Benchmarks

Representative latency and scalability benchmarks for LanceDB Enterprise.

LanceDB Enterprise is designed for low-latency, high-throughput search, but observed performance depends on factors such as dataset shape, index configuration, cache warmth, filter selectivity, concurrency, and cluster sizing. The figures on this page should be read as representative benchmark results for a specific test setup, not as universal guarantees for every deployment.

In our benchmark environment, warmed-cache vector search reached around 25ms P50 latency, and metadata-filtered search reached around 50ms P99 latency for the filter pattern shown below.

If you want performance guidance for your own workload, reach out to [e-mail]. The LanceDB engineering team can help map your latency, throughput, ingestion, and filtering requirements to an appropriate Enterprise cluster design.

Percentile	Vector Search	Vector Search w. Filtering	Full-Text Search
P50	25ms	30ms	26ms
P90	26ms	39ms	37ms
P99	35ms	50ms	42ms

Depending on workload and tuning, Enterprise clusters can also be configured for high concurrency, including thousands of QPS in some deployments, but the right configuration varies by use case. Training, search, and analytics workloads often benefit from different cluster shapes and resource allocation strategies. To understand which parts of the system influence these results, see the Enterprise architecture guide.

Dataset

We used two datasets for this benchmark: the dbpedia-entities-openai-1M for vector search, and a synthetic dataset for vector search with metadata filtering.

Name	# Vectors	Vector Dimension
dbpedia-entities-openai-1M	1,000,000	1536
synthetic dataset	15,000,000	256

These benchmark results are most useful as a directional baseline. Different data distributions, index choices, cache behavior, and cluster settings can materially change the latency profile.

Vector Search

We ran vector queries against dbpedia-entities-openai-1M with a warmed-up cache. In that benchmark setup, we observed the following latency profile:

Percentile	Latency
P50	25ms
P90	26ms
P99	35ms
Max	49ms

Full-Text Search

With the same dataset and a warmed-up cache, full-text search fell into the following range:

Percentile	Latency
P50	26ms
P90	37ms
P99	42ms
Max	98ms

Vector Search with Metadata Filtering

We created a 15M-vector dataset to evaluate metadata-aware search under more complex filtering conditions. These filters can span a wide range of scalar columns, for example, “find Sci-fi movies since 1900”.

With a warmed-up cache, slightly more selective filters, for example, “find Sci-fi movies between the years 2000 and 2012”, produced the following representative results:

Percentile	Latency
P50	30ms
P90	39ms
P99	50ms

Broader or less selective filters, for example, “find Sci-fi movies since 1900”, moved the latency range to:

Percentile	Latency
P50	65ms
P90	76ms
P99	100ms

These benchmarks are intended to provide consistent, reproducible reference points for LanceDB Enterprise rather than one-size-fits-all promises. We periodically re-run and update numbers as necessary, but production performance will still depend on workload shape, cluster tuning, and the architectural choices described in the Enterprise architecture guide.