This file is generated from data/selected_papers.json.
| Theme | Companies / papers | Repository database objects | Decision rule |
|---|---|---|---|
| Unified optimizer, vector, lineage, and SQL visibility | Alibaba: Learned Query Optimizer in Alibaba MaxCompute: Challenges, Analysis, and Solutions, LindormVector: A Distributed Vector Engine on a Cloud-Native Multi-Model NoSQL Database, SQLens: Continuous Code-to-SQL Visibility in the Wild, NebulaSQL: A Large-scale Feature Computation System for Online Recommendation | query_runs, lineage_edges, vector_index_entries, code_to_sql_edges, optimizer_observations |
Connect execution, lineage, and retrieval. |
| Repository as production-scale graph and multimodal warehouse | ByteDance: ByteHouse: ByteDance’s Cloud-Native Data Warehouse for Real-Time Multimodal Data Analytics, ByteGraph-Dione: An Adaptive Dual-Format Graph Engine with Hotspot Awareness and Transaction Efficiency for Production-Scale Workloads, G2+D: A High Performance Distributed Graph Mining System | repo_nodes, graph_edges, multimodal_chunks, execution_traces, embedding_runs |
Model repo knowledge as a typed graph. |
| Generated artifacts as materialized views | Databricks: Enzyme: Incremental View Maintenance for Data Engineering, Enzyme: Incremental View Maintenance With Spark Declarative Pipelines | generated_artifacts, artifact_dependencies, refresh_plans, refresh_runs, stale_flags |
Generated files are materialized views. |
| Long-lived key-value / wide-column metadata design | Google: Twenty Years of Bigtable, Twenty Years of Bigtable | files, file_versions, path_history, content_hashes, metadata_cells |
Paths change; stable IDs and hashes endure. |
| LLM recommendations must be validated by database tooling | Microsoft: Evaluating the Practical Effectiveness of LLM-Driven Index Tuning with Microsoft Database Tuning Advisor, Bitmap Filtering in the Fabric Data Warehouse, CoddSpeed: Hardware Accelerated Query Processing in Microsoft Fabric, ConDABench: Interactive Evaluation of Language Models for Data Analysis | index_advice, explain_plans, benchmark_runs, workload_replays, validation_results |
Models propose; benchmarks decide. |
| Normalize documents, JSON, Markdown, and YAML into relational projections | Oracle: From JSON to Duality: Automated Application Migration from Document to Relational Databases, From JSON to Duality: Automated Application Migration from Document to Relational Databases | docs, json_objects, schemas, relational_projection, document_projection_map |
Keep raw documents and relational projections. |
| AI operators inside the database optimizer | Snowflake: Streaming Model Cascades for Semantic SQL, Cortex AISQL: A Production SQL Engine for Unstructured Data | ai_runs, model_costs, model_cascade_stages, semantic_operator_results, verification_events |
Store every AI operation as queryable data. |
| PostgreSQL-compatible horizontal scale | AWS: Aurora PostgreSQL Limitless Database: Building a Highly Scalable OLTP Database, Aurora PostgreSQL Limitless Database: Building a Highly Scalable OLTP Database | router_rules, project_shards, replicas, global_transaction_log, shard_rebalancing_events |
Preserve relational compatibility before custom sharding. |
| End-to-end benchmarking for RAG and embeddings | IBM: RAGPerf: An End-to-End Benchmarking Framework for Retrieval-Augmented Generation Systems, Fabric-X: Scaling Hyperledger Fabric for Asset Exchange | rag_eval_runs, retrieval_metrics, generation_metrics, rerank_metrics, factuality_checks |
Measure quality and system cost together. |
| Use graph databases when multi-hop queries become the bottleneck | Neo4j: RIOT: Replicated Independently-Ordered Transactions, Neo4j Research: RIOT, RIOT: Replicated Independently-Ordered Transactions | graph_edges, graph_snapshots, neo4j_mirror_state, multi_hop_query_log |
Add graph DB after multi-hop workload evidence. |
| Distributed consensus for highly available metadata | Cockroach Labs: Scalable Leader Leases For Multi Consensus Groups in CockroachDB, Scalable Leader Leases For Multi Consensus Groups in CockroachDB | metadata_ranges, leaseholders, consensus_groups, range_health_events, lease_transfer_events |
Use distributed SQL for HA metadata services. |
| Distributed design for heterogeneous multi-table workloads | Tencent: TDSQL-Boundless: A Distributed Database System for Large-scale Heterogeneous Multi-Table Workloads, TDSQL-Boundless: A Distributed Database System for Large-scale Heterogeneous Multi-Table Workloads | project_shards, package_groups, heterogeneous_tables, workload_classes, cross_shard_queries |
Classify heterogeneous workloads early. |
- Google-style stable file identity.
- Databricks-style generated artifact freshness.
- Snowflake/Microsoft-style AI logging and validation.
- Alibaba/ByteDance-style graph and lineage tables.
- Oracle-style document-to-relational projection.
- Neo4j mirror for graph traversal.
- AWS-style router/shard layer.
- Cockroach-style consensus metadata ranges.
- Tencent-style heterogeneous distributed routing.
- IBM/RAGPerf-style RAG benchmark suite.
- Microsoft-style optimizer/benchmark gates for LLM proposals.