技能 数据科学 数据系统科学可复现性指南

数据系统科学可复现性指南

v20260724
icde-reproducibility
本指南为数据工程领域的学术论文提供了严格的可复现性标准。它指导作者如何精确记录硬件环境、数据来源、软件版本和实验参数,确保论文中的所有性能指标和结论都是可被同行验证和重现的,是高质量学术投稿的必备环节。
获取技能
448 次下载
概览

ICDE Reproducibility

Use this before submission and again before camera-ready. ICDE authors are expected to submit supplemental material, and its availability is weighed in the evaluation — so reproducibility is not optional polish, it is scored evidence.

Evidence map

  • Map each performance claim to a verifiable location: a figure regenerated from logged runs, a workload script, or a documented measurement in the supplement.
  • Pin the environment: CPU, memory, storage device (the NVMe/SSD/HDD distinction changes results), OS and kernel, database/library versions, compiler flags, and any cluster topology.
  • Pin the data: dataset provenance, construction steps, scale factors, and for synthetic data the generator with its seeds — a workload nobody can regenerate is not reproducible.
  • Pin the variance protocol: how many runs, warm-up handling, how outliers are treated, and whether reported bars are standard deviations, confidence intervals, or percentiles.
  • Document baseline tuning: the configuration and tuning budget given to each competitor. Reproducibility here means a reader can re-run the fair comparison, not just your system.
  • Trace figures to raw data: emit tables and plots from logged results so the PDF numbers and the supplement cannot drift apart.

Systems-reproducibility audit table

Dimension Weak answer ICDE-ready answer
Hardware "a modern server" Exact CPU, RAM, storage device model, and topology
Data "a large dataset" Named dataset or a seeded generator with scale factors
Variance one median number N runs with declared spread and warm-up policy
Baselines "we compared to X" X's config and tuning budget, re-runnable
Figures hand-entered numbers Plots emitted from logged runs by a script

Degrees of reproducibility

  • Turnkey: run_all.sh regenerates every figure from logged seeds on a documented machine; run_small.sh gives a fast subset for a reviewer with limited hardware.
  • Scripted: scripts exist but need documented manual steps or restricted-data access.
  • Descriptive: prose detailed enough that a competent engineer could rebuild the pipeline.

For ICDE, aim for turnkey on the synthetic experiments — a reviewer will re-run a generator far sooner than they will provision a cluster — and scripted for large real-data or proprietary-hardware runs, with deviations documented. State the level you actually achieved; overpromising turnkey behavior that fails on a clean machine is worse than an honest "scripted."

Vignette: a throughput-plus-latency paper

A submission claims higher ingestion at bounded read-latency cost. Its reproducibility spine: the storage device and queue-depth settings, the workload generator with append-to-scan parameters and seeds, the run count and percentile policy for the latency tails, the baseline LSM's compaction configuration, and a run_small.sh that reproduces the headline crossover on a single machine in minutes — plus one honest sentence on any result that needs the full cluster.

Single-blind note

  • ICDE supplemental material need not be anonymized — names may stay on the repository and in commit history. Spend the saved effort on making the package actually run, not on scrubbing identity a double-blind venue would demand.

Output format

[Claim inventory] <claim -> evidence location>
[Environment pinned] complete / partial / missing <what>
[Variance protocol] <runs / spread type / warm-up>
[Baseline fairness] <tuning budgets documented? y/n>
[Reproduction level] turnkey / scripted / descriptive
[Fixes before submission] <ordered list>
信息
Category 数据科学
Name icde-reproducibility
版本 v20260724
大小 4.02KB
更新时间 2026-07-28
语言