"ERROR: more units requested" when --hyperconverged is used for OpenStack

[marino-mrc]

Hello,
I'm trying to run the script with the --hyperconverged flag and I'm getting the following error:

$ ./generate-bundle.sh -r ussuri -s focal --ceph --ml2-ovs --default-binding oam  --hyperconverged -n teststs
Juju model 'teststs' already exists and is the current context - skipping create

Created focal-ussuri bundle and overlays:
  + openstack/glance.yaml
  + openstack/keystone.yaml
  + openstack/neutron-gateway.yaml
  + openstack/neutron-openvswitch.yaml
  + openstack/rabbitmq-source.yaml
  + ceph/ceph.yaml
  + openstack/openstack-ceph.yaml
  + openstack/neutron-ml2dns.yaml
  + openstack/neutron-ml2dns-gateway.yaml
  + openstack/neutron-ml2dns-openvswitch.yaml
  + mysql-innodb-cluster.yaml
  + mysql-innodb-cluster-router.yaml
  + openstack/placement.yaml

ERROR: more units requested (3) that machines available (2) (template=mysql.yaml) - hint: add more compute nodes/units

I think the problem is in common/generate_bundle_base when ${num_placement_machines} is calculated.
Actually, we have the following:

num_placement_machines=$((${MOD_PARAMS[__NUM_NEUTRON_GATEWAY_UNITS__]} + ${MOD_PARAMS[__NUM_COMPUTE_UNITS__]}))

But this is conceptually wrong because it doesn't consider the number of controller units (I'm assuming 3 because a MySQL cluster is the default) and miscalculates the minimum number of ceph-osd units (think about the case where the number of compute units is less than 3. We still need at least 3 ceph-osd units).

Assuming that we want to use hyper-convergency for storage and compute, we should have the following pseudo-code:

num_controller_units = 3  # We have 3 machines, and on top of them many LXD containers

# Check if the number of compute units has been defined, and use the appropriate number of num_ceph_osd_units
if num_compute_units < 3
num_ceph_osd_units = 3
else
num_ceph_osd_units = num_compute_units
end

num_placement_machines = num_controller_units + num_ceph_osd_units + num_neutron_gateway_units

Also, I tried to work around the code using a fixed value and I noticed that compute and ceph-osd units are misplaced (LXD containers of the control plane are started on top of compute units)

Regards,
Marco

[dosaboy]

@marino-mrc that message is printed when the bundle being deployed does not have enough physical machines to accommodate 3 units of a charm. This is because the default value of num computes is 1. A simple way to fix this would be to default to 3 computes if the deployments is hyperconverged. What do you think?

[marino-mrc]

@dosaboy I think your solution doesn't work because even with 3 compute nodes, we still have an insufficient number of units. I mean, the script could work, but incompatible services will be placed on the same units.

Some assumptions should be made for the controller (I'm trying to align with what I think was the initial idea when --hyperconverged is used in the stsstack-bundles project):

• MySQL is always installed in a cluster, this requires 3 units (which cannot overlap with Ceph-OSD or Compute or neutron-gateway IMO)
• The number of neutron-gateway units is always 1 by default (and cannot be changed from the CLI)
• The neutron-gateway unit cannot be installed on the same units of ceph-osd and controllers
• Compute and ceph-osd will be installed on the same unit by default (and this cannot be changed from the CLI)
• If the number of compute units is less than 3, we still need 3 units for ceph-osd. Note that the number of compute units can be passed through the CLI

Under the previous assumptions, the minimum number of units is 7 and 3 of them MUST support LXD containers (I'm assuming --hyperconverged and --ml2-ovs):

• 3 for the control plane (mysql + OpenStack related services)
• 1 for neutron-gateway
• 3 for ceph-osds (and eventually for the same number of nova-compute units)

If you agree with my vision, the following pseudo-code works (only if you use --hyperconverged and --ml2-ovs):

#define a static value for the number of controller units
num_controller_units = 3  # We have 3 machines, and on top of them many LXD containers
num_neutron_gateway_units = 1  # We don't need this in the code. It's defined in openstack/pipeline/02configure. This line is just for clarity

# Check if the number of compute units has been defined, and use the appropriate number of num_ceph_osd_units
if num_compute_units < 3
  num_ceph_osd_units = 3
else
  num_ceph_osd_units = num_compute_units
end

num_placement_machines = num_controller_units + num_ceph_osd_units + num_neutron_gateway_units

I tried to set the value to 7 and the script works without any error.
But again, the units are misplaced (for example: I have MySQL units on the same machine where nova-compute is running). This should be fixed later because additional work is needed.

EDIT: The previous pseudo-code should be added in substitution of line 182

[dosaboy]

So, the placement of the units is something that you can change to suite your needs by modifying the unit_placement.yaml file. After you make changes to that file you need to use the --replay option to avoid overwriting your changes. Whenever I do a hyperconverged deployment I have to do this anyway because I have to manually set the machine constraints so that MAAS deploys using the correct machines. Does that not work for you?

[marino-mrc]

@dosaboy It works for me. Based on your suggestion, the current workaround is to simply pass --num-compute 3 to the CLI:

./generate-bundle.sh -r ussuri -s focal --ceph --ml2-ovs --default-binding oam  --hyperconverged --num-compute 3 -n teststs

Just tested and it works.

This should be summarized by saying that if --hyperconverged is used, --num-compute is forced to 3.

Thanks!