[feature request] Optimize UnitedDeployment replicas settings

[veophi]

What would you like to be added:
Currently, UnitedDeployment requires the sum of subsets[x].replicas must equal to spec.replicas. Actually, it is very clumsy.

We expect that：

• If spec.replicas is not set (nil), subsets[i] just follows subsets[i].replicas (should validate all of subsets replicas are not nil);
• If spec.replicas is set (not nil), we should allow at most ONE subset replicas is nil, and it should equals to spec.replicas - <sum of other subset replicas>;

[veophi]

/label first-good-issue

[kruise-bot]

@veophi: The label(s) /label first-good-issue cannot be applied. These labels are supported: ``. Is this label configured under labels -> additional_labels or `labels -> restricted_labels` in `plugin.yaml`?

In response to this:

/label first-good-issue

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository.

[y-ykcir]

I want to try this, can you assign this to me?

[y-ykcir]

@veophi Since Subsets[i].replicas can be set to a percentage, if spec.replicas is not set (nil), should we not allow Subsets[i].replicas to be set to a percentage?

[y-ykcir]

@veophi I browsed the code here, and it shows that all subsets which are not set replicas will be averagely allocated by the rest replicas, this means that the current logic allows more than two subset.replicas to be unset. Should we change this logic to only allow at most ONE subset replicas is nil? If so, the code for these average logic may need to be modified. PTAL, thanks.

[veophi]

@veophi Since Subsets[i].replicas can be set to a percentage, if spec.replicas is not set (nil), should we not allow Subsets[i].replicas to be set to a percentage?

No, we should reject this configuration

[veophi]

@veophi I browsed the code here, and it shows that all subsets which are not set replicas will be averagely allocated by the rest replicas, this means that the current logic allows more than two subset.replicas to be unset. Should we change this logic to only allow at most ONE subset replicas is nil? If so, the code for these average logic may need to be modified. PTAL, thanks.

@y-ykcir Sorry, I didn't notice this code before. I think the logic of this code is reasonable and doesn't need any modification. What do you think?

[y-ykcir]

@veophi I browsed the code here, and it shows that all subsets which are not set replicas will be averagely allocated by the rest replicas, this means that the current logic allows more than two subset.replicas to be unset. Should we change this logic to only allow at most ONE subset replicas is nil? If so, the code for these average logic may need to be modified. PTAL, thanks.

@y-ykcir Sorry, I didn't notice this code before. I think the logic of this code is reasonable and doesn't need any modification. What do you think?

@veophi I think if spec.replicas is set, the current logic of averaging may be more reasonable. Additionally, it seems that the original logic did not mention the situation where spec.replicas was not set as mentioned in the issue, but instead defaulted to 1. Maybe we can add optimization for this situation?

[veophi]

I think if spec.replicas is set, the current logic of averaging may be more reasonable. Additionally, it seems that the original logic did not mention the situation where spec.replicas was not set as mentioned in the issue, but instead defaulted to 1. Maybe we can add optimization for this situation?

@y-ykcir I agree.