OLAP Comparison

Netflix Dataset (100M rows)

Query	ParticleDB	DuckDB	PostgreSQL	PDB vs DuckDB
COUNT(*)	0.04ms	<1ms	1,075ms	Metadata
SUM(rating)	0.08ms	7ms	1,379ms	90x
AVG GROUP BY movie (17K groups)	31.9ms	68ms	2,665ms	2.1x
COUNT GROUP BY rating (5 groups)	5.2ms	19ms	2,221ms	3.7x
GROUP BY 500K groups	11.5ms	12ms	—	Parity

ParticleDB wins 5 of 7 queries vs DuckDB.

Yandex Dataset (100M rows)

Query	ParticleDB	DuckDB	PostgreSQL	PDB vs DuckDB
COUNT(*)	0.04ms	<1ms	1,157ms	Metadata
SUM(dwell_time)	0.06ms	10ms	1,651ms	170x
GROUP BY session_id (5M groups)	66ms	89ms	11,137ms	1.3x
GROUP BY action (4 groups)	20ms	24ms	3,045ms	1.2x
Filter + aggregate	8.8ms	9ms	—	Parity

ParticleDB wins 7 of 8 queries vs DuckDB.

Airline Dataset (100M rows)

Query	ParticleDB	DuckDB	PDB vs DuckDB
SUM(distance)	0.06ms	10ms	156x
AVG GROUP BY carrier (20 groups)	27ms	33ms	1.2x
Filter + 3 aggregates	10ms	23ms	2.3x
GROUP BY 500K groups	71ms	96ms	1.4x

Where DuckDB Wins

DuckDB is faster on some query patterns:

• Multi-column GROUP BY with millions of groups: DuckDB’s morsel-driven vectorized execution handles very high cardinality GROUP BY ~2-3x faster
• Complex JOIN + aggregate: DuckDB’s compiled pipeline avoids per-row function dispatch overhead
• String-heavy operations: DuckDB’s string dictionary encoding is more tightly integrated

We’re actively working on vectorized hash aggregation to close these gaps.

Why ParticleDB Excels at Analytics

Technique	Benefit
Zone map precomputation	Ungrouped SUM/AVG/MIN/MAX in O(1)
Contiguous flat columns	Hardware prefetcher-friendly scans
DictU32 string encoding	4-byte keys for string GROUP BY
SIMD mask accumulation	Branchless low-cardinality GROUP BY
Radix-partitioned hash tables	L2 cache-friendly for high cardinality
Tree-reduction merge	O(log N) parallel merge instead of O(N)