til/sqlite/add-check-constraint-to-existing-column.md

Add Check Constraint To Existing Column

To not bury the lede, SQLite does not support adding a check constraint to an existing column. Instead, if I want to add one after the fact, I can reconstruct the table with the check constraint specified up front and then migrate the data.

Here is a sessions table that includes start_time and end_time columns that use the text data type to represent points in time.

sqlite> .schema sessions
CREATE TABLE sessions (
        id integer primary key,
        active integer not null check (active in (0, 1)),
        project_id integer not null references projects(id) on delete cascade,
        start_time text not null,
        end_time text,
        created_at text not null default (datetime('now')),
        updated_at text not null default (datetime('now'))
    );
CREATE UNIQUE INDEX idx_sessions_single_active
        on sessions(active)
        where active = 1;

I want start_time and end_time to both enforce the shape of the timestamp strings with check constraints. To do that, I need to start a transaction, create a new version of the table with the check constraints, migrate the data, rename sessions to sessions_old, rename sessions_new to sessions (that's the in-place swap), and then commit the transaction. sessions_old can be dropped later once I feel good about the migration.

begin transaction;

create table sessions_new (
  id integer primary key,
  active integer not null check (active in (0, 1)),
  project_id integer not null references projects(id) on delete cascade,
  start_time text not null,
  end_time text,
  created_at text not null default (datetime('now')),
  updated_at text not null default (datetime('now')),
  check(start_time is strftime('%Y-%m-%dT%H:%M:%fZ', start_time)),
  check(end_time is strftime('%Y-%m-%dT%H:%M:%fZ', end_time))
);

insert into sessions_new (
  id, active, project_id, start_time, end_time, created_at, updated_at
)
select id, active, project_id, start_time, end_time, created_at, updated_at
from sessions;

alter table sessions rename to sessions_old;

alter table sessions_new rename to sessions;

drop index if exists idx_sessions_single_active; -- on sessions_old

create unique index idx_sessions_single_active
  on sessions(active)
  where active = 1;

commit;

Notice that after the table renames I also drop index (now pointing to sessions_old) and recreate it for the new sessions table.

This approach worked well for my situation, but may not be a one-size-fits-all solution. Depending on how the database is deployed, the size of the tables, and usage, this approach may not scale. Always do a dry-run of database migrations like this.