Region Placement Policy V1.0¶

S1: CLUSTER_SCALE_IN¶

The placement policy first checks if there are policy decisions from other policies by looking into the deletion field of the action’s data property. If there is such a field, the policy attempts to extract the count value from the deletion field. If the count value is not found, 1 is assumed to be the default.

If, however, the policy fails to find the deletion field, it tries to find if there is a count field in the action’s inputs property. If the answer is true, the policy will use it, or else it will fall back to assume 1 as the default count.

After the policy has find out the count value (i.e. number of nodes to be deleted), it validates the list of region names provided to the policy. If for some reason, none of the provided names passed the validation, the policy check fails with the following data recorded in the action’s data property:

{
  "status": "ERROR",
  "reason": "No region is found usable.",
}

With the list of regions known to be good and the map of node distribution specified in the policy spec, senlin engine continues to calculate a placement plan that best matches the desired distribution.

If there are nodes that cannot be fit into the distribution plan, the policy check failed with an error recorded in the action’s data, as shown below:

{
  "status": "ERROR",
  "reason": "There is no feasible plan to handle all nodes."
}

If there is a feasible plan to remove nodes from each region, the policy saves the plan into the data property of the action as exemplified below:

{
  "status": "OK",
  "deletion": {
    "count": 3,
    "regions": {
      "RegionOne": 2,
      "RegionTwo": 1
    }
  }
}

This means in total, 3 nodes should be removed from the cluster. Among them, 2 nodes should be selected from region “RegionOne” and the rest one should be selected from region “RegionTwo”.

NOTE: When there is a deletion policy attached to the same cluster. That deletion policy will be evaluated after the region placement policy and it is expected to rebase its candidate selection on the region distribution enforced here. For example, if the deletion policy is tasked to select the oldest nodes for deletion, it will adapt its behavior to select the oldest nodes from each region. The number of nodes to be chosen from each region would be based on the output from this placement policy.

S2: CLUSTER_SCALE_OUT¶

The placement policy first checks if there are policy decisions from other policies by looking into the creation field of the action’s data property. If there is such a field, the policy attempts to extract the count value from the creation field. If the count value is not found, 1 is assumed to be the default.

If, however, the policy fails to find the creation field, it tries to find if there is a count field in the action’s inputs property. If the answer is true, the policy will use it, or else it will fall back to assume 1 as the default node count.

After the policy has find out the count value (i.e. number of nodes to be created), it validates the list of region names provided to the policy and extracts the current distribution of nodes among those regions.

If for some reason, none of the provided names passed the validation, the policy check fails with the following data recorded in the action’s data property:

{
  "status": "ERROR",
  "reason": "No region is found usable.",
}

The logic of generating a distribution plan is almost identical to what have been described in scenario S1, except for the output format. When there is a feasible plan to accommodate all nodes, the plan is saved into the data property of the action as shown in the following example:

{
  "status": "OK",
  "creation": {
    "count": 3,
    "regions": {
      "RegionOne": 1,
      "RegionTwo": 2
    }
  }
}

This means in total, 3 nodes should be created into the cluster. Among them, 2 nodes should be created at region “RegionOne” and the left one should be created at region “RegionTwo”.

S3: CLUSTER_RESIZE¶

The placement policy first checks if there are policy decisions from other policies by looking into the creation field of the action’s data property. If there is such a field, the policy extracts the count value from the creation field. If the creation field is not found, the policy tries to find if there is a deletion field in the action’s data property. If there is such a field, the policy extracts the count value from the creation field. If neither creation nor deletion is found in the action’s data property, the policy proceeds to parse the raw inputs of the action.

The output from the parser may indicate an invalid combination of input values. If that is the case, the policy check fails with the action’s data set to something like the following example:

{
  "status": "ERROR",
  "reason": <error message from the parser.>
}

If the parser successfully parsed the action’s raw inputs, the policy tries again to find if there is either creation or deletion field in the action’s data property. It will use the count value from the field found as the number of nodes to be handled.

When the placement policy finds out the number of nodes to create (or delete), it proceeds to calculate a distribution plan. If the action is about growing the size of the cluster, the logic and the output format are the same as that have been outlined in scenario S2. Otherwise, the logic and the output format are identical to that have been described in scenario S1.

S4: NODE_CREATE¶

When handling a NODE_CREATE action, the region placement policy only needs to deal with the node associated with the action. If, however, the node is referencing a profile which has a region_name specified in its spec, this policy will avoid choosing deployment region for the node. In other words, the region_name specified in the profile spec used takes precedence.

If the profile spec doesn’t specify a region name, this placement policy will proceed to do an evaluation of current region distribution followed by a calculation of a distribution plan. The logic and the output format are the same as that in scenario S2, although the number of nodes to handle is one in this case.