Start with the logical hierarchy. Then zoom in to see how a table is divided and distributed across the cluster.
Analytical queries often touch a few columns across many rows. Columnar storage lets Doris read and process only the data those queries need.
Scan only the columns referenced by the query.
Similar typed values are easier to encode compactly.
Contiguous vectors improve cache use and enable SIMD.
The table model defines whether Doris keeps every row, combines values, or maintains one logical row for each Key.
Retains every input row for event details and logs.
Merges Value columns with declared aggregation functions.
Keeps one logical row per Key for updates and CDC.
In the integrated architecture, Frontend nodes coordinate the system while Backend nodes execute queries and persist tablet replicas on local storage.
Stateless compute groups can scale independently while durable data remains in shared storage.
FE serves SQL; Meta Service manages data-layer metadata.
Stateless BE nodes execute work and cache hot data.
Object storage or a supported distributed file system persists data.
Integrated and decoupled deployments optimize for different operating environments. Neither architecture is universally better.
Stateful BE nodes compute and persist data locally.
Stateless compute groups share persistent storage.
The storage engine organizes data and access paths. The query engine converts SQL into distributed, vectorized pipelines.
Reads required columns and compresses analytical data.
Prefix, Min/Max, Bloom, and inverted indexes prune scans.
Splits query execution across Backend nodes.
Applies rule-, cost-, and history-based strategies.
A query is validated, optimized, divided into fragments and instances, executed across BE nodes, and then combined into a result.
Doris combines precomputation, logical rewrites, statistics-based planning, and runtime filtering to reduce data before and during execution.
Reuse precomputed results through query rewrite.
Apply deterministic logical transformations.
Use statistics to compare physical plans.
Push filters toward scans to discard rows early.
