ALTER TABLE

Schema evolution is catalog-only: channels are independent columnar streams, so nothing is ever rewritten. Adding a column, changing its read-time mechanics or renaming it is a metadata edit that takes effect immediately; only DROP touches data — it purges the channel's stream.

ALTER TABLE telemetry ADD COLUMN door BOOL FILL HOLD STALENESS 60 s;
ALTER TABLE telemetry ALTER COLUMN speed SET FILL NONE SET STALENESS 30 s;
ALTER TABLE telemetry RENAME COLUMN speed TO velocity;
ALTER TABLE telemetry DROP COLUMN old_sensor;

ADD COLUMN

Declares a new channel — same type and attribute syntax as CREATE TABLE. Existing rows simply read null for it; no backfill, no table rewrite, no locking. (Ingesting an unknown field over ILP auto-registers the channel too; ADD COLUMN is how you declare one with FILL HOLD/staleness.)

ALTER COLUMN SET

SET INDEX ON|OFF declares (or retires) a text/bytes column's secondary exact index. Folds and compactions index partitions written from then on; older partitions stay unindexed until compaction rewrites them or the data is re-imported — until coverage is complete, equality probes fall back to the scan.

SET FILL HOLD|NONE and SET STALENESS <duration>|INFINITE change read-time behavior only — fill never touches disk, so the new policy applies to all existing data on the next query.

SET CALCULATED ON|OFF marks a column as pipeline-written (see Calculated Fields). A calculated column's reads collapse a rewrite at the same (entity, ts) to the newest write — a recompute is a correction, not a second observation. ADD COLUMN <name> <type> CALCULATED declares one directly; the rule engine does this for you when you define a calculated field.

RENAME COLUMN

Names live in the catalog; the engine addresses channels by id. A rename moves no data and is visible to the next statement.

DROP COLUMN

Removes the declaration and purges the channel's stream: its in-memory delta, its registry entry and its on-disk segments. The channel id is never reused, so re-adding a column with the same name starts empty — old values cannot resurface.

SET FLEXIBLE / SET STRICT

A table-level toggle for schema-on-write:

ALTER TABLE devices SET FLEXIBLE;   -- an INSERT may name a new column; it is added
ALTER TABLE devices SET STRICT;     -- unknown columns are refused again (the default)

FLEXIBLE infers a new column's type from the first value written to it. STRICT only governs future writes — columns already added stay.

SET SYNC

Opt a timeseries table into the sync feed:

ALTER TABLE sensor SET SYNC;       -- start stamping x_seq on every fragment
ALTER TABLE sensor SET SYNC OFF;   -- stop; stamped history stays honest

Enabling (and re-enabling) bumps the table's sync epoch, so a cursor from a previous life is told to resync instead of silently missing the gap. Rows tables refuse this — they are sync-native already; kv tables have no feed by design.

SET RETENTION

Rewire a timeseries table's drop horizon in place:

ALTER TABLE telemetry SET RETENTION 90 d;   -- partitions older than the window are dropped
ALTER TABLE telemetry SET RETENTION OFF;    -- keep forever again

Catalog-only: the merge worker reads the table's definition on every GC pass, so the new window takes effect on the next cycle without a restart — shrinking the window starts dropping expired partitions, growing it simply stops. Rows and kv tables refuse it (kv has ttl, a rows table keeps current state), as does a MUTABLE table.

No type changes

ALTER COLUMN ... TYPE does not exist, by design: data on disk already has a shape. The representation-safe conversions, bool/int ↔ f64, are catalog relabels — ask for one by dropping and re-adding only if the history is disposable, or keep the column as f64.

entity and ts are implicit on every table and cannot be added, renamed or dropped.