- Pre-fetch user group IDs in get_*_by_user_id methods across models layer
- Pass user_group_ids to has_access to avoid repeated group queries
- Reduce query count from 1+N to 1+1 pattern for access control validation
- Apply consistent optimization across knowledge, models, notes, prompts, and tools
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
- Pre-fetch user group IDs once per request in get_tools endpoint
- Pass user_group_ids to has_access to avoid repeated group queries
- Optimize access control validation from 1+N to 1+1 query pattern
- Reduce database load when checking multiple tools access permissions
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
- Replace individual user queries with batch fetching
- Use single query to fetch all required users at once
- Implement O(1) user lookup with dictionary mapping
- Reduce query count from 1+N to 1+1 pattern for tools listing
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
- Replace individual user queries with batch fetching
- Use single query to fetch all required users at once
- Implement O(1) user lookup with dictionary mapping
- Reduce query count from 1+N to 1+1 pattern for models with base_model_id
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
- Replace individual user queries with batch fetching
- Use single query to fetch all required users at once
- Implement O(1) user lookup with dictionary mapping
- Reduce query count from 1+N to 1+1 pattern
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
- Replace individual user queries with batch fetching
- Use single query to fetch all required users at once
- Implement O(1) user lookup with dictionary mapping
- Reduce query count from 1+N to 1+1 pattern
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
- Replace Request object with user.id in cache key for get_all_models
- Request objects are new instances per HTTP request, preventing cache hits
- Cache keys now use user.id ensuring proper cache functionality
- Affects both Ollama and OpenAI model list endpoints
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
Guarantees the last partial delta chunk is emitted when the SSE stream closes (EOF, break, or `[DONE]` sentinel).
* Buffer `last_delta_data` and track `delta_count`
* Flush automatically once `delta_count >= chunk_size`
* Perform a final flush after the iterator ends
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
Every call to get_redis_connection() spawned a new pool, so workers slowly accumulated thousands of open sockets. Even though connections were eventually released, skewed release timing still pushed us past Redis’ max-clients and the cluster egress IP cap.
A module-level _CONNECTION_CACHE now memoises pools by (redis_url, sentinel_hosts, async_mode, decode_responses).
Result: flat connection count, no more IP or FD exhaustion. Public API unchanged.
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
- Implement SentinelRedisProxy class with automatic master discovery
- Add retry logic for handling connection failures and read-only errors
- Support both async and sync Redis operations with Sentinel
- Ensure backward compatibility with existing Redis configurations
- Provide seamless failover during master node outages
This enhancement significantly improves system reliability by eliminating
single points of failure in Redis deployments and ensuring continuous
service availability during infrastructure issues.
Signed-off-by: Sihyeon Jang <sihyeon.jang@navercorp.com>
Sihyeon Jang