somandal commented on code in PR #15617:
URL: https://github.com/apache/pinot/pull/15617#discussion_r2084759396
##########
pinot-controller/src/main/java/org/apache/pinot/controller/helix/core/rebalance/TableRebalancer.java:
##########
@@ -1524,67 +1535,276 @@ private static void handleErrorInstance(String
tableNameWithType, String segment
}
}
+ /**
+ * Uses the default LOGGER
+ */
+ @VisibleForTesting
+ static Map<String, Map<String, String>> getNextAssignment(Map<String,
Map<String, String>> currentAssignment,
+ Map<String, Map<String, String>> targetAssignment, int
minAvailableReplicas, boolean enableStrictReplicaGroup,
+ boolean lowDiskMode, int batchSizePerServer,
Object2IntOpenHashMap<String> segmentPartitionIdMap,
+ PartitionIdFetcher partitionIdFetcher) {
+ return getNextAssignment(currentAssignment, targetAssignment,
minAvailableReplicas, enableStrictReplicaGroup,
+ lowDiskMode, batchSizePerServer, segmentPartitionIdMap,
partitionIdFetcher, LOGGER);
+ }
+
/**
* Returns the next assignment for the table based on the current assignment
and the target assignment with regard to
* the minimum available replicas requirement. For strict replica-group
mode, track the available instances for all
* the segments with the same instances in the next assignment, and ensure
the minimum available replicas requirement
* is met. If adding the assignment for a segment breaks the requirement,
use the current assignment for the segment.
+ * <p>
+ * For strict replica group routing only (where the segment assignment is
not StrictRealtimeSegmentAssignment)
+ * if batching is enabled, the instances assigned for the same partitionId
can be different for different segments.
+ * For strict replica group routing with StrictRealtimeSegmentAssignment on
the other hand, the assignment for a given
+ * partitionId will be the same across all segments. We can treat both cases
similarly by creating a mapping from
+ * partitionId -> unique set of instance assignments -> currentAssignment.
With StrictRealtimeSegmentAssignment,
+ * this map will have a single entry for 'unique set of instance
assignments'.
+ * <p>
+ * TODO: Ideally if strict replica group routing is enabled then
StrictRealtimeSegmentAssignment should be used, but
+ * this is not enforced in the code today. Once enforcement is added,
then the routing side and assignment side
+ * will be equivalent and all segments belonging to a given
partitionId will be assigned to the same set of
+ * instances. Special handling to check each group of assigned
instances can be removed in that case. The
+ * strict replica group routing can also be utilized for OFFLINE
tables, thus StrictRealtimeSegmentAssignment
+ * also needs to be made more generic for the OFFLINE case.
*/
- @VisibleForTesting
- static Map<String, Map<String, String>> getNextAssignment(Map<String,
Map<String, String>> currentAssignment,
+ private static Map<String, Map<String, String>>
getNextAssignment(Map<String, Map<String, String>> currentAssignment,
Map<String, Map<String, String>> targetAssignment, int
minAvailableReplicas, boolean enableStrictReplicaGroup,
- boolean lowDiskMode) {
- return enableStrictReplicaGroup ?
getNextStrictReplicaGroupAssignment(currentAssignment, targetAssignment,
- minAvailableReplicas, lowDiskMode)
+ boolean lowDiskMode, int batchSizePerServer,
Object2IntOpenHashMap<String> segmentPartitionIdMap,
+ PartitionIdFetcher partitionIdFetcher, Logger tableRebalanceLogger) {
+ return enableStrictReplicaGroup
+ ? getNextStrictReplicaGroupAssignment(currentAssignment,
targetAssignment, minAvailableReplicas, lowDiskMode,
+ batchSizePerServer, segmentPartitionIdMap, partitionIdFetcher,
tableRebalanceLogger)
: getNextNonStrictReplicaGroupAssignment(currentAssignment,
targetAssignment, minAvailableReplicas,
- lowDiskMode);
+ lowDiskMode, batchSizePerServer);
}
private static Map<String, Map<String, String>>
getNextStrictReplicaGroupAssignment(
Map<String, Map<String, String>> currentAssignment, Map<String,
Map<String, String>> targetAssignment,
- int minAvailableReplicas, boolean lowDiskMode) {
+ int minAvailableReplicas, boolean lowDiskMode, int batchSizePerServer,
+ Object2IntOpenHashMap<String> segmentPartitionIdMap, PartitionIdFetcher
partitionIdFetcher,
+ Logger tableRebalanceLogger) {
Map<String, Map<String, String>> nextAssignment = new TreeMap<>();
Map<String, Integer> numSegmentsToOffloadMap =
getNumSegmentsToOffloadMap(currentAssignment, targetAssignment);
+ Map<Integer, Map<Set<String>, Map<String, Map<String, String>>>>
+ partitionIdToAssignedInstancesToCurrentAssignmentMap;
+ if (batchSizePerServer == RebalanceConfig.DISABLE_BATCH_SIZE_PER_SERVER) {
+ // Don't calculate the partition id to assigned instances to current
assignment mapping if batching is disabled
+ // since we want to update the next assignment based on all partitions
in this case. Use partitionId as 0
+ // and a dummy set for the assigned instances.
+ partitionIdToAssignedInstancesToCurrentAssignmentMap = new TreeMap<>();
+ partitionIdToAssignedInstancesToCurrentAssignmentMap.put(0, new
HashMap<>());
+
partitionIdToAssignedInstancesToCurrentAssignmentMap.get(0).put(Set.of(""),
currentAssignment);
+ } else {
+ partitionIdToAssignedInstancesToCurrentAssignmentMap =
+
getPartitionIdToAssignedInstancesToCurrentAssignmentMap(currentAssignment,
segmentPartitionIdMap,
+ partitionIdFetcher);
+ }
Map<Pair<Set<String>, Set<String>>, Set<String>> assignmentMap = new
HashMap<>();
Map<Set<String>, Set<String>> availableInstancesMap = new HashMap<>();
- for (Map.Entry<String, Map<String, String>> entry :
currentAssignment.entrySet()) {
- String segmentName = entry.getKey();
- Map<String, String> currentInstanceStateMap = entry.getValue();
- Map<String, String> targetInstanceStateMap =
targetAssignment.get(segmentName);
- SingleSegmentAssignment assignment =
- getNextSingleSegmentAssignment(currentInstanceStateMap,
targetInstanceStateMap, minAvailableReplicas,
- lowDiskMode, numSegmentsToOffloadMap, assignmentMap);
- Set<String> assignedInstances = assignment._instanceStateMap.keySet();
- Set<String> availableInstances = assignment._availableInstances;
- availableInstancesMap.compute(assignedInstances, (k,
currentAvailableInstances) -> {
- if (currentAvailableInstances == null) {
- // First segment assigned to these instances, use the new assignment
and update the available instances
- nextAssignment.put(segmentName, assignment._instanceStateMap);
- updateNumSegmentsToOffloadMap(numSegmentsToOffloadMap,
currentInstanceStateMap.keySet(), k);
- return availableInstances;
- } else {
- // There are other segments assigned to the same instances, check
the available instances to see if adding the
- // new assignment can still hold the minimum available replicas
requirement
- availableInstances.retainAll(currentAvailableInstances);
- if (availableInstances.size() >= minAvailableReplicas) {
- // New assignment can be added
+
+ Map<String, Integer> serverToNumSegmentsAddedSoFar = new HashMap<>();
+ for (Map<Set<String>, Map<String, Map<String, String>>>
assignedInstancesToCurrentAssignment
+ : partitionIdToAssignedInstancesToCurrentAssignmentMap.values()) {
+ boolean anyServerExhaustedBatchSize = false;
+ if (batchSizePerServer != RebalanceConfig.DISABLE_BATCH_SIZE_PER_SERVER)
{
+ // Check if the servers of the first assignment for each unique set of
assigned instances has any space left
+ // to move this partition. If so, let's mark the partitions as to be
moved, otherwise we mark the partition
+ // as a whole as not moveable.
+ for (Map<String, Map<String, String>> curAssignment :
assignedInstancesToCurrentAssignment.values()) {
+ Map.Entry<String, Map<String, String>> firstEntry =
curAssignment.entrySet().iterator().next();
+ // It is enough to check for whether any server for one segment is
above the limit or not since all segments
+ // in curAssignment will have the same assigned instances list
+ Map<String, String> firstEntryInstanceStateMap =
firstEntry.getValue();
Review Comment:
Yeah we only check the first in each unique set of assigned instances (and
note that once we assign some servers for next assignment, we do cache that in
`assignmentMap`) to find the set of servers that'll be added.
we move the partition as a whole, looking at the full set of assigned
instances here -> we calculate whether we can move or not walking the
curAssignments once, and then call the assignment based on the flag.
> And we're choosing to move the group as a whole as long as all the new
servers added have enough batch size capacity left for a single segment
regardless of the actual number of segments in this group?
Yes that's correct
> I guess the intention is to make sure that we can progress even when the
batch size is smaller than the number of segments in one of these groups, but
won't it lead to excessively overshooting the batch size limit in some cases
that could otherwise have been split into multiple steps?
Yes we might move a much larger number of segments than the batch size will
allow, but we'll move at most 1 extra partitionId worth of segments. Imagine a
scenario where we have a small batch size of say 2, but each partition has 10
segments. Without this, we'd never be able to move anything. With this
mechanism of choosing, we'd only move that 1 partition, and for the 2nd
partition (if moving to one of the selected servers in the last partition),
we'll keep the assignment as current rather than moving
let me know if you have other ideas though
--
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.
To unsubscribe, e-mail: commits-unsubscr...@pinot.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@pinot.apache.org
For additional commands, e-mail: commits-h...@pinot.apache.org
