Evolving the C++ engine.

Solid plots show measured full-command time at SF 0.1, 0.5 and 1. Dashed boxes show unrun scenarios for nominal 100 GB and 1 TB. Those targets are 100× and 1,000× beyond the largest measured scale.

18 timed commands + 1 explicit preflight · 13.12 seconds total · two fixed runs per query and scale. Shared-machine medians; no significance claim.

Nominal scale describes the source dataset, not bytes loaded. Each plan reads three projected tables. The memory reference is this 128 GiB machine’s total RAM; available memory is lower. A scenario range is a model envelope, not a confidence interval.

The engine is unchanged in this study. These selected native plans do not establish an official TPC-DS score, general SQL support, or a Photon / competitor speed advantage.

Strict streaming JSON writes directly into columns with lazy missing-cell padding; raw CSV dictionary spans avoid one temporary copy while scratch is immediately owned; exact case-sensitive literal replacement uses substring search. The CSV workflow is fixed, but its parser and replacement code change. Same SF 0.1 input bytes and full query graphs; previous winner preserved.

2 explicit correctness warmups + 16 timed native calls · 4.56 seconds total in this final screen. Two explicit Q3 correctness warmups at the final saved executables precede 16 fixed timing calls (two reversed-order pairs per case, repeat = 1). No retries, sample exclusions, pooling or threshold changes. All startup/read/decode/processor/output/cleanup costs stay inside every timed command. Outcome applies to repeated commands after validation, not first install/launch.

All 16 case/metric guards passed. CSV decode +0.26%, processor +2.55%, command +0.06% and memory -1.05% remain within the preset +5% bounds.

Decode excludes raw file reading and times source-to-Table construction. Processor time sums all native stages. Full command includes startup, reading, decode, processors, output, cleanup and existing runner overhead. Peak RSS is the whole process high-water memory footprint. Bars use baseline 100 on a common 0–125 scale. Median of two within-pair candidate/baseline ratios. Separate medians are descriptive and not used to calculate paired percentages. Preliminary shared-Mac screen; no confidence interval.

27 focused cases / 527 assertions passed. All 14 query outputs (including preflights) match full ordered exact-cent SQL references; all 4 timed CSV outputs agree. Prior 6 CSV + 7 Find/Replace goldens are retained for unchanged source-specific implementations.

Selected SF0.1 native plans on a shared Mac, not general SQL/DECIMAL support, an official TPC-DS score, a Photon/language comparison or formal qualification. Energy/power was not measured.

Reuse the existing streaming reader instead of building an intermediate JSON object tree. Same input files and query plans; the previous winner is preserved. The candidate below is not shipped.

16 native calls · 4.42 seconds total · two fixed baseline/candidate pairs per case, reversed order in pair two. Median within-pair time ratios; preliminary descriptive screen.

The CSV decoding control moved +10.41%, exceeding the preset +5% limit. We rolled back. This small screen does not establish why unchanged CSV code moved; query gains remain experimental.

Decode times only source-to-Table construction; raw file reads precede it. Processor time sums all native stage clocks. Full command includes startup, reads, decode, processors, output, cleanup and runner overhead. All boundaries are unchanged. Bars use baseline 100 on a common 0–125 index scale; raw medians above are descriptive and need not produce the paired percentage.

12 focused cases / 226 assertions pass. All 12 timed query outputs match full ordered exact-cent SQL answers; all 4 CSV control outputs agree. No full-suite qualification claimed. Production sources and both active executables exactly match the preserved baseline. Candidate source, patch, binaries and all evidence remain available.

Two fixed pairs per case, both retained. No warmups/retries/exclusions/cacheflush/resampling. Changes use median within-pair ratios, not ratios of separate medians. Unchanged control movement is not an adopted gain or proof of a causal regression. This is not an official TPC-DS score or a Photon comparison.

DuckDB v1.5.6 generated data, scale factor 0.1. All 288,464 sales rows, 73,049 date rows and 1,800 item rows are retained; only required columns are loaded. Hand-mapped native plans cover three related revenue queries.

Median of three runs per query. 9 query timing calls + 6 legacy control calls · 3.89 seconds total. Both clocks below share the same millisecond scale.

Processor time sums all native stage clocks. Input decode is separate. Command wall includes startup, input reading/decoding, processors, output and cleanup, plus runner overhead. Medians are separate measurements and are not additive.

All nine timed outputs match the SQL reference. 11 correctness calls passed, including a separate 101-row limit/tie check. The unchanged legacy control stayed within its 5% bound; its -2.38% movement is not an adopted gain.

The engine is unchanged. Named inputs extend the existing evaluator; native plans use exact integer cents with checked bounds. This is a small selected-query pilot, not general SQL/DECIMAL support, an official TPC-DS score, or a Photon comparison.

Every workflow starts with 1,000,000 input rows. Actual production Find/Replace steps feed numeric or text filtering and aggregation on the existing synthetic CSV. The previous date-string winner is preserved. This change removes copied data that would immediately be discarded.

Baseline: the accepted C++ build before this attempt. The clock sums every processor stage; CSV parsing, final serialization, startup and graph bookkeeping are excluded. 18 native calls · 3.90 seconds total · Median of three within-pair candidate/baseline processor-time ratios per case; preliminary descriptive screen.

Correctness: 48 focused native cases / 2,274 assertions passed. Frozen replay: 7 / 7, with 0 known mismatches and 0 database cases skipped.

This small screen is preliminary evidence for these C++ processor workloads. It is not formal qualification or a new comparison with another language or Photon. Unchanged control movement is not an adopted gain.

Replace repeated date-string rewrites with one append pass, including date/group header aliases. The current grouped numeric-builder winner is preserved as the baseline. Same small workloads and existing tooling; all prior evidence stays available.

Baseline: the accepted C++ build before this attempt. The transform clock includes grouping, date handling, numeric construction and final row gathering; CSV parsing and result serialization are excluded. 18 native calls · 2.40 seconds total · Median of three within-pair candidate/baseline processor-time ratios per case; preliminary descriptive screen.

Correctness: 26 focused native cases / 1,883 assertions passed. Frozen replay: 11 / 11, with 0 known mismatches and 0 database cases skipped.

This small screen is preliminary evidence for these C++ processor workloads. It is not formal qualification or a new comparison with another language or Photon. Unchanged control movement is not an adopted gain.

Grouped numeric result columns reuse existing sum buffers, then append once in original group order. Final reordering, date-string writes and collision fallbacks stay unchanged. Small cases use two input rows per group and 820 / 410 date partitions.

Baseline: the accepted C++ build before this attempt. The transform clock includes grouping, date handling, numeric construction and final row gathering; CSV parsing and result serialization are excluded. 18 native calls · 3.25 seconds total · Median of three within-pair processor-time ratios per case.

Correctness: 26 focused native cases / 1,871 assertions passed. Frozen replay: 11 / 11, with 0 known mismatches and 0 database cases skipped.

This small screen is preliminary evidence for these C++ processor workloads. It is not formal qualification or a new comparison with another language or Photon. Unchanged control movement is not an adopted gain.

Fresh running-sum result columns now reuse their private output storage and append each value once. Three fixed pairs per case use small repetitions of the existing dated example data.

Baseline: the accepted C++ build before this attempt. The transform clock includes date sorting, row gathering and running-sum construction; CSV parsing and result serialization are excluded. 18 native calls · 2.53 seconds total · Median of three within-pair processor-time ratios per case.

Correctness: 24 focused native cases / 1,649 assertions passed. Frozen replay: 11 / 11, with 0 known mismatches and 0 database cases skipped.

This small screen is preliminary evidence for these C++ processor workloads. It is not formal qualification or a new comparison with another language or Photon. Unchanged control movement is not an adopted gain.

Three workflows on the same million-row synthetic CSV using existing production processors. Current winner kept as the baseline; no retries or extra samples.

Baseline: the accepted C++ build before this attempt. The clock covers all processor stages and excludes CSV parsing and result serialization. 18 native calls · 10.59 seconds total · Median of three within-pair time ratios per workflow.

Correctness: 965 native cases / 362,691 assertions passed. Frozen replay: 634 / 641, with 7 known mismatches and 14 database cases skipped.

This small screen is preliminary evidence for these C++ processor workloads. It is not formal qualification or a new comparison with another language or Photon. Unchanged control movement is not an adopted gain.

Earlier sort attempts · Oct 27, 2026, 5:46 PM CDT · three paired runs · 1,000,000 rows

Earlier sort ideas, both rolled back.

Each attempt compares a proposed change with its earlier C++ baseline. Sorting is the target; filtering and summing are controls whose movement is not credited to the sort change. All four measurements are shown. Lower time is better; a small screen is preliminary evidence.

Both sort ideas were rolled back. Filter and sum used unchanged code. Their timing variations—including results previously shown in green—do not represent adopted improvements or evidence of combined “hybrid” gains.

Bars show paired time indices with each attempt’s baseline at 100. These short internal C++ checks are not formal qualification or a fresh comparison with another language or Photon.

1,000,000 rows; saved measurements on Apple M5 Max, 18 cores, 26.4.1

These are the last published language measurements, not a new run today. Core kernel code is unchanged; recent measured improvements are in the Cast workflows below. Saved other-language time divided by saved C++ time. A ratio above 1 favors C++. It is not a new win or a universal language ranking.

Accepted C++ improvements

What improved since then.

Each result uses its own prior C++ baseline, workload and clock. Percentages are separate studies and must not be multiplied or applied to the language ratios. All studies ran on a shared Mac with loopback workers.

The command-completion result removes idle shutdown waits; it does not measure faster filter, sort or sum kernels. These are internal C++ studies, with each study’s machine and execution scope in its evidence.

Against the field, 1,000,000 rows

Against the field, 10,000,000 rows

Against the field, 100,000,000 rows

Against the ports: Go, Rust, Swift, Zig

The four ports were timed together on one day (2026-10-12); only the cold load was run twice, so it is the one kernel of theirs with a line. Measured once and so not drawn: load, filter, count, sort, as a permutation, sort into a new table, sum by region — the Evals tab has those. The load in process is neo's against the Rust port's, the one port bench_load_profile.sh runs, which it has done since 2026-10-10.

On public data, read as text