EDBT Reproducibility
Use this before submission and again before camera-ready. EDBT's community has a
reproducibility-forward culture, and the published record is open access on OpenProceedings —
so an inspectable, re-runnable package raises a paper's standing and, for an Experiments &
Analysis paper, is the contribution. The goal is that a competent reader could rebuild your
measurements and reach your conclusions.
Evidence map
- Map each claim, mechanism, and reported number to a verifiable location — a paper section, a
table generated from a logged run, or a script in the artifact.
- For a mechanism, give enough of the algorithm, data structures, parameters, and system integration
that a reader could reimplement or rebuild it.
- For an evaluation, report workloads and their derivation, dataset versions and sources, the
measurement harness, metrics, and the analysis scripts.
- Keep the availability statement truthful and specific: what is shared, the workloads and data,
the hardware assumptions, and — if something cannot be shared — exactly why.
- Keep the paper and the artifact consistent: a number in the PDF that no script produces is the
contradiction reviewers read as carelessness.
Availability statement audit
| Claim in the paper |
Weak availability answer |
EDBT-ready answer |
| "We evaluate on workload W" |
"Data available on request" |
Archived workload/query-log derivation + the extracted data or a documented access path |
| "Our operator lowers latency" |
"Code will be released" |
Runnable system/prototype with a build, a demo run, and the config |
| "We compare N systems" (E&A) |
Numbers with no harness |
The full comparison harness that regenerates every table |
| "On a 128-node cluster" |
Nothing about environment |
Hardware/cluster spec, engine build/commit, and how to size a smaller reproduction |
"Available on request" is treated as not available; convert every such line into a concrete
package or an explicit, justified exception (licensing, confidentiality).
Provenance pinning (database-systems flavor)
[Data] pin dataset versions and sources; archive the derived workload/query-log, not just a
description; document filtering and sampling
[System] record the engine/prototype build or commit; ship a build recipe or container
[Environment] state hardware, memory, network, and node counts; note what a smaller reproduction changes
[Harness] the measurement scripts that produce each table/figure, with fixed configuration
[Randomness] log seeds for any stochastic step; say what is and is not deterministic
Degrees of reproducibility (state the one you achieved)
- Turnkey: one documented command (or container) regenerates each table/figure from a run or
from logged results.
- Scripted: scripts exist but require documented manual steps, a specific cluster, or external
data access.
- Descriptive: prose detailed enough that a competent reader could rebuild the pipeline.
For EDBT, aim turnkey for anything a reviewer might re-run quickly (a demo run on a small workload, a
plot from logged results); large-cluster or licensed-data experiments may stay scripted with the
environment and access clearly documented. Stating the achieved level honestly beats promising
turnkey behavior that fails on a clean machine.
Vignette: a distributed-operator study
Consider an operator evaluated on a cluster. Its reproducibility spine: a container or build recipe
for the engine plus the operator; the workload-derivation scripts with pinned dataset versions; the
measurement harness that runs the operator and the tuned baseline across node counts; the logged raw
results; and the analysis notebooks that turn them into the paper's tables — plus one honest sentence
about the parts (the full 128-node run, a licensed dataset) that a reader reproduces at reduced scale
and why.
Consistency and camera-ready pass
- Before submission: every reported number traces to the artifact; the availability statement
matches reality; if the cycle is double-blind, the artifact carries no identity strings.
- Before camera-ready: deposit the package in a DOI-issuing archive (Zenodo, figshare, Software
Heritage) with an OSI-approved license, replace any anonymized links with the permanent ones, and
make the statement consistent with the open-access OpenProceedings record
(
edbt-artifact-evaluation, edbt-camera-ready).
Output format
[Claim inventory] <claim -> evidence location>
[Availability] concrete / vague / missing
[Provenance gaps] <dataset versions / engine build / environment / seeds>
[Reproducibility level] turnkey / scripted / descriptive, stated honestly
[Paper fixes] <must appear in the PDF>
[Artifact fixes] <additions before upload>
1---2name: edbt-reproducibility3description: Use when strengthening EDBT reproducibility for a database-systems paper, covering a runnable artifact, pinned environments and workloads, dataset and query-log provenance, claim-to-evidence mapping, honest degrees of reproducibility, and consistency between the paper and the package for the open-access OpenProceedings record.4---56# EDBT Reproducibility78Use this before submission and again before camera-ready. EDBT's community has a9reproducibility-forward culture, and the published record is **open access on OpenProceedings** —10so an inspectable, re-runnable package raises a paper's standing and, for an **Experiments &11Analysis** paper, *is* the contribution. The goal is that a competent reader could rebuild your12measurements and reach your conclusions.1314## Evidence map1516- Map each claim, mechanism, and reported number to a **verifiable location** — a paper section, a17 table generated from a logged run, or a script in the artifact.18- For a mechanism, give enough of the algorithm, data structures, parameters, and system integration19 that a reader could reimplement or rebuild it.20- For an evaluation, report workloads and their derivation, dataset versions and sources, the21 measurement harness, metrics, and the analysis scripts.22- Keep the **availability statement** truthful and specific: what is shared, the workloads and data,23 the hardware assumptions, and — if something cannot be shared — exactly why.24- Keep the paper and the artifact **consistent**: a number in the PDF that no script produces is the25 contradiction reviewers read as carelessness.2627## Availability statement audit2829| Claim in the paper | Weak availability answer | EDBT-ready answer |30|---|---|---|31| "We evaluate on workload W" | "Data available on request" | Archived workload/query-log derivation + the extracted data or a documented access path |32| "Our operator lowers latency" | "Code will be released" | Runnable system/prototype with a build, a demo run, and the config |33| "We compare N systems" (E&A) | Numbers with no harness | The full comparison harness that regenerates every table |34| "On a 128-node cluster" | Nothing about environment | Hardware/cluster spec, engine build/commit, and how to size a smaller reproduction |3536"Available on request" is treated as *not available*; convert every such line into a concrete37package or an explicit, justified exception (licensing, confidentiality).3839## Provenance pinning (database-systems flavor)4041```text42[Data] pin dataset versions and sources; archive the derived workload/query-log, not just a43 description; document filtering and sampling44[System] record the engine/prototype build or commit; ship a build recipe or container45[Environment] state hardware, memory, network, and node counts; note what a smaller reproduction changes46[Harness] the measurement scripts that produce each table/figure, with fixed configuration47[Randomness] log seeds for any stochastic step; say what is and is not deterministic48```4950## Degrees of reproducibility (state the one you achieved)5152- **Turnkey:** one documented command (or container) regenerates each table/figure from a run or53 from logged results.54- **Scripted:** scripts exist but require documented manual steps, a specific cluster, or external55 data access.56- **Descriptive:** prose detailed enough that a competent reader could rebuild the pipeline.5758For EDBT, aim turnkey for anything a reviewer might re-run quickly (a demo run on a small workload, a59plot from logged results); large-cluster or licensed-data experiments may stay scripted with the60environment and access clearly documented. Stating the achieved level honestly beats promising61turnkey behavior that fails on a clean machine.6263## Vignette: a distributed-operator study6465Consider an operator evaluated on a cluster. Its reproducibility spine: a container or build recipe66for the engine plus the operator; the workload-derivation scripts with pinned dataset versions; the67measurement harness that runs the operator and the tuned baseline across node counts; the logged raw68results; and the analysis notebooks that turn them into the paper's tables — plus one honest sentence69about the parts (the full 128-node run, a licensed dataset) that a reader reproduces at reduced scale70and why.7172## Consistency and camera-ready pass7374- Before submission: every reported number traces to the artifact; the availability statement75 matches reality; if the cycle is double-blind, the artifact carries no identity strings.76- Before camera-ready: deposit the package in a **DOI-issuing archive** (Zenodo, figshare, Software77 Heritage) with an OSI-approved license, replace any anonymized links with the permanent ones, and78 make the statement consistent with the open-access OpenProceedings record79 (`edbt-artifact-evaluation`, `edbt-camera-ready`).8081## Output format8283```text84[Claim inventory] <claim -> evidence location>85[Availability] concrete / vague / missing86[Provenance gaps] <dataset versions / engine build / environment / seeds>87[Reproducibility level] turnkey / scripted / descriptive, stated honestly88[Paper fixes] <must appear in the PDF>89[Artifact fixes] <additions before upload>90```
brycewang-stanford