Skip to content
TensorDB

DML — Data Manipulation

INSERT

Every INSERT creates an immutable fact in the ledger.

-- Single row
INSERT INTO users (id, name, email) VALUES ('u1', 'Alice', 'alice@co.com');


-- Multiple columns
INSERT INTO products (id, name, price, category)
VALUES ('p1', 'Widget', 9.99, 'tools');

INSERT behavior

  • Each INSERT assigns a new commit_ts (monotonically increasing)
  • The row is stored as table/<table_name>/<first_column_value>
  • No uniqueness constraint — inserting the same key creates a new version

UPDATE

UPDATE in TensorDB creates a new version of matching rows:

UPDATE users SET name = 'Alice Chen' WHERE id = 'u1';

This doesn’t modify the original row. Instead, it writes a new fact with a higher commit_ts. The original is still accessible via AS OF.

DELETE

DELETE performs a soft delete by writing a tombstone:

DELETE FROM users WHERE id = 'u1';

The row becomes invisible to current queries but remains in the ledger. Time-travel queries (AS OF) can still see it.

Write Batching

For bulk inserts, use the Rust API for better performance:

let batch = vec![
    ("key1", Value::Text("value1".into())),
    ("key2", Value::Text("value2".into())),
    ("key3", Value::Text("value3".into())),
];
db.put_batch(batch)?;

Write batching amortizes the WAL fsync cost across multiple writes.