Set default branch for git review/gerrit.
Switch tests to stable.
Switch to using stable charm-helpers branch.
Switch to using stable charm.openstack branch.
Switch to using stable zaza branch.
Switch to using stable zaza-openstack-tests branch.
The network-get --primary-address juju-info fails on pre-2.8.?
versions of juju. This results in a NoNetworkBinding error.
Fallback to unit_get() if that occurs for local_address().
Includes updates to charmhelpers/charms.openstack for cert_utils
and unit-get for the install hook error on Juju 2.9
* charm-helpers sync for classic charms
* rebuild for reactive charms
* ensure tox.ini is from release-tools
* ensure requirements.txt files are from release-tools
* On reactive charms:
- ensure master branch for charms.openstack
- ensure master branch for charm-helpers
* Remove mock for unit_get as it's no longer in c-h context module
* Remove sections from subordinate context due to c-h change.
The database pool size is per-worker and not shared among the workers.
Currently we incorrectly scale max_pool_size with the worker count,
resulting in an excessive number of database connections that are both
unnecessary and can exhaust the MySQL server max_connections.
With the previous formula max_pool_size = workers * 4, the resulting
connection count was exponential e.g. with 10 workers you get 400
connections but with 20 workers you get 1600. With the commonly deployed
setting of worker-multiplier=0.25 you get 20 workers on 40C/80T machines
and a 3 node HA setup was consuming 3 * 1600 = 4800 MySQL connections.
This customisation was added when rpc_workers support was added (commit
b6ff05ddfe) and I cannot find any evidence that a higher pool count is
needed, the upstream neutron project also removed it's own customisation
of 10 in deference to the oslo.db default of (LP: #1682307). So this
change appears safe and will result in only 100 connections for 20
workers instead of 1600.
This change only replaced the deprecated function to add a new source
in the `do_openstack_upgrade` function. The original idea of this patch
was to fix an issue that was fixed in another patch.
Copy the new version of the requirements.txt and test-requirements.txt
files to avoid pip 20.3.
Neutron uses an AZ-unaware scheduler (LeastRoutersScheduler) by default
in its configuration and the neutron-api charm does not override it.
AZLeastRoutersScheduler inherits from LeastRoutersScheduler and does the
same, plus respects AZ hints when scheduling HA routers.
For --distributed --ha routers using AZLeastRoutersScheduler means that
snat namespaces will be scheduled with respect to the AZ hints specified
during router creation by an operator.
For --ha but not distributed routers using AZLeastRoutersScheduler means
that qrouter namespaces will be scheduled with respect to the AZ hints.
snat namespaces (--ha & --distributed) and qrouter namespaces (--ha
only) are placed by the scheduler to l3 agents that run in the dvr_snat
mode only so the scheduler change will affect both the deployments with
neutron-gateway units and the ones with neutron-openvswitch running with
use-dvr-snat=True.