…olding

Updated 6 high-traffic list endpoints in main.py to use fresh_db_session()
instead of Depends(get_db), preventing database connections from being held
during response serialization.

Endpoints modified:
- list_servers
- list_a2a_agents
- list_tools
- list_resources
- list_prompts
- list_gateways

Each endpoint now:
1. Uses fresh_db_session() context manager for short-lived DB access
2. Converts ORM objects to dicts (model_dump) before session closes
3. Returns dict representations to avoid lazy loading after session close

Closes #2323

Signed-off-by: Mihai Criveti <crivetimihai@gmail.com>

Added set_fresh_session_factory() and reset_fresh_session_factory() functions
to allow tests to redirect fresh_db_session() to use a test database instead
of the production database.

Updated tests/e2e/test_main_apis.py to use this mechanism, fixing the
"no such table" errors that occurred when fresh_db_session() was introduced
in the endpoint handler fix.

Signed-off-by: Mihai Criveti <crivetimihai@gmail.com>

Updated tests/e2e/test_oauth_protected_resource.py to also override the
fresh_db_session factory, ensuring consistency across all e2e tests that
use test databases.

Signed-off-by: Mihai Criveti <crivetimihai@gmail.com>

Apply the fresh_db_session() pattern to prevent connection pool exhaustion
under high load (Issue #2323, #2330):

RPC handler (main.py):
- Remove Depends(get_db) from handle_rpc() signature
- Wrap all database operations in fresh_db_session() context managers
- For tools/call, connection is released before downstream HTTP calls
- Affects: tools/list, resources/list, prompts/list, tools/call,
  resources/read, sampling/createMessage, completion/complete, etc.

Token scoping middleware:
- Use fresh_db_session() for team membership checks
- Use fresh_db_session() for resource ownership checks
- Connection released before call_next() processes the request

docker-compose.yml:
- Increase DB_POOL_SIZE from 20 to 80 per worker
- Increase DB_MAX_OVERFLOW from 10 to 20
- Adjust PostgreSQL and pgBouncer timeout settings

This prevents the pattern where connections were held for the entire
request lifecycle (including 500-5000ms downstream HTTP calls), which
caused QueuePool exhaustion and system crashes under load.

Signed-off-by: Mihai Criveti <crivetimihai@gmail.com>

Load testing with 4000 Locust users revealed gateway workers being OOM
killed when container memory exceeded 8GB limit.

Memory analysis:
- Each worker uses ~350-450MB RSS baseline, spiking to 0.7-1.9GB under load
- 16 workers × 450MB = 7.2GB baseline, only 0.8GB headroom → OOM kills
- 8 workers × 450MB = 3.6GB baseline, 4.4GB headroom → safe

Changes:
- GRANIAN_WORKERS: 16 → 8
- GRANIAN_BACKPRESSURE: 128 → 256 (maintains 2048 concurrent capacity)

With 3 replicas, total worker count is 24 (was 48), which is still
sufficient for high concurrency while avoiding memory pressure.

Related: #2334, #2323, #2330
Signed-off-by: Mihai Criveti <crivetimihai@gmail.com>