Actively monitor nodes status for DaemonSet deployments #881
Conversation
0be5d4c
to
7b6ba59
Compare
| def find_nodes(cache) | ||
| all_nodes = cache.get_all(Node.kind) | ||
| all_nodes = all_nodes.select { |node_data| node_data.dig('spec', 'unschedulable').to_s.downcase != 'true' } |
There was a problem hiding this comment.
We're looping over the same data three times here. Can we do it just once instead? My ruby might be rusty, but something like:
def find_nodes(cache)
all_nodes = cache.get_all(Node.kind)
all_nodes.each_with_object([]) do |node_data, relevant_nodes|
cond = node_data.dig('status', 'conditions').find { |c| c['type'].downcase == 'ready' }
ready = cond.nil? ? true : cond['status'].downcase == 'true'
schedulable = node_data.dig('spec', 'unschedulable').to_s.downcase != 'true'
if schedulable && ready
relevant_nodes << Node.new(definition: node_data)
end
end
endThere was a problem hiding this comment.
Your ruby is less rusty that mine apparently.
We definitely can.
| ds = build_synced_ds(ds_template: ds_template, pod_templates: pod_templates, node_templates: node_templates) | ||
| refute_predicate(ds, :deploy_succeeded?) | ||
|
|
||
| node_added_status = { |
There was a problem hiding this comment.
This is the same as above and we're not adding a node in this section. Unneeded/misnamed var?
There was a problem hiding this comment.
That's expected, with have the same Pod status, but now one of the node became unschedulable. So it should not block anymore.
node_templates[2]['spec']['unschedulable'] = 'true'
| "numberReady": 2, | ||
| } | ||
| ds_template = build_ds_template(filename: 'daemon_set.yml', status: node_added_status) | ||
| pod_templates = load_fixtures(filenames: ['daemon_set_pods.yml']) |
There was a problem hiding this comment.
similarly, do we need to reload these for some reason or could we just use the ones above?
| "numberReady": 2, | ||
| } | ||
| ds_template = build_ds_template(filename: 'daemon_set.yml', status: node_added_status) | ||
| pod_templates = load_fixtures(filenames: ['daemon_set_pods.yml']) |
There was a problem hiding this comment.
same question as above about these three variables/loads
|
@KnVerey thanks for your feedbacks, I've update the PR accordingly. I'm still tracking down the failing test, not sure yet if it's related or not. |
|
I tested it locally with a |
This PR is a completion of the original implementation in #580
In the original implementation, the list of nodes is fetched when the deployment start and then is fixed. if one of the nodes goes away / becomes unschedulable while daemonsets are rolling out, it keeps waiting indefinitely.
This PR aims to make the node list dynamic.
It also now filters out unschedulable / not ready nodes.
This should replace the
required-rolloutPR #878