LanceDB launches Realtime Engine for millisecond-scale live search

LanceDB has introduced the Realtime Engine, a LanceDB Enterprise capability that lets applications search newly written rows within milliseconds. The system combines streaming writes, live scalar, full-text and vector indexes, and durability in object storage.
- The Realtime Engine is being introduced for LanceDB Enterprise.
- New rows can be searched within milliseconds after durable writes.
- The system supports live scalar, full-text and vector indexes.
- A benchmark reported 52 ms median visibility from durable S3 write to readable row.
- MemWAL adds an LSM-tree structure to the open Lance table format.
A streaming alternative to batch commits
AI applications often need to use data while it is still arriving. Robotics and physical AI labs continuously produce images and sensor readings, while annotation and feature pipelines add labels and embeddings. Reinforcement learning systems generate model responses, tool calls and artifacts, then add rewards and training examples.
LanceDB says these workloads need reads to keep up with writes. A corrected label should affect the next filter, and a recomputed embedding should appear in the next vector search, without waiting for a batch commit or index refresh.
Traditional lakehouse designs collect data into large batch commits, write columnar files to object storage and then publish a new table version. Frequent commits can create small files and metadata updates, while larger batches make readers wait. The Realtime Engine is designed to remove that tradeoff by allowing writes and reads to proceed without waiting for a table-wide batch commit.
Durable writes and live indexes
The capability is built on MemWAL, a format LanceDB developed as part of the Lance table format. MemWAL gives Lance a log-structured merge tree on object storage, combining a write-ahead log, in-memory indexed buffers and indexed data files.
New rows are logged in object storage and added to an indexed MemTable. When MemTables fill, they are written as indexed Lance files called SSTables, which are merged into the base table in the background. Queries search MemTables alongside SSTables and the base table, supporting vector search, full-text search, point lookups and scalar filters.
According to LanceDB, a write returns only after its data is durable in object storage. LanceDB Enterprise manages shard placement, writer recovery, memory limits and backpressure, while separate workers handle compaction, index updates and garbage collection. Queries continue using SSTable indexes while the base table's indexes are updated.
Benchmark results across three datasets
In a benchmark on Amazon EKS, LanceDB reports a median time of 52 ms from a durable write in S3 to a readable row. The tests used FineWeb-Edu, MS MARCO and LAION. Each single-node run used one m8gd.8xlarge node with 30 vCPU and 112 GiB.
With live primary-key indexing, durable write throughput increased from 79-83 MB/s on one shard to 312-334 MB/s on four shards on a single node. Across four nodes, each serving its own four-shard table, combined throughput reached 946 MB/s. During ingestion, median query latency was 12-14 ms for key lookups, 15-28 ms for full-text search and 26-30 ms for vector search.
LanceDB also reports that, on LAION with only the primary-key index live, peak node memory fell from 74 GiB to 18 GiB while throughput stayed within 3%. Tables can choose which indexes remain live, allowing users to balance write throughput against search speed for the newest rows.
What the announcement means for AI data pipelines
The Realtime Engine is aimed at teams that need one table to support incoming data, durable storage and searches across recent and historical records. Robotics teams can find new frames and annotations as they arrive, while reinforcement learning pipelines can read fresh training signals and preserve which rollouts contributed to a checkpoint.
LanceDB positions the design against separate buffering and search-synchronization layers. Because the log, metadata, files and indexes are part of the open Lance table format, the company says the approach avoids keeping newly acknowledged data outside the table until a later flush. LanceDB Enterprise provides the managed execution and maintenance layer for this streaming format.


