selected_papers.csv

company,theme,adoptable_idea,why_it_matters,repo_db_translation,first_schema_objects,do_now,do_later,watch_out,priority,key_papers
Alibaba,"Unified optimizer, vector, lineage, and SQL visibility","Connect query execution, vector search, lineage, and code-to-SQL visibility into one metadata fabric.","A useful repository database should answer not only “what files exist?” but also “which code produced this SQL, which query produced this table, which vector index serves this search?”","Store `graph_edges`, `lineage_edges`, `query_runs`, `vector_index_entries`, and `code_to_sql_edges` as connected concepts.",query_runs; lineage_edges; vector_index_entries; code_to_sql_edges; optimizer_observations,"Log every generated SQL statement with source file, function/symbol, input artifact, execution plan, output artifact, and cost.",Use historical executions and repository context to rank candidate plans and queries.,Vector search without lineage becomes hard to debug; lineage without query runs becomes hard to trust.,P0,"Learned Query Optimizer in Alibaba MaxCompute: Challenges, Analysis, and Solutions <https://arxiv.org/abs/2602.07336>; LindormVector: A Distributed Vector Engine on a Cloud-Native Multi-Model NoSQL Database <https://2026.sigmod.org/sigmod_industry_papers.shtml>; SQLens: Continuous Code-to-SQL Visibility in the Wild <https://2026.sigmod.org/sigmod_industry_papers.shtml>; NebulaSQL: A Large-scale Feature Computation System for Online Recommendation <https://2026.sigmod.org/sigmod_industry_papers.shtml>"
ByteDance,Repository as production-scale graph and multimodal warehouse,"Represent code, docs, schemas, tests, proofs, embeddings, and execution traces as graph-connected multimodal objects.","Large repositories are not just folders. They are multimodal graphs: code calls code, docs describe APIs, tests verify behavior, schemas constrain data, embeddings support retrieval.","Model repository entities in `repo_nodes`, connect them through `graph_edges`, and store multimodal evidence in `multimodal_chunks`.",repo_nodes; graph_edges; multimodal_chunks; execution_traces; embedding_runs,"Create typed nodes and edges for files, symbols, imports, tests, docs, schemas, and embeddings.","Add graph-mining jobs for hotspot files, central APIs, dead edges, and risky dependency clusters.",A graph without stable IDs becomes noisy. Anchor nodes to stable file/symbol IDs and content hashes.,P0,ByteHouse: ByteDance’s Cloud-Native Data Warehouse for Real-Time Multimodal Data Analytics <https://arxiv.org/abs/2602.08226>; ByteGraph-Dione: An Adaptive Dual-Format Graph Engine with Hotspot Awareness and Transaction Efficiency for Production-Scale Workloads <https://2026.sigmod.org/sigmod_industry_papers.shtml>; G2+D: A High Performance Distributed Graph Mining System <https://2026.sigmod.org/sigmod_industry_papers.shtml>
Databricks,Generated artifacts as materialized views,"Treat generated files, summaries, diagrams, JSONL, TSV, and Mermaid outputs as materialized views over source repository state.","A repository with generated artifacts needs freshness, dependency tracking, and refresh plans. Materialized-view thinking gives a clean model for rebuilds.","`generated_artifacts`, `artifact_dependencies`, `refresh_plans`, and `stale_flags` should track source hashes, refresh plans, stale status, and incremental refresh metadata.",generated_artifacts; artifact_dependencies; refresh_plans; refresh_runs; stale_flags,Record source file hashes for every generated artifact and mark artifacts stale when dependencies change.,Add an incremental refresh planner that refreshes only affected artifacts rather than rebuilding everything.,"Without lineage, generated docs silently drift from the source repository.",P0,Enzyme: Incremental View Maintenance for Data Engineering <https://arxiv.org/abs/2603.27775>; Enzyme: Incremental View Maintenance With Spark Declarative Pipelines <https://2026.sigmod.org/sigmod_industry_papers.shtml>
Google,Long-lived key-value / wide-column metadata design,"Optimize metadata identity for decades: stable keys, versioned paths, immutable content hashes, and forward-compatible columns.","Repository metadata is historical by nature. Paths move, files rename, schemas evolve, and hashes remain the best anchor for content identity.","Use `file_id`, `path`, `sha256`, `version_id`, `valid_from`, `valid_to`, and `path_history` so metadata remains stable across renames and rebuilds.",files; file_versions; path_history; content_hashes; metadata_cells,Separate logical file identity from mutable path and immutable content hash.,"Partition metadata by repository, time, and content-hash prefix for large-scale scans.","Do not use path as the only primary key. Paths are names, not identities.",P0,Twenty Years of Bigtable <https://research.google/pubs/twenty-years-of-bigtable/>; Twenty Years of Bigtable <https://2026.sigmod.org/sigmod_industry_papers.shtml>
Microsoft,LLM recommendations must be validated by database tooling,"Use LLMs for proposal generation, but rely on EXPLAIN plans, replay, benchmarks, and optimizer tooling for acceptance.","LLMs can suggest useful index or schema changes, but production databases need repeatable validation because model output has variance and can regress workloads.","LLM-generated index/schema advice must pass `EXPLAIN`, benchmark runs, workload replay, and rollback checks before being merged.",index_advice; explain_plans; benchmark_runs; workload_replays; validation_results,Create a gate: LLM proposal -> generated migration -> EXPLAIN delta -> benchmark -> replay -> approval.,Train a local rule library from LLM-generated hypotheses that repeatedly pass validation.,"Never merge index advice directly from a model. Treat it as a candidate patch, not a decision.",P0,Evaluating the Practical Effectiveness of LLM-Driven Index Tuning with Microsoft Database Tuning Advisor <https://arxiv.org/abs/2603.09181>; Bitmap Filtering in the Fabric Data Warehouse <https://2026.sigmod.org/sigmod_industry_papers.shtml>; CoddSpeed: Hardware Accelerated Query Processing in Microsoft Fabric <https://2026.sigmod.org/sigmod_industry_papers.shtml>; ConDABench: Interactive Evaluation of Language Models for Data Analysis <https://2026.sigmod.org/sigmod_industry_papers.shtml>
Oracle,"Normalize documents, JSON, Markdown, and YAML into relational projections","Keep developer-friendly document formats, but project them into relational tables for querying, constraints, migration, and governance.","Markdown, JSON, and YAML are convenient for authors but weak for joins, constraints, and governance. Relational projection gives queryability without deleting the original documents.","Normalize raw documents into `docs`, parsed objects into `json_objects`, schema definitions into `schemas`, and dual views through `document_projection_map`.",docs; json_objects; schemas; relational_projection; document_projection_map,"Parse frontmatter, JSON, YAML, headings, links, and code blocks into typed relational projections.",Generate migration plans from loose documents to constrained schemas with reversible mappings.,"Do not throw away source documents. Store raw content, parsed structure, and normalized projection together.",P0,From JSON to Duality: Automated Application Migration from Document to Relational Databases <https://dl.acm.org/doi/10.1145/3788853.3803096>; From JSON to Duality: Automated Application Migration from Document to Relational Databases <https://2026.sigmod.org/sigmod_industry_papers.shtml>
Snowflake,AI operators inside the database optimizer,"Treat LLM/AI calls as first-class database operators with measurable cost, quality, confidence, and fallback behavior.","Semantic SQL makes AI inference part of query execution. That means the repository database should not hide prompts, models, costs, confidence, or verification state in logs only; they belong in normalized tables.","`ai_runs`, `model_costs`, `confidence`, `cascade_stage`, and `verified` should be stored alongside query and artifact metadata.",ai_runs; model_costs; model_cascade_stages; semantic_operator_results; verification_events,"Log every AI call with model, prompt hash, input row/object id, output hash, confidence, latency, token usage, price estimate, cascade stage, and verification status.","Let the optimizer choose between small/fast model, large/oracle model, cached output, and human review based on cost and confidence.",Do not trust a semantic operator just because it returns an answer. Persist uncertainty and verification state so downstream artifacts can be rebuilt or invalidated.,P0,Streaming Model Cascades for Semantic SQL <https://arxiv.org/abs/2604.00660>; Cortex AISQL: A Production SQL Engine for Unstructured Data <https://2026.sigmod.org/sigmod_industry_papers.shtml>
AWS,PostgreSQL-compatible horizontal scale,"Start with normal PostgreSQL semantics, then add router, shard, replica, and distributed transaction concepts only when scale demands it.",Most repo metadata projects should begin with one reliable relational database. Horizontal scaling should preserve compatibility instead of forcing application-level sharding too early.,"Begin with single PostgreSQL. When repositories or organizations grow, introduce logical routers, project shards, replicas, and distributed transaction logs.",router_rules; project_shards; replicas; global_transaction_log; shard_rebalancing_events,"Design tables with stable IDs and clear shard keys, even before sharding is implemented.","Route by organization, repository, package, or project boundary when metadata volume becomes large.",Avoid premature sharding. The first goal is correctness and compatibility; scale-out is a later migration path.,P1,Aurora PostgreSQL Limitless Database: Building a Highly Scalable OLTP Database <https://www.amazon.science/publications/aurora-postgresql-limitless-database-building-a-highly-scalable-oltp-database>; Aurora PostgreSQL Limitless Database: Building a Highly Scalable OLTP Database <https://2026.sigmod.org/sigmod_industry_papers.shtml>
IBM,End-to-end benchmarking for RAG and embeddings,"Measure RAG as a whole system: embedding, indexing, retrieval, reranking, generation, latency, memory, update cost, and factual consistency.",A repository knowledge base is useful only if retrieval quality and system cost are measured together. Recall alone is not enough.,"Persist `rag_eval_runs`, `context_recall`, `latency_ms`, `memory_mb`, `update_cost_ms`, `groundedness`, and `factual_consistency`.",rag_eval_runs; retrieval_metrics; generation_metrics; rerank_metrics; factuality_checks,"Benchmark at least five dimensions: recall, answer quality, latency, memory, and index update cost.","Run benchmark suites whenever embedding model, chunker, index, or reranker changes.",RAG quality can improve while latency or update cost becomes unacceptable. Track both quality and systems metrics.,P1,RAGPerf: An End-to-End Benchmarking Framework for Retrieval-Augmented Generation Systems <https://arxiv.org/abs/2603.10765>; Fabric-X: Scaling Hyperledger Fabric for Asset Exchange <https://2026.sigmod.org/sigmod_industry_papers.shtml>
Neo4j,Use graph databases when multi-hop queries become the bottleneck,"Start with `graph_edges` in PostgreSQL, then mirror to Neo4j when multi-hop traversal and graph algorithms dominate workload cost.","Repository graphs become valuable when questions require several hops: which tests cover a changed symbol, which docs mention it, which generated artifacts depend on it?","Use PostgreSQL `graph_edges` first. Add a Neo4j mirror only after multi-hop queries, path ranking, or graph algorithms justify it.",graph_edges; graph_snapshots; neo4j_mirror_state; multi_hop_query_log,"Store edges in a relational table with `src_id`, `dst_id`, `edge_type`, `source`, `valid_from`, and `confidence`.","Export edges to Neo4j for path queries, centrality, community detection, and impact analysis.",Adding a graph DB too early creates operational overhead. Make the graph model portable first.,P1,RIOT: Replicated Independently-Ordered Transactions <https://dl.acm.org/doi/10.1145/3788853.3803094>; Neo4j Research: RIOT <https://neo4j.com/research/>; RIOT: Replicated Independently-Ordered Transactions <https://2026.sigmod.org/sigmod_industry_papers.shtml>
Cockroach Labs,Distributed consensus for highly available metadata,"Leader leases and range-level consensus matter once repository metadata becomes a shared, high-availability service.","Metadata services can become critical infrastructure. Leaseholders, ranges, and consensus groups become relevant when many services read and update metadata concurrently.","Use `metadata_ranges`, `leaseholders`, `consensus_groups`, and range-health observations when evolving toward distributed metadata storage.",metadata_ranges; leaseholders; consensus_groups; range_health_events; lease_transfer_events,"Keep metadata boundaries clean so a future distributed database can partition them by repository, organization, package, or time.",Adopt CockroachDB-like distributed SQL only when availability and horizontal scale matter more than local simplicity.,"Consensus solves availability and coordination problems, not bad schema design. Fix identity, lineage, and validation first.",P2,Scalable Leader Leases For Multi Consensus Groups in CockroachDB <https://www.cockroachlabs.com/blog/distributed-database-leader-leases/>; Scalable Leader Leases For Multi Consensus Groups in CockroachDB <https://2026.sigmod.org/sigmod_industry_papers.shtml>
Tencent,Distributed design for heterogeneous multi-table workloads,"Plan for multiple languages, package managers, schema styles, artifact types, and workload classes in large monorepos.","Large monorepos look like heterogeneous database workloads: many schemas, many access patterns, many teams, and uneven hotspots.","Use `project_shards`, `package_groups`, `heterogeneous_tables`, and workload-class routing for multi-language monorepo metadata.",project_shards; package_groups; heterogeneous_tables; workload_classes; cross_shard_queries,"Tag every file, package, generated artifact, and query with project, language, package manager, and workload class.",Shard by project or package group when single-node metadata operations become a bottleneck.,Heterogeneity is a schema problem before it is a distributed-systems problem. Classify workload types early.,P2,TDSQL-Boundless: A Distributed Database System for Large-scale Heterogeneous Multi-Table Workloads <https://dl.acm.org/doi/abs/10.1145/3788853.3803090>; TDSQL-Boundless: A Distributed Database System for Large-scale Heterogeneous Multi-Table Workloads <https://2026.sigmod.org/sigmod_industry_papers.shtml>