[openstack-dev] [nova] Pike PTG recap - cells
mriedemos at gmail.com
Mon Feb 27 22:33:20 UTC 2017
We talked about cells on Wednesday morning at the PTG. The full etherpad
is here .
We talked a bit about what needs to happen for this, and it starts with
getting the data into searchlight so that it can serve the REST API,
which is being worked in this blueprint . We want to get that done
early in Pike.
We plan on making the use of Searchlight configurable in Nova since at
first you might not even have anything in it, so listing instances
wouldn't work. We're also going to attempt to merge-sort when listing
instances across multiple cells but it's going to be a known issue that
it will be slow.
For testing Nova with Searchlight, we need to start by enabling the
Searchlight devstack plugin in the nova-next CI job, which I'll work on.
I'm going to talk to Kevin Zheng about seeing if he can spend some time
on getting Nova to use Searchlight if it's (1) configured for use and
(2) is available (the endpoint is in the service catalog). Kevin is a
Searchlight core and familiar with the Nova API code, so he's a good
candidate for working on this (assuming he's available and willing to
Cells-aware gaps in the API
Dan Smith has started a blueprint  for closing gaps in the API which
break in a multi-cell deployment. He has a test patch  to expose the
failures and then they can be worked on individually. The pattern of the
work is in . Help is welcome here, so please attend the weekly cells
meeting  if you want to help out.
Auto-discovery of compute hosts
The "discover_hosts_in_cells_interval" config option was introduced in
Ocata which controls a periodic task in the scheduler to discover new
unmapped compute hosts but it's not very efficient since it queries all
cell mappings and then all compute nodes in each cell mapping and checks
to see if those compute nodes are yet mapped to the cell in the nova_api
database. Dan Smith has a series of changes  which should make that
discovery process more efficient, it just needs to be cleaned up a bit.
Dan Smith is working on a series of changes in both Nova and devstack
for testing with multiple cells . The general idea is that there will
still be two nodes and two nova-compute services. There will be three
nova-conductor services, one per cell, and then another top-level "super
conductor" which is there for building instances and sending the server
create down to one of the cells. All three conductors are going to be
running in the subnode just to balance the resources a bit otherwise the
primary node is going to be starved. The multi-cell job won't be running
migration tests since we don't currently support instance move
operations between cells. We're going to work a hack into the scheduler
to restrict a move operation to the same cell the instance is already
in. This means the live migration job will still be a single-cell setup
where both nova-computes are in the same cell.
Getting rid of nova-consoleauth
There is an unfinished blueprint  from Paul Murray which melwitt is
going to pick up for Pike. The idea is to move the tokens into the
database so we don't care where the consoleauth service lives and then
we can also kill the service.
More information about the OpenStack-dev