Add an in-memory `Widen` encoding for integers by connortsui20 · Pull Request #10864 · vortex-data/vortex

connortsui20 · 2026-10-27T14:31:27Z

Summary

Tracking Issue: Tracking Issue: Add Widen encoding for integers and decimals #10733

Read the tracking issue for context.

Note that this does not add Decimal support yet, though it will later.

Replaces #10223 and #10332, which are closed.

Changes

WidenArray::encode stores an integer array at the smallest width with the same signedness that holds its valid values. Slice, filter, mask, cast, fill null, take and compare run on the narrow values. Everything else, including aggregates and arithmetic until the next PRs in the stack, decodes through a cast, so results match the canonical array. Take executes eagerly instead of rewriting lazily, so that a Dict stays at the root, where the DuckDB exporter and the dictionary layout writer look for it.

Widen has no wire form yet, so ArraySession does not register it until the wire format lands, and serializing an array that holds a Widen fails like it does for any other unregistered encoding. Registering it now would also add a reader entry that could only fail.

WidenArray::encode_values is the deprecated variant that changes the dtype, for internal children such as list offsets and dictionary codes. The deprecation marks each future caller for its own decision.

For review, read the commits one at a time.

A registered VTable also gets a reader entry under its wire ID. For Narrow, which has no wire form yet, that entry can only fail, so it hides the allow_unknown foreign fallback and lists vortex.narrow in footers written without editions. NR1 (#10864) therefore leaves Narrow out of ArraySession, lowers every Narrow before the serializer lookup, and makes Narrow::serialize panic with a TODO. WR1 adds the registration together with the working deserializer. D26, the NR1 brief and the wire brief now say so. The NR1 brief also gets the final deprecation note of encode_values. Signed-off-by: Connor Tsui <[email protected]>

NR1 (#10864) no longer lowers a Narrow in ArraySerializationTree. In NR1 nothing outside the tests creates a Narrow, so the branch protected no real path, and a branch specific to Narrow does not belong in the generic serializer while nothing needs it. Serializing a Narrow now fails like it does for any other unregistered in-memory encoding. D26 is now deferred to CD4, the first PR where a Narrow reaches writers in ordinary use. CD4 chooses between a lowering specific to Narrow, a generic lowering for arrays that cannot be written, and no lowering. The NR1 brief, the CD4 and WR1 briefs, the design and the plan now say so. Signed-off-by: Connor Tsui <[email protected]>

codspeed · 2026-10-27T14:56:17Z

Merging this PR will regress 4 benchmarks

3 warnings, 1 note

⚠️ Walltime results may be unreliable

Some benchmarks used the Walltime instrument outside CodSpeed Macro Runners, which can make the results inconsistent.

Learn more about Macro Runners

See affected benchmarks

⚠️ 1 benchmark is dominated by system calls

System calls cannot be simulated consistently and are left out of the measure. The reported value understates the real cost. Switch to the Walltime instrument to measure them.

Learn more about system calls

See affected benchmarks

⚠️ Base and head ran on different runner types

Some benchmarks with significant performance changes were compared across different runner types, so the changes might not come from the code. Run base and head on the same runner type, or use CodSpeed Macro Runners for more stable results.

Learn more about Macro Runners

Open the report in CodSpeed to investigate

ℹ️ This run is excluded from the branch history

This commit was force-pushed, so this run no longer appears in the branch's performance history.

⚡ 8 improved benchmarks
❌ 4 regressed benchmarks
✅ 2212 untouched benchmarks
⏩ 370 skipped benchmarks¹
🗄️ 30 archived benchmarks run²

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	`BASE`	`HEAD`	Efficiency
❌	CPU Simulation	`filter_powerlaw_by_random[10000]`	889.4 ns	1,477.5 ns	-39.8%
❌	CPU Simulation	`take_filter_primitive_small_uncached_random_mask_random_indices[4096, 10]`	20.4 µs	24.7 µs	-17.44%
❌	Walltime	`threshold_avx2[80]`	2.6 µs	3 µs	-11.42%
❌	Walltime	`add_constant_shapes_neon[(32768, PerRowConstant)]`	14.2 µs	15.8 µs	-10.2%
⚡	Walltime	`mul_u64_nonnull_neon`	39.6 µs	20.6 µs	+92.05%
⚡	CPU Simulation	`bench_compare_varbin[(50000, 2)]`	164.5 µs	135.5 µs	+21.42%
⚡	CPU Simulation	`bench_compare_varbin[(50000, 64)]`	166.2 µs	137.4 µs	+21.02%
⚡	CPU Simulation	`encode_primitives[u8, (2000, 4)]`	163.7 µs	135.9 µs	+20.47%
⚡	CPU Simulation	`bench_compare_sliced_dict_varbinview[(7500, 10000)]`	230 µs	196.5 µs	+17.04%
⚡	CPU Simulation	`bench_compare_sliced_dict_varbinview[(10000, 10000)]`	236.1 µs	208.7 µs	+13.16%
⚡	CPU Simulation	`bench_compare_sliced_dict_varbinview[(1000, 10000)]`	92.2 µs	81.7 µs	+12.89%
⚡	Walltime	`compare_u8_neon`	1.7 µs	1.6 µs	+10.44%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

_{Comparing ct/narrow-in-memory (75c7f4a) with develop (c45aa5d)}

370 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩
30 benchmarks were run, but are now archived. If they were deleted in another branch, consider rebasing to remove them from the report. Instead if they were added back, click here to restore them. ↩

connortsui20 · 2026-10-27T14:56:18Z

+        {
+            (*values_ptype, *logical_ptype)
+        }
+        _ => vortex_bail!("Narrow requires integer dtypes, got {values} values for {logical}"),


Note that in the future this will also check for Decimal

NR1 (#10864) gave vortex-array/src/arrays/narrow/ the layout of the other encodings: one file per operation in compute/, and the vtable in vtable/. The integer Narrow brief now describes that layout, and the NR3, NR4 and NR5 file tables put their code there. The CD4 and wire briefs point to the same section. Signed-off-by: Connor Tsui <[email protected]>

connortsui20 · 2026-10-27T15:16:48Z

+//! Take executes its gather instead of leaving a lazy [`Dict`] under a `Narrow`, because code such
+//! as the DuckDB exporter and the dict layout writer looks for a `Dict` at the root of an array.


This is related to #10748

Widen stores an integer array in a strictly narrower integer type with the same signedness and nullability. The values child holds the validity and can have any encoding. try_new and a reduce rule keep Widen arrays from nesting. Execution casts the values to the logical dtype, and probe_scalar casts each child scalar. WidenArray::encode stores a PrimitiveArray at the smallest integer type of the same signedness that holds its valid values. The deprecated encode_values returns the same narrower values without the Widen, for internal children whose parent chooses their dtype, so that each future caller is marked for its own decision. Widen has no wire form yet, so ArraySession does not register it, and serializing an array that holds a Widen fails as it does for any other encoding that is not registered. A registered Widen would add a reader entry that can only fail, which hides the allow_unknown foreign fallback and lists vortex.widen in footers written without editions. Signed-off-by: Connor Tsui <[email protected]>

Slice, filter, mask, cast and fill null run on the values and wrap the result in a Widen again, with the nullability of the result. A cast to a wider integer target of the same signedness keeps the values at their own type, so a cast to a non-nullable target still fails on a null. A cast to a float or to the other signedness, and a fill value outside the range of the values, fall back to canonical execution. Take is an execution kernel for the Dict parent. It executes the gather of the values, so that no lazy Dict sits under a Widen, where the DuckDB exporter and the dict layout writer would not find it. The compare kernel compares two Widen operands at the wider value type, and a constant inside the range of the values with the values. A constant outside that range gives the same answer for every valid row, so the kernel returns that answer with the validity of the values, as a ConstantArray when there are no nulls. The compute conformance suites run on each shared fixture. Signed-off-by: Connor Tsui <[email protected]>

A check of the briefs against NR1 (#10864) found nine mismatches. The NR3 kernels need a pub(crate) compute module, and the decode function moves to array.rs, because the vtable modules are private. The fastlanes test directory does not exist yet. There is no ZipFalseReduce trait, so NR4 writes that rule by hand, and it reuses integer_value instead of copying it. Narrow is not registered in ArraySession. The other fixes update moved line numbers and the NR2 session recipe. Signed-off-by: Connor Tsui <[email protected]>

The cast conformance suite casts an integer array only to integer types, so this test is the only one that covers the cast rule declining a float target, including the integer guard for unsigned values. It keeps one signed and one unsigned case. Signed-off-by: Connor Tsui <[email protected]>

robert3005

This should be called Widen and should be a ScalarFn

connortsui20 · 2026-10-27T17:01:05Z

once someone fixes everything in here we can make this a scalar fn too: #9156

connortsui20 added the changelog/feature A new feature label Oct 27, 2026

connortsui20 mentioned this pull request Oct 27, 2026

Tracking Issue: Add Widen encoding for integers and decimals #10733

Open

12 of 23 tasks

connortsui20 force-pushed the ct/narrow-in-memory branch from 0f4366f to 479232c Compare October 27, 2026 14:51

connortsui20 commented Oct 27, 2026

View reviewed changes

connortsui20 force-pushed the ct/narrow-in-memory branch from 479232c to 76a4cce Compare October 27, 2026 15:14

connortsui20 commented Oct 27, 2026

View reviewed changes

connortsui20 added 2 commits October 27, 2026 11:19

connortsui20 force-pushed the ct/narrow-in-memory branch from 76a4cce to 81f227f Compare October 27, 2026 15:20

connortsui20 requested review from joseph-isaacs, mhk197 and robert3005 October 27, 2026 15:38

connortsui20 marked this pull request as ready for review October 27, 2026 15:38

robert3005 reviewed Oct 27, 2026

View reviewed changes

This was referenced Oct 27, 2026

Add compute benchmarks for Widen #10888

Merged

Run Widen arithmetic, cast, fill null, zip and between on the values #10891

Merged

connortsui20 added this pull request to stack #10892 October 27, 2026 17:49

connortsui20 enabled stacked auto-merge (squash) October 28, 2026 14:42

stacked auto-merge was automatically disabled October 28, 2026 14:43
Replaced by a newer auto-merge request

connortsui20 enabled stacked auto-merge (squash) October 28, 2026 14:43

connortsui20 requested a review from robert3005 October 28, 2026 14:45

connortsui20 force-pushed the ct/narrow-in-memory branch from fcb0046 to 75c7f4a Compare October 28, 2026 15:39

connortsui20 changed the title ~~Add an in-memory Narrow encoding for integers~~ Add an in-memory Widen encoding for integers Oct 28, 2026

robert3005 approved these changes Oct 28, 2026

View reviewed changes

connortsui20 force-pushed the ct/narrow-in-memory branch from 75c7f4a to a56a48e Compare October 28, 2026 15:54

stacked auto-merge was automatically disabled October 28, 2026 16:16
Auto-merge was disabled for another pull request in the stack

connortsui20 merged commit 6005d19 into develop Oct 28, 2026
110 of 112 checks passed

connortsui20 deleted the ct/narrow-in-memory branch October 28, 2026 16:17

Navigation Menu

Repository navigation

Add an in-memory `Widen` encoding for integers - #10864

Add an in-memory `Widen` encoding for integers#10864
connortsui20 merged 3 commits into
developfrom
ct/narrow-in-memory

connortsui20 commented Oct 27, 2026 •

edited

Loading

Uh oh!

codspeed Bot commented Oct 27, 2026 •

edited

Loading

Uh oh!

connortsui20 Oct 27, 2026

Uh oh!

connortsui20 Oct 27, 2026

Uh oh!

robert3005 left a comment

Uh oh!

connortsui20 commented Oct 27, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		//! Take executes its gather instead of leaving a lazy [`Dict`] under a `Narrow`, because code such
		//! as the DuckDB exporter and the dict layout writer looks for a `Dict` at the root of an array.

Conversation

connortsui20 commented Oct 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Changes

Uh oh!

codspeed Bot commented Oct 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Merging this PR will regress 4 benchmarks

Performance Changes

Footnotes

Uh oh!

connortsui20 Oct 27, 2026

Choose a reason for hiding this comment

Uh oh!

connortsui20 Oct 27, 2026

Choose a reason for hiding this comment

Uh oh!

robert3005 left a comment

Choose a reason for hiding this comment

Uh oh!

connortsui20 commented Oct 27, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

connortsui20 commented Oct 27, 2026 •

edited

Loading

codspeed Bot commented Oct 27, 2026 •

edited

Loading