Implementation:Lance format Lance MemtableReadBench

Knowledge Sources	Lance
Domains	Benchmarking, Performance
Last Updated	2026-02-08 19:33 GMT

Overview

Description

MemtableReadBench is a Criterion-based benchmark that compares read performance between in-memory MemTables (using MemTableScanner) and in-memory Lance tables. It exercises four distinct read operations, each comparing the MemTable path against a conventional Lance dataset:

• Scan — Full table scan returning all rows, measuring total throughput.
• Point Lookup — Scalar index-based point lookups using a BTree index on the primary key column.
• Full-Text Search (FTS) — Token-based text search using an inverted index on a text column.
• Vector Search — IVF-PQ vector similarity search, comparing a MemTable with trained IVF centroids and PQ codebook against a Lance dataset with a materialized IVF-PQ index.

The data schema includes three columns:

• id (Int64) — Primary key
• text (Utf8) — Text column with common word patterns for FTS testing
• vector (FixedSizeList<Float32>) — Normalized random vectors

The benchmark trains IVF centroids and a PQ codebook from the generated data to build MemTable-level vector search indexes. For Lance dataset comparisons, it creates equivalent materialized indexes.

Usage

This benchmark is used to evaluate whether the MemTable in-memory data structure can match or exceed the read performance of a fully materialized Lance table for different access patterns. It helps developers identify MemTable performance bottlenecks and optimize the MemTable, IndexStore, and CacheConfig components.

Code Reference

Source Location

rust/lance/benches/memtable_read.rs (1119 lines)

Signature

The benchmark defines a single orchestrator that delegates to four sub-benchmarks:

fn all_benchmarks(c: &mut Criterion) {
    bench_scan(c);
    bench_point_lookup(c);
    bench_fts(c);
    bench_vector_search(c);
}

Import

use lance::dataset::mem_wal::write::{CacheConfig, IndexStore, MemTable};
use lance::dataset::{Dataset, WriteParams};
use lance::index::vector::VectorIndexParams;
use lance_index::scalar::inverted::tokenizer::InvertedIndexParams;
use lance_index::scalar::FullTextSearchQuery;
use lance_index::vector::ivf::storage::IvfModel;
use lance_index::vector::ivf::IvfBuildParams;
use lance_index::vector::kmeans::{train_kmeans, KMeansParams};
use lance_index::vector::pq::builder::PQBuildParams;
use lance_index::{DatasetIndexExt, IndexType};
use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};

I/O Contract

Inputs

Parameter	Type	Default	Description
NUM_ROWS	Environment variable	10000	Total number of rows in the dataset
BATCH_SIZE	Environment variable	100	Number of rows per batch
VECTOR_DIM	Environment variable	128	Dimension of the vector column
SAMPLE_SIZE	Environment variable	100	Number of benchmark iterations (minimum 10)

Outputs

Output	Type	Description
Criterion report	HTML/JSON	Throughput statistics (rows/sec) for each benchmark group comparing MemTable vs Lance, with optional flamegraph profiling on Linux
Comparison data	Criterion stats	Side-by-side MemTable vs Lance dataset performance for scan, point lookup, FTS, and vector search

Usage Examples

Run the full benchmark suite:

cargo bench -p lance --bench memtable_read

Run with custom data size:

export NUM_ROWS=50000
export VECTOR_DIM=256
cargo bench -p lance --bench memtable_read

Run only the vector search sub-benchmark:

cargo bench -p lance --bench memtable_read -- "vector"

Run only the FTS sub-benchmark:

cargo bench -p lance --bench memtable_read -- "fts"

Related Pages

• Implementation:Lance_format_Lance_MemWalReadBench — Benchmark for LSM scanner read performance with MemWAL
• Implementation:Lance_format_Lance_MemWalWriteBench — Benchmark for MemWAL write throughput