address #6924 review nits - cargo fmt: wrap long lines in the agg_groupby_inline dispatch chain and the bench_packed_u64_two_strings helper call. - Hoist `use std::sync::Arc;` to the `bench` module top instead of inside `make_two_long_string_batch`. - Update doc comments for `test_inline_multi_col_long_strings` and `test_inline_multi_col_short_strings` — both were inherited from PR #6748 but now route through `agg_packed_u64_path` first under this PR's dispatch order, so the comments mentioning "active symbolized path" and "symbolization skipped via gate" were stale. Addresses Greptile P2 comments on #6924.
BABTUNA:perf/packed-symbol-groupby
9 hours ago
Latest Branches
CodSpeed Performance Gauge
0%
feat(ext): run native extensions on the Ray runner
#6760
6 hours ago
4b681ea
desmond/native-extensions-ray-support
CodSpeed Performance Gauge
-13%
feat(distributed): add bisect autoscaling strategy for faster cluster…
#6929
6 hours ago
59b1356
DogerW666:add-bisect-autoscaling-strategy
CodSpeed Performance Gauge
-12%
perf(inline-agg): pack two-string-column keys into u64 for typed FNV grouping
#6924
7 hours ago
ee0657c
BABTUNA:perf/packed-symbol-groupby
© 2026 CodSpeed Technology
Home
Terms
Privacy
Docs