LianeMarilin/swe-marathon-datafusion-sample

DataFusion: Incremental Spilling FULL OUTER JOIN

A software-engineering benchmark sample focused on bounded-memory, spill-backed, incremental query execution in Apache DataFusion. The task combines asynchronous execution, exact join semantics, resource management, and compatibility with an existing database engine.

Status: locally validated. This is a SWE-Marathon-style sample, not an accepted official SWE-Marathon entry. Local technical checks do not establish compliance with every official quality requirement.

This repository publishes the complete reference solution, verifier, and execution evidence. It supports review and reproduction. Runs exposed to these answers must not be presented as uncontaminated blind evaluations. Author and contact fields remain blank; the upload account is not a verified authorship attribution.

Downloads

• Complete sample ZIP: 197,901,242 bytes, approximately 198 MB
• SHA256 checksums
• Original archive integrity report
• Post-extraction report for the 18 original static checks

The ZIP is byte-for-byte identical to the frozen September 12, 2026 delivery. Publication and dataset-card edits do not repackage the archive or modify its evidence.

cd2f7c7522f36ecd95a21af589daa2c6241bc29e06de46922eb04b59e2e7d051  swe-marathon-datafusion.zip

Original logs and reports retain development-machine paths, run identifiers, and upstream contributor information for traceability. Local absolute paths are not download links and may require environment-specific adjustment during reproduction. This is a task archive, not a tabular dataset directly loadable with load_dataset.

Dataset Scope

Property	Description
Unit of data	One software-engineering task, not 169 tasks or training examples
Acceptance cases	169: 100 fixed-seed randomized small cases, 48 compatibility cases, and 21 boundary, resource, and lifecycle cases; security preflight is an additional gate
Dataset splits	No train, validation, or test corpus split is defined; the task's tests/ directory contains its program verifier
Implementation languages	Rust implementation; Python, C, and Shell verification and environment tooling
Documentation languages	This dataset card and the task instructions are in English; the unchanged archive also contains Chinese delivery documentation
Source and inputs	A pinned upstream codebase with deterministic synthetic evaluation inputs
Release dates	Archive frozen on September 12, 2026; first published on Hugging Face on September 14, 2026

The archive extracts into swe-marathon-datafusion/:

swe-marathon-datafusion/
├── README.md                 # Full delivery notes
├── PACKAGE_MANIFEST.json     # Member inventory and file hashes
├── tasks/                    # Task, starting environment, solution, and verifier
├── authoring/                # Standards, proposal, and freeze approval records
├── validation/               # Execution evidence, reviews, and failure history
├── scripts/                  # Freeze, environment, Harbor, and archive checks
└── licenses/                 # Upstream licenses and NOTICE files

Task Description

Extend the existing HashJoinExec in a pinned Apache DataFusion 54.0.0 codebase to support bounded-memory, spill-backed FULL OUTER JOIN execution within the specified domain. The implementation must continuously emit determined results before the probe input reaches EOF.

Key requirements include:

• A 64 MiB DataFusion memory pool and a 256 MiB process peak-RSS limit.
• Exact duplicate-row and NULL semantics.
• Four timed probe windows with output-progress requirements.
• Resource cleanup after errors and cancellation.
• Preservation of behavior covered by 48 public-API compatibility cases.

The complete domain, limits, and acceptance rules are specified by the archived task instructions and environment/CONTRACT.md.

The baseline is Apache DataFusion commit 45d943dfb8699dc9cb9ef2320e955b73e3e6c03b. Quality requirements were pinned to SWE-Marathon commit 8e264e9. The six proposal criteria, 29 implementation criteria, and acceptance rules were frozen before implementation of the reference solution and verifier. Internal design approval is not official approval.

Validation Evidence

Check	Recorded result
Formal Harbor NOP run	Reward 0; 169 cases executed, with 161 passing and 8 failing
Formal Harbor Oracle run	Reward 1; all 169 cases passing
Additional reference runs	Two full diagnostic runs, each passing 169/169 cases
Legal parameter variant	A 32-bucket, 512-row-output variant passing 169/169 cases; a parameter variant of the same algorithm, not an independent algorithm
Negative controls	Four incorrect semantic or progress behaviors rejected by the verifier
Original static checks	All 18 checks from the pinned upstream version passed, including a rerun against task files extracted from the final ZIP
Archive integrity	23,560 files, including freeze records, source, environments, solution, verifier, raw logs, failure history, and review evidence

These are local validation results recorded on September 12, 2026. The two diagnostic runs and the formal Oracle run constitute three complete reference executions, not three independent model-solving trials. Publication did not rerun a model or the complete benchmark. Remote file contents were checked against local SHA256 hashes after upload.

Formal scoring uses a binary reward: 1 only when all applicable acceptance gates pass, otherwise 0. partial_score is diagnostic. Case pass rates describe the reference implementation, not model performance.

The source used in the two diagnostic builds differs from the packaged implementation only by the required Canary comment. The formal Harbor run uses the delivered source; the 22 final verifier files are byte-identical across the final runs. Exact source and evidence bindings are documented in the original reports.

Reproduction

With the Hugging Face CLI installed, download and verify the archive in a dedicated working directory:

hf download LianeMarilin/swe-marathon-datafusion-sample swe-marathon-datafusion.zip SHA256SUMS --repo-type dataset --local-dir .
sha256sum -c SHA256SUMS
unzip swe-marathon-datafusion.zip
cd swe-marathon-datafusion
python3 scripts/verify_freeze.py

On macOS, replace the checksum command with shasum -a 256 -c SHA256SUMS. The freeze check verifies that files match the recorded specification; it does not re-establish runtime correctness.

After reading the archived README.md and satisfying the environment requirements below, optionally run the NOP and Oracle checks:

python3 scripts/validate_harbor.py --agent both

This command builds Docker images, executes the unmodified baseline and reference solution, and writes new validation evidence into the extracted directory. Dependency installation during image builds may require network access; evaluation runs disable network access. It does not invoke a solving model or a paid model service, and it is not a frontier-model blind evaluation.

Environment and Limitations

Reproduction requires Docker, Harbor 0.22.0, and a Linux kernel supporting Landlock ABI >= 3 and seccomp. The task is configured for 4 vCPU, 16 GiB RAM, and 20 GiB storage. The original validation host provided approximately 7.65 GiB of physical memory to Docker; a 16 GiB container limit does not establish the availability of 16 GiB of physical memory. The recorded aarch64/Linux results are not official-environment or cross-platform validation.

Human authorship or substantive human revision has not been verified. Author and contact information, frontier-model difficulty calibration, and official pre-review and acceptance remain unresolved. The 3-5 hour estimate is an internal design-review estimate of best-case completion by a proficient specialist, not a measured completion time or calibrated model difficulty.

The public archive retains its original provenance records. This English dataset card does not alter or supersede those records.

Intended Uses

Suitable uses include database execution-engine task-design research, inspection of verifier behavior and negative controls, reproduction of reference results, and Harbor integration testing.

This sample is not a confidential, uncontaminated evaluation set. It does not establish model rankings, long-horizon model capabilities, or performance on other engineering tasks. No independent model-solving trajectories or measured token-budget results are provided.

Sources and Licensing

The upstream Apache DataFusion source retains its Apache-2.0 license and NOTICE. Other included materials remain subject to the licenses in licenses/ and their respective upstream sources. This publication does not grant or assume a new uniform license for the entire archive and does not change existing licensing or contributor attribution.