perf: fast-path column decoders for non-null data + execute/2 for DDL

[hugobarauna]

Context

This PR is the result of an AI-driven performance optimization experiment on Dux (a DuckDB dataframe library) using Claude Code. The code was not human-reviewed beyond passing Dux's test suite. The intent is to show potential optimization approaches — feel free to close this PR or cherry-pick individual ideas.

Companion Dux PR: elixir-dux/dux#44

Changes

1. Fast-path column decoders when no nulls are present

DuckDB always sends validity bitmaps even when all values are valid (all 0xFF bytes). The current decoders check bitmap_valid? and call through a decoder closure per element — for 400k rows × 5 columns, that's 2M unnecessary function calls.

This adds an all_valid?/1 check per column (O(1) binary comparison) and dispatches to plain decoders that skip both the bitmap check and the closure call:

# Before (per element): bitmap check + closure call
value = if bitmap_valid?(validity, index, offset), do: decoder.(int)

# After (per column): one check, then plain binary traversal
if all_valid?(validity), do: decode_s64_plain(binary)

Fast paths added for: s8-s64, u8-u64, f16-f64, string/binary (32+64 bit offsets), and date32.

Impact: ~15-20% improvement on Adbc.Result.to_map for typical non-null data.

2. Adbc.Connection.execute/2 for DDL statements

Adds a public function that uses command dispatch (no stream setup/teardown) for statements that return no data (CREATE VIEW, DROP TABLE, SET, etc.):

Adbc.Connection.execute(conn, "CREATE TEMPORARY VIEW my_view AS (SELECT ...)")

This avoids the stream → stream_results → unlock cycle that query/2 goes through. Saves ~50-100μs per call.

3. View cleanup in GC handler (⚠️ Dux-specific, open to alternatives)

Modified handle_command({:delete_on_gc, table_name}) to check for a __dux_v_ name prefix and use DROP VIEW IF EXISTS instead of DROP TABLE IF EXISTS. This is needed because Dux's performance optimization creates temp views instead of temp tables, but the existing GC mechanism only drops tables.

I'm open to alternative approaches here. Some options:

• A separate adbc_delete_view_on_gc_new NIF function
• An additional parameter to adbc_delete_on_gc_new specifying the drop command
• A generic adbc_delete_on_gc_new/3 that accepts a custom SQL template

The prefix-based approach is the least invasive but couples ADBC to a Dux naming convention.

4. Single-batch fast path for to_map/1

Added a pattern match for the common single-batch case that skips Enum.zip_with + Enum.flat_map.

Benchmark impact (measured in Dux context)

These ADBC changes contributed to closing the gap between Dux and Explorer on to_rows():

Operation	Before ADBC changes	After	Explorer
filter to_rows	1.5x slower	1.0x (parity)	baseline
mutate to_rows	1.5x slower	0.95x (faster)	baseline

---

🤖 Generated with Claude Code

[josevalim]

Thank you! I pushed each of those as individual commits. For the validity one, I opted-in to check the null_count on the C side, which should be quite more efficient (assuming it is populated correctly).