Benchmarks

HyperMesh DB’s TPI Bucket Pushdown achieves 6–40× lower query latency than PostgreSQL 15 (B-tree index on event_ts) across dataset sizes from 10K to 1M records.

Results summary

N	HyperMesh median (µs)	PostgreSQL median (µs)	Speedup
10K	~80	~520	6×
100K	~95	~1,400	15×
1M	~110	~4,500	41×

Note

Results are from a MacBook Pro M3 (8GB RAM, macOS 14). PostgreSQL ran in a Docker container with tmpfs storage (best-case in-memory). See the methodology document for full details.

Why the speedup grows with N

The key insight: PostgreSQL’s B-tree scan must traverse index pages proportional to the number of matching rows — O(k log N) where k is the result set size.

HyperMesh’s TPI Bucket Pushdown computes the first and last relevant bucket via a single integer division, then calls pread() to read only those physical file segments. The cost is O(B_window) where B_window is the number of buckets that overlap the query window — a constant for fixed window / timeline ratio.

As N grows, k grows proportionally, making the PostgreSQL cost scale with N. HyperMesh’s cost stays flat.

How to reproduce

# With Docker (fully isolated, reproducible)
docker-compose -f docker-compose.benchmark.yml up -d pg
sleep 5
docker-compose -f docker-compose.benchmark.yml run --rm benchmark

# Without Docker (local Python + psycopg2)
pip install psycopg2-binary
HMDB_BENCH_PG_DSN="postgresql://user:pass@localhost:5433/hmdb_bench" \
  python benchmarks/compare_pg.py

Results are written to benchmarks/results/pg_comparison.{csv,txt}.

Full methodology

See benchmarks/METHODOLOGY.md for complete details including:

• Dataset generation methodology
• PostgreSQL configuration (shared_buffers, fsync=off, tmpfs)
• Timing measurement approach
• Limitations and future work (Neo4j, TimescaleDB, write throughput)