Bench: Lance point lookup vs LMDB

[wojiaodoubao]

Environment

• lance 0.30: https://pypi.org/project/pylance/0.30.0/
• pytest & pytest-benchmark
• ECS instance (volcano engine ecs.r3il.2xlarge)
  • CPU: 8vCPU
  • Mem: 64GB
  • OS: Linux version 5.4.250-9-velinux1u1-amd6

Bench code

Code, designed by @ddupg.

The bench is ran on "memory://test.lance", which means all data is stored in memory.

Bench result

link

Simple Analysis

1. As the batch size increases, the gap between Lance and LMDB narrows.
2. With the increase in batch size and value size, version 2.1 shows a performance regression compared to version 2.0.

Future work

Recently, there has been a performance optimization(#4090) where Lance can benefit a lot in single-point lookup scenarios. I'm going to test it.

Related to #3983