pgvector HNSW Index Build Slow

How to fix it

// try these in order, the most common fix is first

1. FIX 1Raise maintenance_work_mem for the session

   Copy steps
   1. 01Run SET maintenance_work_mem = '8GB'; in the session that builds the index
   2. 02Neon suggests staying under about 50 to 60 percent of compute RAM
   3. 03Do not set it so high that the server runs out of memory
   Did this fix it? Yes No

2. FIX 2Add parallel workers

   Copy steps
   1. 01SET max_parallel_maintenance_workers = 7;
   2. 02Raise max_parallel_workers (default 8) if needed
   Did this fix it? Yes No

3. FIX 3Load before you index

   Copy steps
   1. 01Insert data first, then create the index
   2. 02Use CREATE INDEX CONCURRENTLY in production
   3. 03Monitor with pg_stat_progress_create_index
   Did this fix it? Yes No

Symptoms

CREATE INDEX USING hnsw runs for hours and the log shows: hnsw graph no longer fits into maintenance_work_mem.

Why it happens

• C-01maintenance_work_mem too low
• C-02Default parallel maintenance workers (2)
• C-03Large table with high dimensions
• C-04Small instance

// this guide also covers

• pgvector hnsw index slow
• hnsw graph no longer fits into maintenance_work_mem
• pgvector index build time

The same problem in other apps

All export, render & quality loss guides →

• Chroma pip install "Failed to build hnswlib" Error (Chroma)

Other pgvector errors

• pgvector Error: extension "vector" is not available
• pgvector "operator does not exist: vector <=>" Error
• pgvector Dimension Limit: Cannot Index More Than 2000 Dimensions
• pgvector "different vector dimensions" Error
• pgvector Query Returns Fewer Results Than LIMIT With Filter
• pgvector Index Not Being Used
• pgvector HNSW Build Out of Memory or Shared Memory Error
• How to Enable pgvector on Supabase