jiajunwang commented on a change in pull request #362: The WAGED rebalancer
cluster model implementation
URL: https://github.com/apache/helix/pull/362#discussion_r309905879
##########
File path:
helix-core/src/main/java/org/apache/helix/controller/rebalancer/waged/model/AssignableNode.java
##########
@@ -19,10 +19,293 @@
* under the License.
*/
+import org.apache.helix.HelixException;
+import
org.apache.helix.controller.dataproviders.ResourceControllerDataProvider;
+import org.apache.helix.model.ClusterConfig;
+import org.apache.helix.model.InstanceConfig;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.util.Arrays;
+import java.util.Collection;
+import java.util.Collections;
+import java.util.HashMap;
+import java.util.HashSet;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+
+import static java.lang.Math.max;
+
/**
- * A placeholder before we have the implementation.
- *
- * This class represents a potential allocation of the replication.
- * Note that AssignableNode is not thread safe.
+ * This class represents a possible allocation of the replication.
+ * Note that any usage updates to the AssignableNode are not thread safe.
*/
-public class AssignableNode { }
+public class AssignableNode {
+ private static final Logger _logger =
LoggerFactory.getLogger(AssignableNode.class.getName());
+
+ // basic node information
+ private final String _instanceName;
+ private Set<String> _instanceTags;
+ private String _faultZone;
+ private Map<String, List<String>> _disabledPartitionsMap;
+ private Map<String, Integer> _maxCapacity;
+ private int _maxPartition;
+
+ // proposed assignment tracking
+ // <resource name, partition name>
+ private Map<String, Set<String>> _currentAssignments;
+ // <resource name, top state partition name>
+ private Map<String, Set<String>> _currentTopStateAssignments;
+ // <capacity key, capacity value>
+ private Map<String, Integer> _currentCapacity;
+ // runtime usage tracking
+ private int _totalReplicaAssignmentCount;
+ private float _highestCapacityUtilization;
+
+ AssignableNode(ResourceControllerDataProvider clusterCache, String
instanceName,
+ Collection<AssignableReplica> existingAssignment) {
+ _instanceName = instanceName;
+ refresh(clusterCache, existingAssignment);
+ }
+
+ private void reset() {
+ _currentAssignments = new HashMap<>();
+ _currentTopStateAssignments = new HashMap<>();
+ _currentCapacity = new HashMap<>();
+ _totalReplicaAssignmentCount = 0;
+ _highestCapacityUtilization = 0;
+ }
+
+ /**
+ * Update the node with a ClusterDataCache. This resets the current
assignment and recalculate currentCapacity.
+ * NOTE: While this is required to be used in the constructor, this can also
be used when the clusterCache needs to be
+ * refreshed. This is under the assumption that the capacity mappings of
InstanceConfig and ResourceConfig could
+ * subject to changes. If the assumption is no longer true, this function
should become private.
+ *
+ * @param clusterCache - the current cluster cache to initial the
AssignableNode.
+ */
+ private void refresh(ResourceControllerDataProvider clusterCache,
+ Collection<AssignableReplica> existingAssignment) {
+ reset();
+
+ InstanceConfig instanceConfig =
clusterCache.getInstanceConfigMap().get(_instanceName);
+ ClusterConfig clusterConfig = clusterCache.getClusterConfig();
+
+ _currentCapacity.putAll(instanceConfig.getInstanceCapacityMap());
+ _faultZone = computeFaultZone(clusterConfig, instanceConfig);
+ _instanceTags = new HashSet<>(instanceConfig.getTags());
+ _disabledPartitionsMap = instanceConfig.getDisabledPartitionsMap();
+ _maxCapacity = instanceConfig.getInstanceCapacityMap();
+ _maxPartition = clusterConfig.getMaxPartitionsPerInstance();
+
+ assignNewBatch(existingAssignment);
+ }
+
+ /**
+ * Assign a replica to the node.
+ *
+ * @param assignableReplica - the replica to be assigned
+ */
+ void assign(AssignableReplica assignableReplica) {
+ if (!addToAssignmentRecord(assignableReplica, _currentAssignments)) {
+ throw new HelixException(String
+ .format("Resource %s already has a replica from partition %s on this
node",
+ assignableReplica.getResourceName(),
assignableReplica.getPartitionName()));
+ } else {
+ if (assignableReplica.isReplicaTopState()) {
+ addToAssignmentRecord(assignableReplica, _currentTopStateAssignments);
+ }
+ _totalReplicaAssignmentCount += 1;
+ assignableReplica.getCapacity().entrySet().stream()
+ .forEach(entry -> updateCapacityAndUtilization(entry.getKey(),
entry.getValue()));
+ }
+ }
+
+ /**
+ * Release a replica from the node.
+ * If the replication is not on this node, the assignable node is not
updated.
+ *
+ * @param assignableReplica - the replica to be released
+ */
+ void release(AssignableReplica assignableReplica) throws
IllegalArgumentException {
+ String resourceName = assignableReplica.getResourceName();
+ String partitionName = assignableReplica.getPartitionName();
+
+ // Check if the release is necessary
+ if (!_currentAssignments.containsKey(resourceName)) {
+ _logger.warn("Resource " + resourceName + " is not on this node. Ignore
the release call.");
+ return;
+ }
+ Set<String> partitions = _currentAssignments.get(resourceName);
+ if (!partitions.contains(partitionName)) {
+ _logger.warn(String
+ .format("Resource %s does not have a replica from partition %s on
this node",
+ resourceName, partitionName));
+ return;
+ }
+
+ partitions.remove(assignableReplica.getPartitionName());
+ if (assignableReplica.isReplicaTopState()) {
+ _currentTopStateAssignments.get(resourceName).remove(partitionName);
+ }
+ _totalReplicaAssignmentCount -= 1;
+ // Recalculate utilization because of release
+ _highestCapacityUtilization = 0;
+ assignableReplica.getCapacity().entrySet().stream()
+ .forEach(entry -> updateCapacityAndUtilization(entry.getKey(), -1 *
entry.getValue()));
+ }
+
+ public Map<String, Set<String>> getCurrentAssignmentsMap() {
+ return _currentAssignments;
+ }
+
+ public Set<String> getCurrentAssignmentsByResource(String resource) {
+ return _currentAssignments.getOrDefault(resource, Collections.emptySet());
+ }
+
+ public Set<String> getCurrentTopStateAssignmentsByResource(String resource) {
+ return _currentTopStateAssignments.getOrDefault(resource,
Collections.emptySet());
+ }
+
+ public int getTopStateAssignmentTotalSize() {
+ return
_currentTopStateAssignments.values().stream().mapToInt(Set::size).sum();
+ }
+
+ public int getCurrentAssignmentCount() {
+ return _totalReplicaAssignmentCount;
+ }
+
+ public Map<String, Integer> getCurrentCapacity() {
+ return _currentCapacity;
+ }
+
+ public float getHighestCapacityUtilization() {
+ return _highestCapacityUtilization;
+ }
+
+ public String getInstanceName() {
+ return _instanceName;
+ }
+
+ public Set<String> getInstanceTags() {
+ return _instanceTags;
+ }
+
+ public String getFaultZone() {
+ return _faultZone;
+ }
+
+ public Map<String, List<String>> getDisabledPartitionsMap() {
+ return _disabledPartitionsMap;
+ }
+
+ public Map<String, Integer> getMaxCapacity() {
+ return _maxCapacity;
+ }
+
+ public int getMaxPartition() {
+ return _maxPartition;
+ }
+
+ /**
+ * Computes the fault zone id based on the domain and fault zone type when
topology is enabled. For example, when
+ * the domain is "zone=2, instance=testInstance" and the fault zone type is
"zone", this function returns "2".
+ * If cannot find the fault zone id, this function leaves the fault zone id
as the instance name.
+ *
+ * TODO merge this logic with Topology.java tree building logic.
+ * For now, the WAGED rebalancer has a more strict topology def requirement.
+ * Any missing field will cause an invalid topology config exception.
+ */
+ private String computeFaultZone(ClusterConfig clusterConfig, InstanceConfig
instanceConfig) {
+ if (clusterConfig.isTopologyAwareEnabled()) {
+ String topologyStr = clusterConfig.getTopology();
+ String faultZoneType = clusterConfig.getFaultZoneType();
+ if (topologyStr == null || faultZoneType == null) {
+ throw new HelixException("Fault zone or cluster topology information
is not configured.");
+ }
+
+ String[] topologyDef = topologyStr.trim().split("/");
+ if (topologyDef.length == 0 || Arrays.stream(topologyDef)
+ .noneMatch(type -> type.equals(faultZoneType))) {
+ throw new HelixException(
+ "The configured topology definition is empty or does not contain
the fault zone type.");
+ }
+
+ Map<String, String> domainAsMap = instanceConfig.getDomainAsMap();
+ if (domainAsMap == null) {
+ throw new HelixException(String
+ .format("The domain configuration of instance %s is not
configured", _instanceName));
+ } else {
+ StringBuilder faultZoneStringBuilder = new StringBuilder();
+ for (String key : topologyDef) {
+ if (!key.isEmpty()) {
+ if (domainAsMap.containsKey(key)) {
+ faultZoneStringBuilder.append(domainAsMap.get(key));
+ faultZoneStringBuilder.append('/');
+ } else {
+ throw new HelixException(String.format(
+ "The domain configuration of instance %s is not complete.
Type %s is not found.",
+ _instanceName, key));
+ }
+ if (key.equals(faultZoneType)) {
+ break;
+ }
+ }
+ }
+ return faultZoneStringBuilder.toString();
+ }
+ } else {
+ // For backward compatibility
+ String zoneId = instanceConfig.getZoneId();
+ return zoneId == null ? instanceConfig.getInstanceName() : zoneId;
+ }
+ }
+
+ /**
+ * This function should only be used to assign a set of partitions that
doesn't exist before.
+ * Using this function avoids the overhead of updating capacity repeatedly.
+ */
+ private void assignNewBatch(Collection<AssignableReplica> replicas) {
+ Map<String, Integer> totalPartitionCapacity = new HashMap<>();
+ for (AssignableReplica replica : replicas) {
+ addToAssignmentRecord(replica, _currentAssignments);
+ if (replica.isReplicaTopState()) {
+ addToAssignmentRecord(replica, _currentTopStateAssignments);
+ }
+ // increment the capacity requirement according to partition's capacity
configure.
+ for (Map.Entry<String, Integer> capacity :
replica.getCapacity().entrySet()) {
+ totalPartitionCapacity.compute(capacity.getKey(),
+ (k, v) -> (v == null) ? capacity.getValue() : v +
capacity.getValue());
+ }
+ }
+ _totalReplicaAssignmentCount += replicas.size();
+
+ // Update to the global state after all single replications' calculation
is done.
+ for (String key : totalPartitionCapacity.keySet()) {
+ updateCapacityAndUtilization(key, totalPartitionCapacity.get(key));
+ }
+ }
+
+ private boolean addToAssignmentRecord(AssignableReplica replica,
+ Map<String, Set<String>> currentAssignments) {
+ return currentAssignments.computeIfAbsent(replica.getResourceName(), k ->
new HashSet<>())
+ .add(replica.getPartitionName());
+ }
+
+ private void updateCapacityAndUtilization(String capacityKey, int
valueToSubtract) {
+ if (_currentCapacity.containsKey(capacityKey)) {
Review comment:
That basically means this category is unlimited on this node. This shall be
allowed since node capacity and partition weight might be configured in a
different logic. There is no need to require strictly consistent.
If adding any warning log here, the result will be very verbose.
The only restrict now is that node capacity categories have to match the
cluster level definition.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
With regards,
Apache Git Services
