ruvector-postgres:2.0.1 Docker image ships ruvector extension 0.3.0, not 2.0.0

[stuinfla]

Bug Report

Image: ruvnet/ruvector-postgres:2.0.1 (Docker Hub)
Platform: Fly.io (iad region), PostgreSQL 17.7

Problem

The Docker image ruvnet/ruvector-postgres:2.0.1 installs the ruvector extension at version 0.3.0, not 2.0.0 or 2.0.1 as the image tag suggests.

postgres=# SELECT extversion FROM pg_extension WHERE extname='ruvector';
 0.3.0

postgres=# SELECT * FROM pg_available_extension_versions WHERE name = 'ruvector';
 ruvector | 0.1.0 | f
 ruvector | 0.3.0 | t   -- installed
 ruvector | 2.0.0 | f   -- available but no update path from 0.3.0

Impact

1. HNSW indexes break non-vector queries

In ruvector 0.3.0, HNSW indexes interfere with the query planner for non-vector operations. Simple COUNT(*) queries on tables with HNSW indexes fail with:

ERROR: HNSW: Could not extract query vector from parameter.

This makes diagnostic queries, health checks, and any aggregation queries fail on tables that have HNSW indexes. The workaround is to drop all HNSW indexes, which removes O(log n) search capability.

2. No upgrade path from 0.3.0 to 2.0.0

ALTER EXTENSION ruvector UPDATE TO '2.0.0' returns:

ERROR: extension "ruvector" has no update path from version "0.3.0" to version "2.0.0"

The only way to upgrade is DROP EXTENSION ruvector CASCADE which destroys ALL columns of type ruvector across ALL schemas — a destructive operation that requires full re-embedding.

3. The latest image has the same problem

ruvnet/ruvector-postgres:latest (updated 2026-03-03) ships the same 0.3.0 extension .so file. Even after deploying latest, the shared library at /usr/lib/postgresql/17/lib/ruvector.so has 0.3.0 internals. Attempting to create the extension at version 2.0.0 with this .so results in missing functions:

ERROR: could not find function "ruvector_embed_wrapper" in file "/usr/lib/postgresql/17/lib/ruvector.so"

Expected Behavior

• ruvnet/ruvector-postgres:2.0.1 should ship ruvector extension 2.0.x
• ruvnet/ruvector-postgres:latest should ship the latest stable ruvector extension
• There should be an upgrade path from 0.3.0 to 2.0.0 (or clear migration documentation)
• HNSW indexes should not interfere with non-vector queries

Requested Fix

1. Rebuild Docker images with the correct ruvector .so matching the image tag version
2. Add an update path SQL script (0.3.0 → 2.0.0) or document the migration procedure
3. Fix the HNSW planner interference bug in the extension

Environment

• Docker image: ruvnet/ruvector-postgres:2.0.1 and :latest
• PostgreSQL: 17.7 (Debian 17.7-3.pgdg12+1)
• Host: Fly.io (iad region, 256MB RAM)
• Local comparison: ruvector 2.0.0 extension on local Docker (same image tag 2.0.1) works correctly — suggesting the images on Docker Hub may have been built at different times with different .so files

[stuinfla]

Additional Finding: HNSW bug exists on both 0.3.0 and 2.0.0

After further testing, the HNSW planner interference bug exists in the shared library (.so) file shipped with ALL current Docker images, regardless of SQL extension version:

• ruvnet/ruvector-postgres:2.0.1 with extension 0.3.0 — HNSW bug present
• ruvnet/ruvector-postgres:2.0.1 with extension 2.0.0 — HNSW bug still present
• ruvnet/ruvector-postgres:2.0.4 with extension 2.0.0 — HNSW bug still present

The bug is in the compiled Rust code (ruvector.so), not the SQL extension scripts.

Additional issue: v2.0.4 image missing v0.3.0 SQL files

The 2.0.4 image only ships ruvector--2.0.0.sql (no 0.3.0 scripts). If a database was previously running 0.3.0, there is:

• No upgrade path (ALTER EXTENSION UPDATE TO '2.0.0' fails)
• No rollback path (cannot CREATE EXTENSION VERSION '0.3.0')

The only option is DROP CASCADE which destroys all ruvector-type columns.

Additional issue: v2.0.4 image .so missing functions

The ruvector.so in the 2.0.4 image does not contain ruvector_embed_wrapper, so ruvector_embed() fails:

ERROR: could not find function "ruvector_embed_wrapper" in file "/usr/lib/postgresql/17/lib/ruvector.so"

Working workaround found

Extension SQL at version 2.0.0, running on the 2.0.1 image's .so file — this combination has working ruvector_embed(). HNSW indexes still break COUNT(*) but vector search via sequential scan works at 243ms for 87 rows.

[stuinfla]

Summary: All Workarounds Blocked

This issue blocks production use of HNSW indexes on ruvector-postgres Docker deployments. Here is every approach we attempted and why each fails:

Attempted Workarounds (all blocked)

#	Approach	Result	Blocker
1	Use HNSW indexes on ruvector-postgres:2.0.1	COUNT(*) and non-vector queries fail	HNSW planner bug in .so
2	Upgrade extension 0.3.0 → 2.0.0 in-place	ERROR: no update path from version "0.3.0" to version "2.0.0"	Missing migration SQL
3	DROP CASCADE + CREATE at 2.0.0 on same image	2.0.0 creates but same .so = same HNSW bug	Bug is in compiled Rust, not SQL
4	Switch to ruvector-postgres:2.0.4 image	ruvector_embed() broken: could not find function "ruvector_embed_wrapper"	.so compiled without embed functions
5	Switch to ruvector-postgres:latest image	Same as #4 — ruvector_embed_wrapper missing	Same .so as 2.0.4
6	Use 2.0.4 image + rollback to extension 0.3.0	ERROR: no installation script for version "0.3.0"	2.0.4 image doesn't ship 0.3.0 SQL
7	Cross-build local ARM image for AMD64 Fly.io	Architecture mismatch (local is Apple Silicon ARM64, Fly.io is AMD64)	Would need full Rust nightly build chain

Current Workaround (degraded)

Extension at 2.0.0, running on the 2.0.1 image's .so file. This gives us:

• ruvector_embed() — works (384-dim vectors)
• Vector similarity search (<=>) — works via sequential scan (243ms for 87 rows)
• HNSW indexes — broken (creates successfully but interferes with ALL queries including COUNT(*), SELECT *, UNION ALL)
• All HNSW indexes dropped from production

Business Impact

• Production database serving a client CMO platform on Vercel + Fly.io
• 65,657 rows across 5 schemas with vector embeddings
• Daily autonomous agent cycle relies on vector search
• Cannot use HNSW indexing → forced to sequential scan → will not scale past ~1,000 knowledge entries
• 50+ minutes of production downtime caused by attempting workaround Reorganize repo structure and update documentation #3 (DROP CASCADE destroyed all embedding columns, required manual re-embedding of 1,996 rows)

What We Need

1. A Docker image where HNSW indexes don't break non-vector queries — this is the critical fix
2. A migration path from 0.3.0 to 2.0.0 — either an update SQL script or documentation
3. Consistent .so files — ruvector_embed() should work on all image tags

Reproduction

# On any ruvector-postgres Docker image (2.0.1, 2.0.4, latest)
docker run -d --name test-rv -e POSTGRES_HOST_AUTH_METHOD=trust ruvnet/ruvector-postgres:2.0.1
docker exec test-rv psql -U postgres -c "CREATE EXTENSION ruvector;"
docker exec test-rv psql -U postgres -c "CREATE TABLE t(id serial, v ruvector(3));"
docker exec test-rv psql -U postgres -c "INSERT INTO t(v) VALUES ('[1,2,3]'), ('[4,5,6]');"
docker exec test-rv psql -U postgres -c "CREATE INDEX ON t USING hnsw (v ruvector_cosine_ops);"
docker exec test-rv psql -U postgres -c "SELECT count(*) FROM t;"
-- Expected: 2
-- Actual: ERROR: HNSW: Could not extract query vector from parameter