A Hybrid Dependency-Aware Framework for Task Decomposition and Dynamic Agent Generation in Oracle-to-PostgreSQL Migration
Oleg Grynets, Oleg Kaskun, Alona Seletska, Daryna Tukalo, Vasyl Lyashkevych
cs.LO · cs.AI · cs.MA · cs.SE
Quick verdict
Honest framework paper: solid deterministic decomposition results, but titular dynamic-agent layer unvalidated and the decisive ablation remains unexecuted.
Summary
The paper proposes a hybrid, deterministic-first framework for Oracle-to-PostgreSQL migration: ANTLR-based parsing decomposes source artifacts into typed migration units, a global symbol table resolves cross-file dependencies into a graph, Tarjan SCC condensation handles cycles, and the resulting DAG phases feed a proposed Dynamic Agent Factory that compiles task specifications into runtime agent configurations. On a proprietary 116-file Oracle corpus the deterministic pipeline yields 1,037 units with zero coverage gaps and 1,271 AST-derived dependencies; a selective LLM fallback over 165 parse-error units adds 496 validated dependencies, removes all dependency-blind units, and raises resolved internal edges from 446 to 527. No cycles occurred in the real corpus, so cycle handling is validated only on a synthetic fixture. A complementary, separately sourced experiment on 1,006 PL/SQL files reports 623 regenerated scripts (~62%), of which 380 (~61%) executed in PostgreSQL 16 (37.8% end-to-end); tables reached ~85% regeneration while no queries succeeded under the specification-mediated baseline. The paper's honesty is notable — it explicitly states the C0–C3 ablation designed to test whether dependency-aware dynamic agents outperform monolithic or static decomposition has not been executed, so the titular dynamic-agent contribution remains a design. Key limitations: a private, non-redistributable corpus, single-run LLM results without statistical treatment, unverified quality of LLM-recovered edges, and a decisive experiment that exists only as a protocol.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A deterministic-first hybrid pipeline decomposes 116 real Oracle files into 1,037 typed units with zero coverage gaps and 1,271 AST-derived dependencies.
- A selective LLM fallback restricted to 165 parse-error units adds 496 validated dependencies, eliminates all dependency-blind units, and raises resolved internal edges from 446 to 527 while leaving the AST contribution unchanged.
- Tarjan SCC condensation produces a dependency-respecting execution DAG with four phases; cycle handling is validated on a controlled fixture because the real corpus contained no cycles.
- A complementary 1,006-file specification-first baseline regenerates 623 PL/SQL scripts (~62%), of which 380 (~61%) execute in PostgreSQL 16 (37.8% end-to-end), with tables ~85% and queries 0%.
- Task-conditioned dynamic agent generation is proposed as a routing policy with a designed C0–C3 ablation, but the paper explicitly does not claim it improves migration outcomes yet.
Strengths
Attribution-controlled fallback experiment
By freezing the AST path, Section VII-B / Table V cleanly attribute the +496 dependencies and the 165→0 elimination of dependency-blind units to the LLM fallback alone; AST dependencies stayed exactly 1,271 in both runs — rare experimental discipline in agentic-LLM work.
Honest claim boundaries
Abstract, Section VI and Section XVII explicitly state the C0–C3 comparison is unexecuted and that the 1,006-file baseline 'does not by itself prove that dynamic-agent routing is superior'; the paper also discloses the failed test (564/565) and the 623→380 execution attrition instead of hiding them.
Reproducibility engineering on the deterministic side
Extraction vs resolution separation
Useful negative finding
Weaknesses
Titular contribution never exercised
Unauditable evidence base
No statistical treatment
LLM edge quality unverified
Complexity labels underived
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a framework paper from an EPAM Systems team proposing a hybrid, deterministic-first architecture for Oracle-to-PostgreSQL migration (Sections III–V), accompanied by two empirical layers: a measured decomposition study on a proprietary 116-file Oracle corpus (Section VII) and a separately sourced, 'incorporated' migration-generation experiment on 1,006 PL/SQL files (Section XII). The claims are carefully scoped. The paper claims (a) lossless decomposition — 1,037 units, zero coverage gaps, 1,271 AST-derived dependencies; (b) a clean fallback effect — +496 LLM-recovered dependencies over 165 parse-error units (15.9% of units), unresolved units 165→0, resolved internal edges 446→527 (+18.2%); (c) a valid dependency-respecting DAG with four phases; and (d) an execution-level baseline where 623/1,006 scripts regenerated (~62%) and 380/623 executed in PostgreSQL 16 (~61%; 37.8% end-to-end funnel). Crucially, it explicitly disclaims the claim its title gestures at: whether dependency-aware dynamic agents beat monolithic or static decomposition is deferred to a designed-but-unexecuted C0–C3 ablation (Sections VI, XV).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does this paper show that multi-agent LLM systems beat a single agent for database migration?
No. The paper designs a C0–C3 ablation (monolithic vs decomposed vs dependency-aware vs dynamic agents) but explicitly states it has not been executed; dynamic-agent superiority remains unproven.
What concrete results does the paper report?
On a private 116-file Oracle corpus: 1,037 migration units with zero coverage gaps and 1,271 AST-derived dependencies; an LLM fallback on 165 parse-error units added 496 validated dependencies, cut dependency-blind units to zero, and raised resolved internal edges from 446 to 527. A separately sourced 1,006-file PL/SQL experiment regenerated 623 scripts (~62%), of which 380 ran successfully in PostgreSQL 16 (~61% of regenerated; 37.8% end-to-end).
Is the code or dataset available?
No public release is stated. The 116-file corpus is company-internal and cannot be redistributed, and the 1,006-file experiment is described as 'supplied'; only PlantUML figure sources are said to accompany the manuscript.
Why did zero query migrations succeed?
Under the evaluated specification-mediated pipeline, queries had no persistent target object to regenerate, and the authors argue a direct dialect-translation route would suit queries better than writing a specification first.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–10 are referenced with PlantUML sources 'supplied with the manuscript' but are not analyzable from the extracted text.
- Equations (1)–(9) are garbled by text extraction (task tuple, Gainedges, DOR, Efficiency); their meaning was reconstructed from surrounding prose.
- The 1,006-file experiment is 'incorporated/supplied' and its full provenance is not in the text; per-object counts beyond tables (~85%) and queries (0%) are unavailable.
- The paper itself states the C0–C3 comparison is unexecuted; judgments about the dynamic-agent layer assess design, not results.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.