Use concat_elements_dyn from arrow-rs

[pepijnve]

Which issue does this PR close?

• Closes Replace custom ByteView concat kernel with the implementation from arrow-rs #23210

Rationale for this change

apache/arrow-rs#9876 added ByteView and FixedSizeBinary support to concat_elements_dyn in arrow-rs. As a consequence the extended implementation in DataFusion can now be replaced by a call to the arrow-rs implementation.

What changes are included in this PR?

• Remove the kernels::concat_elements_utf8view and kernels::concat_elements_binary_view_array
• Replace implementation of binary::concat_elements with a call to arrow::compute::kernels::concat_elements::concat_elements_dyn

Are these changes tested?

• Usage is covered by existing tests
• The kernels themselves are tested in arrow-rs

Are there any user-facing changes?

No, two pub functions have been removed from kernel, but kernel itself is not pub.

[pepijnve]

@alamb I kept this as draft for now, but while reviewing the changes I noticed the implementations from DataFusion are almost identical to the versions in the current release of arrow-rs. Could you run the string_concat benchmark? If there is no performance difference, perhaps this can already be merged.

[pepijnve]

On further inspection it turns out that apache/arrow-rs#9876 added the missing parts to concat_elements_dyn that DataFusion had. I've broadened this PR a bit to remove the custom concat_elements_dyn entirely instead. The one in arrow is actually more capable at this point since it also supports FixedBinary.

[alamb]

DataFusion had. I've broadened this PR a bit to remove the custom concat_elements_dyn entirely instead. The one in arrow is actually more capable at this point since it also supports FixedBinary.

The plan is working!

[alamb]

run benchmarks string_concat

[adriangbot]

🤖 Benchmark running (GKE) | trigger
Instance: c4a-highmem-16 (12 vCPU / 65 GiB) | Linux bench-c4812851016-718-42bd4 6.12.85+ #1 SMP Mon May 11 08:17:35 UTC 2026 aarch64 GNU/Linux

CPU Details (lscpu)

Architecture:                            aarch64
CPU op-mode(s):                          64-bit
Byte Order:                              Little Endian
CPU(s):                                  16
On-line CPU(s) list:                     0-15
Vendor ID:                               ARM
Model name:                              Neoverse-V2
Model:                                   1
Thread(s) per core:                      1
Core(s) per cluster:                     16
Socket(s):                               -
Cluster(s):                              1
Stepping:                                r0p1
BogoMIPS:                                2000.00
Flags:                                   fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh rng bti
L1d cache:                               1 MiB (16 instances)
L1i cache:                               1 MiB (16 instances)
L2 cache:                                32 MiB (16 instances)
L3 cache:                                80 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; CSV2, BHB
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

Comparing string_concat (b1d3f05) to 1fd29c9 (merge-base) diff using: string_concat
Results will be posted here when complete

---

File an issue against this benchmark runner

[alamb]

(I will be pretty stoked if we get more features and faster performance by deleting code 😎 )

[adriangbot]

🤖 Benchmark completed (GKE) | trigger

Instance: c4a-highmem-16 (12 vCPU / 65 GiB)

CPU Details (lscpu)

Architecture:                            aarch64
CPU op-mode(s):                          64-bit
Byte Order:                              Little Endian
CPU(s):                                  16
On-line CPU(s) list:                     0-15
Vendor ID:                               ARM
Model name:                              Neoverse-V2
Model:                                   1
Thread(s) per core:                      1
Core(s) per cluster:                     16
Socket(s):                               -
Cluster(s):                              1
Stepping:                                r0p1
BogoMIPS:                                2000.00
Flags:                                   fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh rng bti
L1d cache:                               1 MiB (16 instances)
L1i cache:                               1 MiB (16 instances)
L2 cache:                                32 MiB (16 instances)
L3 cache:                                80 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; CSV2, BHB
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

Details

group                              HEAD                                   string_concat
-----                              ----                                   -------------
concat_utf8view/concat/nulls_0     1.00     59.5±0.12µs        ? ?/sec    1.35     80.6±2.29µs        ? ?/sec
concat_utf8view/concat/nulls_10    1.00     67.2±0.10µs        ? ?/sec    1.30     87.5±0.38µs        ? ?/sec
concat_utf8view/concat/nulls_50    1.00     32.6±0.10µs        ? ?/sec    1.20     39.1±0.17µs        ? ?/sec

Resource Usage

string_concat — base (merge-base)

Metric	Value
Wall time	240.1s
Peak memory	25.9 MiB
Avg memory	2.3 MiB
CPU user	35.9s
CPU sys	0.0s
Peak spill	0 B

string_concat — branch

Metric	Value
Wall time	245.1s
Peak memory	29.1 MiB
Avg memory	2.7 MiB
CPU user	37.0s
CPU sys	0.0s
Peak spill	0 B

---

File an issue against this benchmark runner

[pepijnve]

more features

I was a little bit too optimistic there. While the arrow implementation can in theory do FixedBinary(n) || FixedBinary(m) => FixedBinary(n + m), there's a left.data_type() == right.data_type() guard that gets in the way of actually doing so. More PRs to prepare.

string_concat_coercion prevents hitting that code path at the moment by coercing the two FixedBinary types to variable length Binary.

[pepijnve]

Benchmark results do show a significant slowdown, so there must be something I overlooked. Let's wait for the arrow changes to land and then reevaluate.

[pepijnve]

I did the experiment of running the benchmark with the concat_elements_dyn implementation from arrow-rs main. That shows the speedup we were aiming for rather than a regression.

[pepijnve]

#21883 introduced string_concat operand coercion specifically for FixedSizeBinary to Binary. This causes DataFusion to not use the FixedSizeBinary code path that's now enabled by apache/arrow-rs#10222. Is there a reason to keep this or would it be preferable to make || return FixedSizeBinary(l + r)?