Why this change is needed:
After implementing model isolation, two critical bugs were discovered that would cause data access failures:
Bug 1: In delete_entity_relation(), the SQL query uses positional parameters
($1, $2) but the parameter dict was not converted to a list of values before
passing to db.execute(). This caused parameter binding failures when trying to
delete entity relations.
Bug 2: Four read methods (get_by_id, get_by_ids, get_vectors_by_ids, drop)
were still using namespace_to_table_name(self.namespace) to get legacy table
names instead of self.table_name with model suffix. This meant these methods
would query the wrong table (legacy without suffix) while data was being
inserted into the new table (with suffix), causing data not found errors.
How it solves it:
- Bug 1: Convert parameter dict to list using list(params.values()) before
passing to db.execute(), matching the pattern used in other methods
- Bug 2: Replace all namespace_to_table_name(self.namespace) calls with
self.table_name in the four affected methods, ensuring they query the
correct model-specific table
Impact:
- delete_entity_relation now correctly deletes relations by entity name
- All read operations now correctly query model-specific tables
- Data written with model isolation can now be properly retrieved
- Maintains consistency with write operations using self.table_name
Testing:
- All 6 PostgreSQL migration tests pass (test_postgres_migration.py)
- All 6 Qdrant migration tests pass (test_qdrant_migration.py)
- Verified parameter binding works correctly
- Verified read methods access correct tables
Why this change is needed:
PostgreSQL vector storage needs model isolation to prevent dimension
conflicts when different workspaces use different embedding models.
Without this, the first workspace locks the vector dimension for all
subsequent workspaces, causing failures.
How it solves it:
- Implements dynamic table naming with model suffix: {table}_{model}_{dim}d
- Adds setup_table() method mirroring Qdrant's approach for consistency
- Implements 4-branch migration logic: both exist -> warn, only new -> use,
neither -> create, only legacy -> migrate
- Batch migration: 500 records/batch (same as Qdrant)
- No automatic rollback to support idempotent re-runs
Impact:
- PostgreSQL tables now isolated by embedding model and dimension
- Automatic data migration from legacy tables on startup
- Backward compatible: model_name=None defaults to "unknown"
- All SQL operations use dynamic table names
Testing:
- 6 new tests for PostgreSQL migration (100% pass)
- Tests cover: naming, migration trigger, scenarios 1-3
- 3 additional scenario tests added for Qdrant completeness
Co-Authored-By: Claude <noreply@anthropic.com>
Why this change is needed:
To implement vector storage model isolation for Qdrant, allowing different workspaces to use different embedding models without conflict, and automatically migrating existing data.
How it solves it:
- Modified QdrantVectorDBStorage to use model-specific collection suffixes
- Implemented automated migration logic from legacy collections to new schema
- Fixed Shared-Data lock re-entrancy issue in multiprocess mode
- Added comprehensive tests for collection naming and migration triggers
Impact:
- Existing users will have data automatically migrated on next startup
- New workspaces will use isolated collections based on embedding model
- Fixes potential lock-related bugs in shared storage
Testing:
- Added tests/test_qdrant_migration.py passing
- Verified migration logic covers all 4 states (New/Legacy existence combinations)
Previously, configure_vchordrq would fail silently when probes was empty
(the default), preventing epsilon from being configured. Now each parameter
is handled independently with conditional execution, and configuration
errors fail-fast instead of being swallowed.
This fixes the documented epsilon setting being impossible to use in the
default configuration.
- Add _default_workspace to global vars
- Set _default_workspace to None on cleanup
- Ensure complete resource cleanup
- Fix missing workspace finalization
* Acquire lock before setting ContextVar
* Prevent state corruption on cancellation
* Fix permanent lock brick scenario
* Store context only after success
* Handle acquisition failure properly
• Replace truthy checks with `is not None`
• Handle empty dict edge case properly
• Prevent data reload failures
• Add comprehensive test coverage
• Fix JsonKVStorage and DocStatusStorage
• Reload cleaned data after sanitization
• Update shared memory with clean data
• Add specific surrogate char tests
• Test migration sanitization flow
• Prevent dirty data in memory
- Fast path for clean data (no sanitization)
- Slow path sanitizes during encoding
- Reload shared memory after sanitization
- Custom encoder avoids deep copies
- Comprehensive test coverage
Fixes two compatibility issues in workspace isolation:
1. Problem: lightrag_server.py calls initialize_pipeline_status()
without workspace parameter, causing pipeline to initialize in
global namespace instead of rag's workspace.
Solution: Add set_default_workspace() mechanism in shared_storage.
LightRAG.initialize_storages() now sets default workspace, which
initialize_pipeline_status() uses when called without parameters.
2. Problem: /health endpoint hardcoded to use "pipeline_status",
cannot return workspace-specific status or support frontend
workspace selection.
Solution: Add LIGHTRAG-WORKSPACE header support. Endpoint now
extracts workspace from header or falls back to server default,
returning correct workspace-specific pipeline status.
Changes:
- lightrag/kg/shared_storage.py: Add set/get_default_workspace()
- lightrag/lightrag.py: Call set_default_workspace() in initialize_storages()
- lightrag/api/lightrag_server.py: Add get_workspace_from_request() helper,
update /health endpoint to support LIGHTRAG-WORKSPACE header
Testing:
- Backward compatibility: Old code works without modification
- Multi-instance safety: Explicit workspace passing preserved
- /health endpoint: Supports both default and header-specified workspaces
Related: #2353
Problem:
In multi-tenant scenarios, different workspaces share a single global
pipeline_status namespace, causing pipelines from different tenants to
block each other, severely impacting concurrent processing performance.
Solution:
- Extended get_namespace_data() to recognize workspace-specific pipeline
namespaces with pattern "{workspace}:pipeline" (following GraphDB pattern)
- Added workspace parameter to initialize_pipeline_status() for per-tenant
isolated pipeline namespaces
- Updated all 7 call sites to use workspace-aware locks:
* lightrag.py: process_document_queue(), aremove_document()
* document_routes.py: background_delete_documents(), clear_documents(),
cancel_pipeline(), get_pipeline_status(), delete_documents()
Impact:
- Different workspaces can process documents concurrently without blocking
- Backward compatible: empty workspace defaults to "pipeline_status"
- Maintains fail-fast: uninitialized pipeline raises clear error
- Expected N× performance improvement for N concurrent tenants
Bug fixes:
- Fixed AttributeError by using self.workspace instead of self.global_config
- Fixed pipeline status endpoint to show workspace-specific status
- Fixed delete endpoint to check workspace-specific busy flag
Code changes: 4 files, 141 insertions(+), 28 deletions(-)
Testing: All syntax checks passed, comprehensive workspace isolation tests completed
• Replace truthy checks with `is not None`
• Handle empty dict edge case properly
• Prevent data reload failures
• Add comprehensive test coverage
• Fix JsonKVStorage and DocStatusStorage
• Reload cleaned data after sanitization
• Update shared memory with clean data
• Add specific surrogate char tests
• Test migration sanitization flow
• Prevent dirty data in memory
- Fast path for clean data (no sanitization)
- Slow path sanitizes during encoding
- Reload shared memory after sanitization
- Custom encoder avoids deep copies
- Comprehensive test coverage
• Remove premature ID normalization
• Add lookup mapping for node resolution
• Filter results by requested nodes only
• Improve error logging with workspace
- Batch index existence checks into single query (16+ queries -> 1 query)
- Batch timestamp column checks into single query (8 queries -> 1 query)
- Batch field length checks into single query (5 queries -> 1 query)
Performance improvement: ~70-80% faster initialization (35s -> 5-10s)
Key optimizations:
1. check_tables(): Use ANY($1) to check all indexes at once
2. _migrate_timestamp_columns(): Batch all column type checks
3. _migrate_field_lengths(): Batch all field definition checks
All changes are backward compatible with no schema or API changes.
Reduces database round-trips by batching information_schema queries.
