A durable table design starts with row semantics, then selects sort keys, partition boundaries, and physical distribution.
Duplicate keeps every row, Unique keeps one logical row per key, and Aggregate combines value columns with declared functions.
Append facts, logs, and detailed events without key-based merging.
UPSERT mutable entities and preserve the latest logical row.
Merge measures with SUM, MAX, MIN, REPLACE, or related functions.
Doris stores data ordered by key columns, so the leading key prefix has an outsized effect on pruning, compression, and scan cost.
Compact columns aligned with common predicates.
Wide columns unrelated to normal access paths.
Partitions isolate data ranges or value groups, support lifecycle operations, and let queries avoid unrelated data.
Explicit RANGE or LIST boundaries with direct operator control.
Scheduled time partitions with rolling creation and expiration.
Create RANGE or LIST partitions from values during ingestion.
Hash bucketing gives deterministic placement by key; random bucketing spreads rows without requiring a distribution key.
The same key values map deterministically.
Rows spread without a chosen key.
The strongest DDL is internally consistent: row semantics, pruning, distribution, and ingestion behavior all reinforce the workload.
Duplicate rows, time pruning, and even parallel scans.
Unique business keys with UPSERT-oriented ingestion.
Stable dimensions with declared value aggregation.
Check partition pruning, tablet size, row distribution, query plans, and load behavior before scaling the workload.
