Description of problem:
I installed OpenShift 3.5 and logging's installation worked fine from scratch. I then deleted the logging project and re-ran the logging installation (to test something else). 

The logging project is not created with a null nodeselector in this case, which subsequently causes component deployment to fail.

Version-Release number of selected component (if applicable):
openshift v3.5.5.26
kubernetes v1.5.2+43a9be4

Here are some relevant snippets from the installer config:
osm_default_node_selector='env=user'
openshift_hosted_logging_curator_nodeselector='env=infra'
openshift_hosted_logging_elasticsearch_nodeselector='env=infra'
openshift_hosted_logging_kibana_nodeselector='env=infra'

Here is the project yaml that is created:
apiVersion: v1
kind: Project
metadata:
  annotations:
    openshift.io/description: ""
    openshift.io/display-name: ""
    openshift.io/sa.scc.mcs: s0:c10,c0
    openshift.io/sa.scc.supplemental-groups: 1000090000/10000
    openshift.io/sa.scc.uid-range: 1000090000/10000
  creationTimestamp: 2017-07-12T20:05:17Z
  name: logging
  resourceVersion: "8970"
  selfLink: /oapi/v1/projects/logging
  uid: 6d14c3f6-673d-11e7-8b26-123e6dadc042
spec:
  finalizers:
  - openshift.io/origin
  - kubernetes
status:
  phase: Active


Here is an example problem in the deployment (fluentd)
Status:                 Failed
Reason:                 MatchNodeSelector
Message:                Pod Predicate MatchNodeSelector failed

nodeSelector:
    env: user
    fluentd: "true"

Notice that the resulting nodeselector includes the system default, because the project doesn't override the selector with null.

I am not sure why this works on install from scratch (perhaps the default nodeselector comes in way later), but running the playbook directly is broken:

ansible-playbook -i /etc/ansible/hosts /usr/share/ansible/openshift-ansible/playbooks/byo/openshift-cluster/openshift-logging.yml -e openshift_logging_install_logging=true

You can see here that "create project" is run simply:
https://github.com/openshift/openshift-ansible/blob/master/roles/openshift_hosted_logging/tasks/deploy_logging.yaml#L26-L29

Unfortunately, this will inherit the default nodeselector.

It appears that we need to modify the create project behavior to specify that there should be a null nodeselector:

{{ openshift.common.client_binary }} adm --config={{ mktemp.stdout }}/admin.kubeconfig new-project logging --node-selector=""

Again - I am not sure why this works with a from-scratch installation, but it definitely does not work with a post-install logging installation.