Vector Plus

The EDB Vector Plus (edb_vectorplus) extension adds the ivfplus index access method for vector similarity search. Vector Plus uses hierarchical Inverted File (IVF) clustering and RaBitQ quantization to deliver faster queries with a lower memory footprint at large row counts.

On WarehousePG (WHPG), Vector Plus builds on the pgvector extension, and each segment builds and scans its own ivfplus index over its own slice of the table.

Vector Plus requires pgvector 0.7.0 or later. The WarehousePG pgvector extension meets this requirement. Install pgvector before you enable Vector Plus.

Installing Vector Plus

Refer to Downloading and installing an extension for installation and setup instructions. The package name is edb-whpg7-vectorplus.

Enabling the extension

Enable the extension in each database where you want to use it. CASCADE creates pgvector's vector type first, which edb_vectorplus requires:

CREATE EXTENSION edb_vectorplus CASCADE;

For more information, see Configuring Vector Plus.

Using Vector Plus

For guidance on creating ivfplus indexes, tuning queries, filtering, and measuring recall, see Using Vector Plus. For the full list of operator classes, index options, parameters, and functions, see the Vector Plus reference.

Considerations for WarehousePG

Account for the following differences when you run Vector Plus on a WarehousePG cluster. Each segment holds its own index, so Vector Plus behaves differently than on a single Postgres node.

• Set optimizer = off before running queries. The ORCA optimizer has no path for a custom index access method, so it plans a sequential scan and never uses ivfplus.
• The ivfplus.probes and ivfplus.max_probes configuration parameters apply per segment. A query over an N-segment cluster visits up to probes × N lists in total, which increases both recall and work compared to the same setting on a single node.
• The lists index option also applies per segment. Each segment clusters roughly rows / N rows into lists lists, so divide the usual rows / 1000 sizing rule by N.
• The ivfplus_metapage_info() and check_ivfplus_index_health() functions describe the coordinator's copy of the index, which holds no rows. They report an empty, healthy index regardless of the state of the segments. To see the real state, see Checking index health across segments.
• Index builds are serial.

Checking index health across segments

To check the real state of an ivfplus index, query each primary segment directly and sum the counters.

1. From the coordinator, list the primary segments using the gp_segment_configuration catalog:

   → WrapCopy

   SELECT hostname, port FROM gp_segment_configuration
     WHERE role = 'p' AND content >= 0 ORDER BY content;

2. For each segment, connect in utility mode and run the health check. Replace <index_name> with the name of your index:

   → WrapCopy

   PGOPTIONS='-c gp_role=utility' psql -h <hostname> -p <port> -d <database_name> \
     -c "SELECT * FROM check_ivfplus_index_health('<index_name>');"

3. Sum the n_frozen_live, n_dead_lanes, and n_pending counters across all segments.