Performance Tuning

Data Plane Cores

[server]
data_plane_cores = 4    # default: CPUs - 1

More cores = more parallel shards. Each core is an isolated, shared-nothing shard.

io_uring

The Data Plane uses io_uring for NVMe I/O. Ensure:

  • Linux kernel 5.1+ (5.11+ recommended for full io_uring features)
  • NVMe storage (io_uring benefits are minimal on spinning disks)
  • ulimit -l unlimited for locked memory

Vector Engine

ParameterDefaultTune for
M16Higher = more recall
EF_CONSTRUCTION200Higher = better graph
QuantizationnoneSQ8/PQ for large datasets

FTS Engine

ParameterDefaultTune for
Memtable threshold32M postingsHigher = fewer flushes
Compaction levels8x8More levels = less write amplification

Query Optimization

EXPLAIN SELECT * FROM users WHERE age > 30;

Check that:

  • Vector searches use the HNSW index (not brute force)
  • Spatial queries use the R*-tree (check for SpatialOp)
  • Columnar queries show predicate pushdown and block skip
  • FTS queries use the inverted index (not full scan)
View page sourceLast updated on Apr 18, 2026 by Farhan Syah