No valid host was found, although capacity exists¶
Instances fail to schedule while hypervisors show free CPU, memory and disk. The classic cause on a platform that just added hosts is a compute node that exists but is not mapped to a cell; the others are Placement inventory, disabled services and filters.
Symptom¶
$ openstack server show <server-id> -c status -c fault
status | ERROR
fault | {'code': 500, 'message': 'No valid host was found. There are not enough hosts available.', ...}
OS-EXT-SRV-ATTR:host is empty: no host was ever selected. Meanwhile Placement shows hosts with room
(openstack resource provider usage show <rp-uuid> against its inventory; openstack hypervisor list no longer
shows capacity with recent compute API microversions).
The message depends on where the request stopped (observed on a lab):
No valid host was found. There are not enough hosts available.: a filter rejected every host.No valid host was found.: Placement returned no candidate.Host '<host>' is not mapped to any cell: the scheduler chose a host that is not mapped to a cell. This one names the host instead of saying "No valid host".
Impact¶
Provisioning fails for some or all flavors, projects or availability zones. Running instances are unaffected. Operators waste capacity they paid for; users see random failures if only some hosts are affected.
Quick checks¶
# 1. Are all compute services up and enabled?
openstack compute service list --service nova-compute -c Host -c Zone -c Status -c State
# 2. Are all compute hosts mapped to a cell? (run where nova-manage and nova.conf are available)
nova-manage cell_v2 list_hosts
# Compare with the hosts in step 1. A host present in step 1 and absent here is unmapped.
# 3. Does Placement have candidates for the flavor's resources? (needs the osc-placement plugin)
openstack allocation candidate list --resource VCPU=<vcpus> --resource MEMORY_MB=<ram> --resource DISK_GB=<disk>
# Nova leaves out disabled computes through the COMPUTE_STATUS_DISABLED trait; this command does not. To see
# the candidates as Nova does:
openstack --os-placement-api-version 1.22 allocation candidate list --resource VCPU=<vcpus> \
--resource MEMORY_MB=<ram> --resource DISK_GB=<disk> --forbidden COMPUTE_STATUS_DISABLED
# 4. What did the scheduler say?
# nova-scheduler.log for the request id of the failed boot (openstack server event list <server-id>)
Kolla-Ansible: nova-manage runs inside the nova_api container: sudo docker exec nova_api nova-manage ....
Without -it: it is not needed, and -it fails outside an interactive terminal (cannot attach stdin to a
TTY-enabled container because stdin is not a terminal).
Diagnosis¶
Unmapped host¶
nova-scheduler.log or nova-conductor.log for the request id shows the host being selected, then the conductor
failing with:
Host '<host>' is not mapped to any cell
The scheduler picked the host (it exists in Placement) but the conductor could not find which cell database holds
it. This happens when nova-manage cell_v2 discover_hosts was not run after adding the host and
[scheduler] discover_hosts_in_cells_interval is at its default of -1 (disabled). Kolla-Ansible 22.2.0 writes
discover_hosts_in_cells_interval = -1 in the scheduler's nova.conf, and maps new computes itself when it deploys
them. A host added or re-created outside Kolla-Ansible needs discover_hosts.
Placement has no candidates¶
If step 3 returns nothing, look at the providers:
openstack resource provider list
openstack resource provider inventory list <rp-uuid> # capacity, reserved, allocation_ratio
openstack resource provider usage show <rp-uuid> # what Placement believes is used
Usual findings: allocation_ratio lower than you think (for example cpu_allocation_ratio left at 1.0 on new
hosts), reserved too high, usage inflated by allocations of instances that no longer exist (see
Stale records), or the flavor asks for a
trait or a resource class (PCPU, a custom class) the hosts do not expose.
Filters rejected every candidate¶
The scheduler log line Filter <Name> returned 0 hosts names the filter. Common ones: AggregateInstanceExtraSpecsFilter
(flavor extra specs versus aggregate metadata), ServerGroupAntiAffinityFilter (an anti-affinity group already
has a member on every candidate), AvailabilityZoneFilter (the requested zone has no enabled host),
NUMATopologyFilter (pinned or hugepage flavor with no matching host), ComputeFilter (service down),
ImagePropertiesFilter (image requires a property the hosts lack).
Service disabled or down¶
Step 1 shows disabled (someone disabled it, read the reason) or down (the service stopped reporting: host
problem, or the service lost its message bus connection).
A disabled compute does not reach the filters: Nova sets the COMPUTE_STATUS_DISABLED trait on its resource
provider and asks Placement to leave it out. The scheduler log then says Got no allocation candidates from the
Placement API, while quick check 3 without --forbidden COMPUTE_STATUS_DISABLED still lists the host.
Possible causes¶
| Cause | How to tell |
|---|---|
| Host not mapped to a cell | Conductor log: is not mapped to any cell; host missing from cell_v2 list_hosts |
| Placement inventory or ratios | allocation candidate list empty; inventory shows low ratio or high reserved |
| Stale allocations inflating usage | usage show higher than instances actually on the host; nova-manage placement audit reports orphans |
| Filter rejects everything | Scheduler log names the filter returning 0 hosts |
| Compute service disabled or down | compute service list |
| Flavor asks for a resource class or trait no host has | Candidate list empty for that flavor only; resource provider trait list |
Resolution¶
Map the host¶
⚠️ State-changing. Writes the host mapping to the Nova API database. Reversible (
delete_host). No impact on running instances.
nova-manage cell_v2 discover_hosts --verbose
nova-manage cell_v2 list_hosts
Consider setting [scheduler] discover_hosts_in_cells_interval = 300 on nova-scheduler so that new hosts are
mapped automatically. Verify how your deployment tool handles this before changing the file by hand.
Fix inventory or ratios¶
Set cpu_allocation_ratio, ram_allocation_ratio, disk_allocation_ratio and reserved_host_* in the
[DEFAULT] section of nova.conf on the compute host, restart nova-compute, and check the
inventory again.
Kolla-Ansible: put them in /etc/kolla/config/nova/nova-compute.conf (or /etc/kolla/config/nova.conf), then:
⚠️ State-changing.
reconfigurerecreates thenova_computecontainers whose configuration changed; running instances are not affected.
kolla-ansible reconfigure -i <inventory> --tags nova
openstack resource provider inventory list <rp-uuid> # the new reserved value or ratio, about a minute later
Changing a ratio directly in Placement is overridden by the compute on its next update unless the
initial_*_allocation_ratio logic applies; prefer the Nova configuration.
Heal stale allocations¶
(Not run on this page's lab.)
See Stale records; in short,
nova-manage placement audit to list, nova-manage placement heal_allocations to repair, with the warnings there.
Re-enable a service¶
⚠️ State-changing. Instances will be scheduled on the host again. Make sure the reason it was disabled no longer applies.
openstack compute service set --enable <compute-host> nova-compute
Verification¶
openstack allocation candidate list --resource VCPU=<vcpus> --resource MEMORY_MB=<ram> --resource DISK_GB=<disk> | head
openstack server create --flavor <flavor> --image <image> --network <network> sched-test
openstack server show sched-test -c status -c OS-EXT-SRV-ATTR:host
openstack server delete sched-test
Lab observations¶
On a lab: OpenStack 2026.1, Kolla-Ansible 22.2.0, all-in-one (one compute), ML2/OVS, Docker, Ubuntu 24.04, osc-placement 4.8.0. Each cause was produced on purpose, then repaired:
| Cause produced | Fault | Quick checks | Repair |
|---|---|---|---|
Host mapping deleted (nova-manage cell_v2 delete_host, the host's instances shelved first) |
Host '<host>' is not mapped to any cell, no host |
1: enabled, up. 2: list_hosts empty. 3: one candidate. 4: HostMappingNotFound in the conductor log |
discover_hosts --verbose: Creating host mapping for compute host '<host>'; a test instance ACTIVE |
reserved_host_memory_mb = 15000 on a 16 GB compute |
No candidate in Placement (see note) | 3: empty; inventory list: MEMORY_MB reserved 15000; usage show: 1024 used. Scheduler log: Got no allocation candidates from the Placement API |
Override removed, reconfigure: reserved back to 512, one candidate |
| Compute service disabled | No valid host was found. |
1: disabled. 3: one candidate without --forbidden, none with it. Scheduler log: Got no allocation candidates from the Placement API |
compute service set --enable; a test instance ACTIVE |
| A second member of an anti-affinity group, one host | No valid host was found. There are not enough hosts available. |
Scheduler log: Filter ServerGroupAntiAffinityFilter returned 0 hosts, and the filter results |
Members deleted |
Note: during the reserved-memory test, the test instance ended with MessagingTimeout, not with "No valid host".
The scheduler answered 28 s after the conductor had stopped waiting, and its log shows the
Got no allocation candidates answer for that request. Why it answered late was not established: reconfigure had
not restarted it.
Validity and limits¶
Derived from an operator's account of "capacity exists but nothing lands on it"; release not recorded. Observed on
one all-in-one lab on 2026.1 (see Lab observations): the unmapped host, a reserved value in the inventory, a
disabled service and a filter, with their repairs. Not run: stale allocations and heal_allocations, a resource
class or trait no host has, a down service, and anything that needs several computes. Cells v2
and Placement have been mandatory since Ocata (15.0), so the mapping mechanism applies to every current release.
The osc-placement plugin is required for the openstack resource provider and allocation candidate commands.
Not covered: scheduling failures specific to Ironic bare-metal nodes.
Related¶
- Instance stuck in BUILD waiting for vif-plugged: the failure that happens after a host was selected.
- Stale records.
- Source: OpenStack, the Day After Tomorrow (S. Zaouam), chapter 5.
Provenance: written by agent:ozarland-author, from the book OpenStack, the Day After Tomorrow. Reviewed in its pull request (see the page history). Checked by a human operator: not yet.
Worked for you, or not? Tell us in one line: your release, your deployment method, and what happened.