fix(core): create vector storage at initialization, never at runtime by phernandez · Pull Request #1656 · basicmachines-co/basic-memory

phernandez · 2026-10-06T00:54:03Z

Problem

Every embedding job builds a fresh PostgresSearchRepository, and its first vector call ran CREATE EXTENSION / CREATE TABLE IF NOT EXISTS / CREATE INDEX IF NOT EXISTS for the chunk and embedding tables. CREATE INDEX IF NOT EXISTS takes a SHARE lock on the table before it sees the index exists, so under concurrent writers every chunk/embedding write queued behind it. In production one tenant had ~28k index_embeddings* jobs time out (asyncpg 30s) in two days; pg_stat_activity showed 9 of these CREATE INDEX statements and ~50 writes waiting on relation locks.

Fix: no DDL at runtime

Runtime (_ensure_vector_tables, used by search and vector sync) only binds the vector adapter. PgVectorIndex.upsert/search/delete* no longer call initialize(); PgVectorIndex.initialize() is now a runtime no-op. A database without vector storage fails on its first query.
Initialization (init_search_index() → _create_vector_storage() → PgVectorIndex.create_storage()) creates the chunk and embedding tables. Local startup already calls this from db.run_migrations; cloud will call it from its once-per-tenant-per-process schema hook (companion cloud PR).
create_storage() reads the catalog first (pg_class/pg_attribute/pg_extension, scoped to current_schema()), so initializing an already-initialized database — every new worker — takes no locks on our tables. Only a fresh or outdated database runs DDL.

Tests

Real Postgres: with another connection holding ROW EXCLUSIVE on both vector tables, runtime binding and re-initialization both finish within 5s. On main this times out, the production failure.
Real Postgres: a layered search_path (tenant, public) with tables only in public still creates the tenant schema's own tables (Codex P1).
Unit: runtime upsert/delete issue no DDL; current storage needs exactly one catalog read.
Tests that relied on lazy runtime creation now call init_search_index().
Postgres repository/services: 1750 passed; SQLite: 1773 passed; just lint, just typecheck clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01T8wjd6HrtSA2LN9ssC4NzF

Every embedding job builds a fresh search repository, so its per-instance "tables ready" flags start false and setup re-ran CREATE TABLE/INDEX IF NOT EXISTS on every job. CREATE INDEX IF NOT EXISTS takes a SHARE lock before it sees the index exists, so under concurrent writers it queued every chunk and embedding write behind it until queries hit the 30s timeout (one cloud tenant: ~28k embedding job timeouts in two days). Setup now reads the catalog first (to_regclass/pg_attribute, no table locks) and only runs DDL when the chunk table, the embeddings table, or their indexes are missing or out of date. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T8wjd6HrtSA2LN9ssC4NzF Signed-off-by: phernandez <paul@basicmachines.co>

chatgpt-codex-connector · 2026-10-06T00:54:13Z

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-06T01:52:02.140264Z	`1910011`	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7d27cbafa9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-10-06T00:56:50Z

+                            "SELECT to_regclass('search_vector_chunks') IS NOT NULL "
+                            "AND to_regclass('idx_search_vector_chunks_project_entity') "
+                            "IS NOT NULL"


Keep readiness probes in the current schema

When a deployment uses a layered search path such as tenant,public and the tenant schema is being bootstrapped while public already contains these objects, to_regclass searches the entire path, so this probe evaluates true for the public table and index. Unlike the previous unconditional CREATE TABLE/CREATE INDEX, which targets the first writable schema, setup now skips tenant-local creation and later unqualified vector writes fall through to the public tables, mixing search state across schema boundaries. Verify that both OIDs belong to current_schema() before taking the fast path.

AGENTS.md reference: AGENTS.md:L392-L396

Useful? React with 👍 / 👎.

to_regclass searches the whole search_path, so with a layered path like `tenant, public` the vector tables in public satisfied the probe and setup skipped creating the tenant schema's own. The probes now read pg_class for current_schema() only, matching where unqualified DDL and writes land. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T8wjd6HrtSA2LN9ssC4NzF Signed-off-by: phernandez <paul@basicmachines.co>

Runtime code no longer issues DDL. Search and vector-sync paths only bind the vector adapter; PgVectorIndex upsert/search/delete no longer call initialize(). Storage is created by init_search_index() -> create_storage(), which local startup already calls and cloud will call once per tenant per process alongside migrations. A database missing vector storage now fails on its first query instead of creating tables mid-job. create_storage() still reads the catalog first, so initializing an already initialized database (every new worker) takes no locks on the tables. Tests that relied on runtime creation now initialize the database first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T8wjd6HrtSA2LN9ssC4NzF Signed-off-by: phernandez <paul@basicmachines.co>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2b89294ab5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-10-06T01:45:23Z

+            existing_dimensions = await self._existing_dimensions(session)
+            storage_missing = existing_dimensions is None
+            source_hash_missing = (
+                existing_dimensions is not None and not await self._has_source_hash_column(session)


Keep mismatch checks in the target schema

With a layered search path such as tenant,public, when the tenant embedding table is absent but public.search_vector_embeddings has different dimensions or lacks source_hash, the new local _read_storage probe correctly takes the creation path, but _existing_dimensions() and _has_source_hash_column() still resolve the public table. The resulting mismatch reaches the unqualified DROP TABLE IF EXISTS search_vector_embeddings, which resolves to and deletes the public table rather than creating isolated tenant storage. Fresh evidence beyond the earlier comment is that the fast probe is now schema-scoped while these follow-up probes remain search-path-scoped; scope them to current_schema() as well.

Useful? React with 👍 / 👎.

create_storage() already reads pg_class/pg_attribute for current_schema(); the dimension and source_hash checks used separate lookups that searched the whole search_path, so a mismatched same-named table in another schema could be treated as ours and dropped. Reuse the schema-local read and delete the two path-wide helpers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T8wjd6HrtSA2LN9ssC4NzF Signed-off-by: phernandez <paul@basicmachines.co>

chatgpt-codex-connector Bot reviewed Oct 6, 2026

View reviewed changes

phernandez changed the title ~~Skip vector DDL when the storage already exists~~ fix(core): skip vector DDL when the storage already exists Oct 6, 2026

phernandez and others added 2 commits October 5, 2026 20:07

phernandez changed the title ~~fix(core): skip vector DDL when the storage already exists~~ fix(core): create vector storage at initialization, never at runtime Oct 6, 2026

chatgpt-codex-connector Bot reviewed Oct 6, 2026

View reviewed changes

phernandez merged commit dcbf55d into main Oct 6, 2026
51 of 52 checks passed

phernandez deleted the vector-schema-precheck branch October 6, 2026 02:35

saintar1997 mentioned this pull request Oct 6, 2026

fix(core): re-plan SQLite vector preparation under the write lock saintar1997/basic-memory#1

Open

This was referenced Oct 8, 2026

docs(core): bring the v0.24.0 changelog up to date #1680

Merged

Avoid redundant vector index DDL blocking behind ordinary writes #1526

Closed

Navigation Menu

Repository navigation

fix(core): create vector storage at initialization, never at runtime - #1656

fix(core): create vector storage at initialization, never at runtime#1656
phernandez merged 4 commits into
mainfrom
vector-schema-precheck

phernandez commented Oct 6, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot commented Oct 6, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Oct 6, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Oct 6, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

phernandez commented Oct 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Fix: no DDL at runtime

Tests

Uh oh!

chatgpt-codex-connector Bot commented Oct 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codex Review Summary

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

phernandez commented Oct 6, 2026 •

edited

Loading

chatgpt-codex-connector Bot commented Oct 6, 2026 •

edited

Loading