Columnar Engine

Typed columns with per-column compression — the same approach as ClickHouse and DuckDB, living alongside your OLTP data.

When to Use

Analytical queries (GROUP BY, aggregations, window functions)
Reporting dashboards
Data science workloads
Any scan-heavy workload reading few columns from many rows
HTAP: pair with strict documents for combined OLTP + OLAP

Compression Codecs

Each column gets a codec chain tuned for its data type:

Codec	Target	Approach
ALP	Floats	Lossless float-to-integer conversion
FastLanes	Integers	SIMD bit-packing
FSST	Strings	Substring dictionary compression
Gorilla	Metrics	XOR-based compression
Pcodec	Complex numerics	Adaptive numeric compression
rANS	Cold-tier data	Entropy coding
LZ4	All	Terminal stage compression

Multi-stage pipeline achieves 20-40x compression on typical workloads.

Block Statistics

Data is stored in 1024-row blocks. Each block has min/max/null-count statistics. The query engine skips blocks that can't match the predicate — no decompression needed.

DDL

-- Plain columnar
CREATE COLLECTION logs TYPE COLUMNAR (
    ts TIMESTAMP TIME_KEY,
    host VARCHAR,
    level VARCHAR,
    message VARCHAR
);

-- Timeseries profile
CREATE COLLECTION metrics TYPE COLUMNAR (
    ts TIMESTAMP TIME_KEY, host VARCHAR, cpu FLOAT
) WITH profile = 'timeseries', partition_by = '1h';

-- Spatial profile
CREATE COLLECTION locations TYPE COLUMNAR (
    geom GEOMETRY SPATIAL_INDEX, name VARCHAR
);

Queries

SELECT level, COUNT(*) FROM logs
WHERE ts > now() - INTERVAL '1 hour'
GROUP BY level ORDER BY COUNT(*) DESC;

-- Window functions
SELECT host, message,
       ROW_NUMBER() OVER (PARTITION BY host ORDER BY ts DESC) AS rank
FROM logs;

HTAP Bridge

Combine strict (OLTP) with columnar (OLAP):

CREATE COLLECTION orders TYPE DOCUMENT STRICT (...);

CREATE MATERIALIZED VIEW order_analytics AS
SELECT status, DATE_TRUNC('day', created_at) AS day, COUNT(*), SUM(total)
FROM orders GROUP BY status, day;

-- Point lookups → strict engine. Scans → columnar engine. Automatic routing.

Delete Bitmaps

Deleted rows are tracked with Roaring Bitmaps. Three-phase crash-safe compaction reclaims space: write new segments → swap references → delete old segments.