Examples with Distribution com.facebook.presto.raptor.metadata.Distribution used on opensource projects

Example 6 with Distribution

use of com.facebook.presto.raptor.metadata.Distribution in project presto by prestodb.

the class BucketBalancer method computeAssignmentChanges.

private static Multimap<String, BucketAssignment> computeAssignmentChanges(ClusterState clusterState) {
    Multimap<String, BucketAssignment> sourceToAllocationChanges = HashMultimap.create();
    Map<String, Long> allocationBytes = new HashMap<>(clusterState.getAssignedBytes());
    Set<String> activeNodes = clusterState.getActiveNodes();
    for (Distribution distribution : clusterState.getDistributionAssignments().keySet()) {
        // number of buckets in this distribution assigned to a node
        Multiset<String> allocationCounts = HashMultiset.create();
        Collection<BucketAssignment> distributionAssignments = clusterState.getDistributionAssignments().get(distribution);
        distributionAssignments.stream().map(BucketAssignment::getNodeIdentifier).forEach(allocationCounts::add);
        int currentMin = allocationBytes.keySet().stream().mapToInt(allocationCounts::count).min().getAsInt();
        int currentMax = allocationBytes.keySet().stream().mapToInt(allocationCounts::count).max().getAsInt();
        int numBuckets = distributionAssignments.size();
        int targetMin = (int) Math.floor((numBuckets * 1.0) / clusterState.getActiveNodes().size());
        int targetMax = (int) Math.ceil((numBuckets * 1.0) / clusterState.getActiveNodes().size());
        log.info("Distribution %s: Current bucket skew: min %s, max %s. Target bucket skew: min %s, max %s", distribution.getId(), currentMin, currentMax, targetMin, targetMax);
        for (String source : ImmutableSet.copyOf(allocationCounts)) {
            List<BucketAssignment> existingAssignments = distributionAssignments.stream().filter(assignment -> assignment.getNodeIdentifier().equals(source)).collect(toList());
            for (BucketAssignment existingAssignment : existingAssignments) {
                if (activeNodes.contains(source) && allocationCounts.count(source) <= targetMin) {
                    break;
                }
                // identify nodes with bucket counts lower than the computed target, and greedily select from this set based on projected disk utilization.
                // greediness means that this may produce decidedly non-optimal results if one looks at the global distribution of buckets->nodes.
                // also, this assumes that nodes in a cluster have identical storage capacity
                String target = activeNodes.stream().filter(candidate -> !candidate.equals(source) && allocationCounts.count(candidate) < targetMax).sorted(comparingInt(allocationCounts::count)).min(Comparator.comparingDouble(allocationBytes::get)).orElseThrow(() -> new VerifyException("unable to find target for rebalancing"));
                long bucketSize = clusterState.getDistributionBucketSize().get(distribution);
                // only move bucket if it reduces imbalance
                if (activeNodes.contains(source) && (allocationCounts.count(source) == targetMax && allocationCounts.count(target) == targetMin)) {
                    break;
                }
                allocationCounts.remove(source);
                allocationCounts.add(target);
                allocationBytes.compute(source, (k, v) -> v - bucketSize);
                allocationBytes.compute(target, (k, v) -> v + bucketSize);
                sourceToAllocationChanges.put(existingAssignment.getNodeIdentifier(), new BucketAssignment(existingAssignment.getDistributionId(), existingAssignment.getBucketNumber(), target));
            }
        }
    }
    return sourceToAllocationChanges;
}