|
Fixes bug 1128605
The dbpool code in sqlalchemy session is the wrong place to implement
thread pooling as it wraps each individual SQL call to run in its own
thread. When combined with SQL server locking, all threads can be eaten
waiting on locks with none available to run a 'COMMIT'.
The correct place to do thread pooling is around each DB API call.
This patch removes dbpool from sqlalchemy and creates a common DB API
loader for all openstack projects which implements the following
configuration options:
db_backend: Full path to DB API backend module (or a known short name if
a project chooses to implement a mapping)
dbapi_use_tpool: True or False whether to use thread pooling around all
DB API calls.
DB backend modules must implement a 'get_backend()' method.
Example usage for nova/db/api.py would be:
"""
from nova.openstack.common.db import api as db_api
_KNOWN_BACKENDS = {'sqlalchemy': 'nova.db.sqlalchemy.api'}
IMPL = db_api.DBAPI(backend_mapping=_KNOWN_BACKENDS)
"""
NOTE: Enabling thread pooling will be broken until this issue is
resolved in eventlet _OR_ until we modify our eventlet.monkey_patch()
calls to include 'thread=False':
https://bitbucket.org/eventlet/eventlet/issue/137/
Change-Id: Idf14563ea07cf8ccf2a77b3f53659d8528927fc7
|