Due to change I8d426f2635232ffc4b510548a905794ca88d7f99 in Pike,
which ironically was meant to avoid up-calls (I think), it
introduced an up-call during reschedules for server create and
resize to set the instance.availability_zone based on the
alternate host selected during the reschedule.
This adds the up-call to our list of known issues in the cells
v2 docs so we can track the issue and make people aware of it.
Change-Id: Id819f91477613a013b89b1fb0b2def3b0fd4b08c
Related-Bug: #1781286
As of Rocky, the cell databases are being used for storing console
authorizations, so the layout for running console proxies has
changed from global to a deployment to per cell.
Part of blueprint convert-consoles-to-objects
Change-Id: If5c7b081682f0cb36a678d0e21e26b7985cd8527
This ensures we have version-specific references to other projects [1].
Note that this doesn't mean the URLs are actually valid - we need to do
more work (linkcheck?) here, but it's an improvement nonetheless.
[1] https://docs.openstack.org/openstackdocstheme/latest/#external-link-helper
Change-Id: Ifb99e727110c4904a85bc4a13366c2cae300b8df
Change I2c492c46e85c1df96aa7fdc12cdee0b1c7ba775e removed the
need to the host aggregate up-call when doing non-block
live migration in the xenapi driver. This change adds a note
in the cells v2 layout docs saying the upcall is resolved.
Related to blueprint live-migration-in-xapi-pool
Change-Id: I7fe6ed428c77a0f719afb957d74e517f1ab9a991
The affinity and anti-affinity server group policy is enforced by the
scheduler but two parallel scheduling could cause that such policy is
violated. During instance boot a late policy check was performed in
the compute manager to prevent this. This check was missing in case
of rebuild. Therefore two parallel evacuate command could cause that
the server group policy is violated. This patch introduces the late
policy check to rebuild to prevent such situation. When the violation
is detected during boot a re-scheduling happens. However the rebuild
action does not have the re-scheduling implementation so in this case
the rebuild will fail and the evacuation needs to be retried by the
user. Still this is better than allowing a parallel evacuation to
break the server group affinity policy.
To make the late policy check possible in the compute/manager the
rebuild_instance compute RPC call was extended with a request_spec
parameter.
Co-Authored-By: Richard Zsarnoczai <richard.zsarnoczai@ericsson.com>
Change-Id: I752617066bb2167b49239ab9d17b0c89754a3e12
Closes-Bug: #1735407
The instance list performance and reschedules issues have
been fixed in Queens so this updates the caveats section
of the cells v2 layout doc to point out those changes since
Pike.
Change-Id: I882f185554b6d2781534fa93e41e5010ea3a641d
The cells v2 layout documentation clearly states that there are no
upcalls from cells back to the central API services. This mislead
me for sometime as I could not fathom how a compute node in a cell
was supposed to report its resource info.
It turns out nova looks up the placement service in the keystone
catalogue and contacts it directly which to my mind is an upcall. I
wonder if the author of the not felt that the placement service is
not really part of nova?
Change-Id: If14be8b182f0af4e4e6641046fec638c07e26546
Closes-Bug: #1742421
Underscores make for really ugly URLs and every other document here is
now using hyphens for this very reason. Let's try to keep that going.
Change-Id: I5c99ff6b04ee97bac210a0d6762015225775c5ee
Matt Riedemann