Docs
  • 0.49.1latest
  • 0.48.0
  • 0.47.0
  • 0.46.0
  • 0.45.0
  • 0.44.0
  • 0.43.0
  • 0.42.0
  • 0.41.0
  • 0.40.0
  • 0.39.6
  • 0.39.5
  • 0.39.3
  • 0.39.2
  • 0.39.1
  • 0.39.0
  • 0.38.4
  • 0.38.2
  • 0.38.1
  • 0.37.1
  • 0.37.0
  • 0.36.2
  • 0.36.1
  • 0.36.0
  • 0.35.3
  • 0.35.2
  • 0.35.1
  • 0.35.0
  • 0.34.4
  • 0.34.3-beta
  • 0.34.2-beta
  • 0.34.1-beta
  • 0.34.0-beta
  • 0.33.4-beta
  • 0.33.3-beta
  • 0.33.2-beta
  • 0.33.1-beta
  • 0.33.0-beta
  • 0.32.0-beta

Columnar tables

DEFINE TABLE ... STORE COLUMN keeps a columnar projection beside the row authority — per-chunk statistics prune analytical reads, and a GROUP BY on the partition key folds each lane and merges.

DEFINE TABLE events (seq int, kind string) STORE COLUMN;
DEFINE TABLE events (seq int, kind string) STORE COLUMN PARTITIONS 4 BY kind;
DEFINE SERIES readings RETAIN 30d TIME at STORE COLUMN PARTITIONS 4 BY sensor;

A table stores its records as versioned rows. Saying STORE COLUMN beside the definition adds a columnar projection of the same records, written inside the record's own commit batch — every chunk of the projection lands in the same atomic write as the rows it projects, and a statement that refuses refuses both. The row stays the authority: a read served through the projection names candidates, and the row at the snapshot confirms each one, so a columnar table answers exactly what the row store answers — the plan and the answer's note say which path it took and how many chunks were examined against skipped.

What the projection buys

Analytical reads over many rows skip what they do not need. Every sealed chunk carries min/max bounds and a bloom filter as write-path invariants, so a selective predicate prunes whole chunks without decoding them: measured over 1 048 576 rows, a top-1% range aggregate answers in 75 ms against the row walk's 2.7 s, skipping 126 of 128 chunks. A GROUP BY on the partition key folds each partition's lanes and merges the partials — the plan reports lanes: n.

The declaration

  • STORE ROW is the default, and the word INFO FOR TABLE and BACKUP SCRIPT render — a script that never says the word reads back with it said.
  • The kind is immutable after define. ALTER TABLE … STORE … refuses StorageKindImmutable: the migration between layouts is DROP and DEFINE, not a clause.
  • PARTITIONS n BY field routes each row's projection to one of n lanes by the hash of the key field's value; a row whose key is null or absent routes to partition 0. The count and key are definition-time constants.
  • Only record-shaped kinds take the clause: a plain table and a series. An edge table, queue, topic or space refuses StorageKindOnKind.

The stated limits

Each is deliberate rather than implicit:

  • Feeds refuse on a columnar table (FeedOnColumnar) — the feed captures from the row authority, which a columnar table fully retains, so the composition is designed but not built yet.
  • Single-row updates cost a chunk copy-on-write. Columnar is not free for OLTP-shaped writes: measured at one shape (100 000 rows in one commit), the columnar write costs 2.11× the row write. The tax is published, not hidden.
  • Version and point reads are always row-path — exact, never pruned. Series reads are row-path at this milestone too; the projection holds the points, the row serves the reads.
  • Projection vacuum rides the store's version reclamation — no separate merge job ships yet.