Contents

• Owned late-interaction storage and write amplification

Owned late-interaction storage and write amplification

The owned late-interaction certification gate records both correctness across a PostgreSQL restart and the incremental storage/write cost of maintaining token rows in the source DML transaction.

Run the reproducible PostgreSQL 17 workload from the repository root:

PG_VERSION=pg17 PG_FEATURE=pg17 \
DBNAME=pgcontext_late_interaction_owned \
tests/heavy/late_interaction_ann_serving.sh

The workload registers four two-dimensional source points, applies an update and a logical point deletion, checkpoints and restarts PostgreSQL, verifies the owned HNSW generation, repairs it, and confirms that repair does not reactivate the deleted point. It then reports:

• logical point and token counts from the maintained registration counters;
• the sum of pg_column_size() for owned token rows in that collection;
• allocated bytes for the collection-scoped partial HNSW index; and
• WAL bytes generated by replacing one source row containing two tokens.

Reference result on 2026-07-21 (PostgreSQL 17.10, Apple arm64, development build):

These tiny-fixture numbers characterize fixed overhead and are not throughput claims. The HNSW value is page-allocation dominated. The gate is valuable because it makes changes to token-row shape, index allocation, or synchronous write amplification visible alongside restart and exact-recheck correctness.