Guarantees and limits

Guarantees

• At least once, in order per row. A batch may be applied twice after a crash. Every write is idempotent, so that’s harmless. Changes to one row arrive on one shard, in order.
• Nothing is skipped. A failing batch is retried until it’s applied or parked in the dead-letter queue. The worker’s position moves past a batch only after one of those happens.
• A failed batch writes nothing. Vectors are written only after every embedding in the batch succeeded.

Consistency

The index trails the table by seconds. A query reads rows back from Keyspaces, so a returned row is current, but a row changed in the last few seconds may still be ranked by its previous text.

Limits

• Vector keys are at most 1,024 characters (S3 Vectors). A row whose encoded primary key is longer is skipped and logged, not retried.
• Filterable metadata is at most 2 KB per vector. Store long text with sync.store-text-as and make that key non-filterable.
• Titan Text Embeddings V2: 60 requests a minute, one text each, not adjustable. See Choose an embedding model.
• First start takes about 12 minutes while KCL prepares its lease table.
• One table per deployment.
• Changing the model means a new index.