NeorunBase Key Features
A PostgreSQL-compatible multi-modal Lakebase that handles relational, JSONB, vector database, full-text search, graph database, and spatial data in a single distributed ACID engine. No separate Pinecone, Elasticsearch, or Neo4j — all signals combine in one SQL statement, with multi-site DR, silent-fail safety, and topology-aware placement built in for enterprise operations.
PostgreSQL Wire Protocol
psql, JDBC, pgAdmin, LangChain, and pgvector clients connect with zero code changes. Move from PostgreSQL to a distributed engine without rewriting applications.
SQL That Never Quietly Lies
The dangerous failure in a distributed engine is not "cannot run this" - it is an answer that looks like data and is not the one you asked for. NeorunBase either produces the answer PostgreSQL would produce, or it raises. A WHERE or HAVING clause is never dropped (on an UPDATE or DELETE that would rewrite every row); an unknown column or function is an error, not NULL; SET c = c + 1 actually computes; COUNT(DISTINCT ...) merges the shards value sets instead of adding their counts. UNIQUE, CHECK and DEFAULT are enforced on write, not just accepted at CREATE TABLE. Two e2e suites ship with the source so you can verify the rule against your own cluster.
Distributed ACID + Self-Healing
Hash-based sharding with online rebalancing, ZooKeeper sticky leader election, automatic shard / disk recovery, and 2PC distributed transactions — all on a RocksDB-backed ACID core.
Multi-Modal in One SQL
Combine relational, vector, full-text, graph, and spatial data inside a single ACID SQL statement. Hard Filter (WHERE) + semantic/keyword hybrid retrieval + graph-aware re-ranking (structural — by graph proximity) expressed in one SELECT — no separate stores, no glue ETL.
Vector Database (pgvector compatible)
pgvector-compatible VECTOR type with distance operators (<->, <#>, <=>) and distributed HNSW ANN indexes. Rows, metadata, and embeddings live in the same ACID transaction — ready as a RAG retrieval backend out of the box.
Full-Text (Lucene BM25)
Lucene-backed BM25 inverted index with PostgreSQL-compatible FTS syntax (@@, ts_rank) and 35+ language analyzers, including Nori morphological analysis for Korean. Indexes are declared the way PostgreSQL declares them — USING gin works — and \di reports each index's method (btree / fts / hnsw), so an index that is not the kind you meant is visible rather than inferred from wrong results. As in PostgreSQL, @@ works without an index at all; the index only makes it fast, and the same query returns the same rows either way. Hybrid retrieval blending BM25 + vector ANN is a single HYBRID_SEARCH(...) TVF call.
Graph Database + Analytics
Multi-hop BFS (GRAPH_NEIGHBORS), PageRank · Personalized PageRank, and reachability fact-checks (GRAPH_PATH_EXISTS) exposed as SQL TVFs over edge tables. A CSR acceleration layer cuts single-hop latency by 100×+, and a single SELECT can re-rank hybrid-search results by graph proximity — the Graph RAG pattern in one query.
Geospatial (PostGIS compatible)
PostGIS-compatible spatial functions (ST_Distance, ST_Contains, …) with Z-order spatial indexing. Location-based services, GIS, and geo-aware RAG, all from one engine.
Iceberg CDC + Kafka Ingest
OLTP table changes auto-sync to Iceberg/Parquet — Ontul reads the same data via the Iceberg catalog, no separate ETL. Synced tables are partitioned by a bucket of the primary key, so the equality delete that supersedes a row always lands in the same partition as the row it replaces. Reads both Iceberg v2 and v3 (deletion vectors), with direct Kafka consumer integration plus REST bulk insert and MERGE / upsert. With Write-Audit-Publish (WAP), CTAS, MERGE, and CDC sync can land on a non-main branch for isolated audit, then publish to main via REST (/admin/api/iceberg/wap/publish).
Row-Level Iceberg DML (v2 position deletes, v3 deletion vectors)
Run INSERT, UPDATE, DELETE and MERGE INTO straight against an Iceberg table — no copy into a native table first, no external engine. UPDATE, DELETE and MERGE are all merge-on-read: matched rows are recorded as (file, position) and written as delete files, the data files are never rewritten, and the whole statement lands as one atomic RowDelta commit. MERGE deletes by position rather than by key, so it removes exactly the rows that matched, reaches a target row whatever partition it lives in, and works on a table that declares no identifier fields. The delete form follows the format version — a parquet position-delete file on v2, a Puffin deletion vector on v3. On v3 the spec allows one vector per data file, so an existing vector is read, merged, and the superseded delete file removed in the same commit. The output is verified cross-engine: Trino has to read back exactly what NeorunBase reports.
Iceberg Serving (LakeBase) — PK point/range index
Serve the open lakehouse (Iceberg) at low latency and high QPS without copying data. A primary-key predicate (= / IN / BETWEEN / range comparison) is answered by a per-snapshot pk → (file, row) index — reading only the matching rows instead of scanning files — at thousands of QPS and 1-2 ms, faster than a native RocksDB read for point lookups. The index is automatic (no DDL), snapshot-invalidated, persisted to local RocksDB (restored on restart), and append-incremental. Declare secondary indexes on non-PK columns with CREATE INDEX, and scope auto-indexing across thousands of tables with per-catalog index.patterns. External applications and dashboards connect directly over the PostgreSQL wire — a true serving layer.
Enterprise Security & Admin Console
AES-256-GCM envelope encryption (data, WAL, metadata), built-in KMS with key rotation, IAM/RBAC + STS, TLS for the pgwire endpoint. Column masking, column deny, and row filters are enforced on the Iceberg serving path too. Alongside ontul, IAM federation makes ontul the policy authority while NeorunBase pulls and enforces it for direct (JDBC) access, with ontul-token SSO. A React-based console + Prometheus metrics covers cluster, shards, IAM, Iceberg, and Kafka in one place.
Multi-Site Disaster Recovery (Site Replication)
Every change on the primary — INSERT, UPDATE, DELETE, even CREATE TABLE — streams to the standby site over WAL-level asynchronous replication. Vector (HNSW) and FTS (Lucene) indexes auto-reconcile. Run a minutes-RPO DR site without Debezium or Golden Gate.
NIC Silent-Fail Safety
Shuts down the trickiest distributed-systems failure mode — "ZooKeeper says alive but the NIC is dead." Selector-driven non-blocking connect / read / write with 10-second deadlines and SO_KEEPALIVE — one sick node can never wedge the whole cluster.
Topology-Aware Placement & Auto-Recovery
A 6-pass cascade spreads shard replicas across Zone → Rack → Host → Node so a rack or zone outage loses at most one copy. Reactive Shard Repair scans every 60 s and re-replicates diverging replicas automatically — self-healing without operator intervention.
Replica Read Placement & Read Failover
Writes replicate synchronously to every replica, so any replica holds the same rows — and the same secondary indexes, HNSW vector indexes and Lucene full-text indexes — as the primary. Spread reads across replicas (round_robin) and the index copies replication already keeps resident stop sitting idle and start serving searches, raising read throughput and search parallelism. Read failover applies under either setting: one unreachable data node no longer fails a SELECT, a full-text query or a vector search — its shards are retried on the surviving replicas.
Vector Index Memory Budget & Prewarm
An HNSW graph is loaded whole into the heap, so each node enforces a resident budget (40% of max heap by default) and evicts the least recently used indexes once it is crossed — flushing them to their encrypted sidecars first, so the cost is a reload, never lost writes, and never an OOM. After a restart the leading coordinator opens the vector and full-text indexes up front, so the first search does not pay decrypt and deserialize inside user latency.
Use Cases
Graph RAG & Agent Retrieval
Hard Filter (SQL WHERE) + semantic/keyword hybrid retrieval + graph-based re-ranking (structural — by graph proximity) + fact-check, all in a single SELECT. Skip the four-system stack (Postgres + Pinecone + Elasticsearch + Neo4j) and run RAG / agent retrieval from one database.
Knowledge Graph & Ontology Store
Model entities and relationships as ordinary tables; traverse with BFS, score with PageRank / Personalized PageRank. RBAC, transactions, and backup come for free — no separate graph DB to operate.
Unified OLTP & Analytics
OLTP transactions auto-sync to Iceberg via CDC, and Ontul / Trino / Spark analyse the same data. No bespoke ETL or CDC pipeline to maintain.
Large-Scale Distributed DB & Real-Time Ingest
Horizontal sharding with automatic rebalancing plus direct Kafka ingestion handles petabyte-scale transactional workloads.
Location-Aware & Geo RAG
Combine PostGIS-compatible spatial functions with vector embeddings for location-based recommendation, GIS, and geo-aware retrieval — all in one engine.
Considering NeorunBase for your data platform?
OLTP · Vector DB · FTS · Graph DB · Spatial · DR — One ACID SQL Engine.
Multi-modal data — vector DB, graph DB, and multi-site DR — in one PostgreSQL-compatible engine.