Examples with Distribution io.trino.plugin.raptor.legacy.metadata.Distribution used on opensource projects

Example 1 with Distribution

use of io.trino.plugin.raptor.legacy.metadata.Distribution in project trino by trinodb.

the class RaptorMetadata method getOrCreateDistribution.

private Optional<DistributionInfo> getOrCreateDistribution(Map<String, RaptorColumnHandle> columnHandleMap, Map<String, Object> properties) {
    OptionalInt bucketCount = getBucketCount(properties);
    List<RaptorColumnHandle> bucketColumnHandles = getBucketColumnHandles(getBucketColumns(properties), columnHandleMap);
    if (bucketCount.isPresent() && bucketColumnHandles.isEmpty()) {
        throw new TrinoException(INVALID_TABLE_PROPERTY, format("Must specify '%s' along with '%s'", BUCKETED_ON_PROPERTY, BUCKET_COUNT_PROPERTY));
    }
    if (bucketCount.isEmpty() && !bucketColumnHandles.isEmpty()) {
        throw new TrinoException(INVALID_TABLE_PROPERTY, format("Must specify '%s' along with '%s'", BUCKET_COUNT_PROPERTY, BUCKETED_ON_PROPERTY));
    }
    ImmutableList.Builder<Type> bucketColumnTypes = ImmutableList.builder();
    for (RaptorColumnHandle column : bucketColumnHandles) {
        validateBucketType(column.getColumnType());
        bucketColumnTypes.add(column.getColumnType());
    }
    long distributionId;
    String distributionName = getDistributionName(properties);
    if (distributionName != null) {
        if (bucketColumnHandles.isEmpty()) {
            throw new TrinoException(INVALID_TABLE_PROPERTY, format("Must specify '%s' along with '%s'", BUCKETED_ON_PROPERTY, DISTRIBUTION_NAME_PROPERTY));
        }
        Distribution distribution = dao.getDistribution(distributionName);
        if (distribution == null) {
            if (bucketCount.isEmpty()) {
                throw new TrinoException(INVALID_TABLE_PROPERTY, "Distribution does not exist and bucket count is not specified");
            }
            distribution = getOrCreateDistribution(distributionName, bucketColumnTypes.build(), bucketCount.getAsInt());
        }
        distributionId = distribution.getId();
        if (bucketCount.isPresent() && (distribution.getBucketCount() != bucketCount.getAsInt())) {
            throw new TrinoException(INVALID_TABLE_PROPERTY, "Bucket count must match distribution");
        }
        if (!distribution.getColumnTypes().equals(bucketColumnTypes.build())) {
            throw new TrinoException(INVALID_TABLE_PROPERTY, "Bucket column types must match distribution");
        }
    } else if (bucketCount.isPresent()) {
        String types = Distribution.serializeColumnTypes(bucketColumnTypes.build());
        distributionId = dao.insertDistribution(null, types, bucketCount.getAsInt());
    } else {
        return Optional.empty();
    }
    shardManager.createBuckets(distributionId, bucketCount.getAsInt());
    return Optional.of(new DistributionInfo(distributionId, bucketCount.getAsInt(), bucketColumnHandles));
}

Example 2 with Distribution

use of io.trino.plugin.raptor.legacy.metadata.Distribution in project trino by trinodb.

the class BucketBalancer method fetchClusterState.

@VisibleForTesting
ClusterState fetchClusterState() {
    Set<String> activeNodes = nodeSupplier.getWorkerNodes().stream().map(Node::getNodeIdentifier).collect(toSet());
    Map<String, Long> assignedNodeSize = new HashMap<>(activeNodes.stream().collect(toMap(node -> node, node -> 0L)));
    ImmutableMultimap.Builder<Distribution, BucketAssignment> distributionAssignments = ImmutableMultimap.builder();
    ImmutableMap.Builder<Distribution, Long> distributionBucketSize = ImmutableMap.builder();
    for (Distribution distribution : shardManager.getDistributions()) {
        long distributionSize = shardManager.getDistributionSizeInBytes(distribution.getId());
        long bucketSize = (long) (1.0 * distributionSize) / distribution.getBucketCount();
        distributionBucketSize.put(distribution, bucketSize);
        for (BucketNode bucketNode : shardManager.getBucketNodes(distribution.getId())) {
            String node = bucketNode.getNodeIdentifier();
            distributionAssignments.put(distribution, new BucketAssignment(distribution.getId(), bucketNode.getBucketNumber(), node));
            assignedNodeSize.merge(node, bucketSize, Math::addExact);
        }
    }
    return new ClusterState(activeNodes, assignedNodeSize, distributionAssignments.build(), distributionBucketSize.buildOrThrow());
}

Example 3 with Distribution

use of io.trino.plugin.raptor.legacy.metadata.Distribution in project trino by trinodb.

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;
}

Example 4 with Distribution

use of io.trino.plugin.raptor.legacy.metadata.Distribution in project trino by trinodb.

the class TestBucketBalancer method assertBalancing.

private static void assertBalancing(BucketBalancer balancer, int expectedMoves) {
    int actualMoves = balancer.balance();
    assertEquals(actualMoves, expectedMoves);
    // check that number of buckets per node is within bounds
    ClusterState clusterState = balancer.fetchClusterState();
    for (Distribution distribution : clusterState.getDistributionAssignments().keySet()) {
        Multiset<String> allocationCounts = HashMultiset.create();
        clusterState.getDistributionAssignments().get(distribution).stream().map(BucketAssignment::getNodeIdentifier).forEach(allocationCounts::add);
        double bucketsPerNode = (1.0 * allocationCounts.size()) / clusterState.getActiveNodes().size();
        for (String node : allocationCounts) {
            assertGreaterThanOrEqual(allocationCounts.count(node), (int) Math.floor(bucketsPerNode), node + " has fewer buckets than expected");
            assertLessThanOrEqual(allocationCounts.count(node), (int) Math.ceil(bucketsPerNode), node + " has more buckets than expected");
        }
    }
    // check stability
    assertEquals(balancer.balance(), 0);
}

Example 5 with Distribution

use of io.trino.plugin.raptor.legacy.metadata.Distribution in project trino by trinodb.

the class RaptorMetadata method getOrCreateDistribution.

private Distribution getOrCreateDistribution(String name, List<Type> columnTypes, int bucketCount) {
    String types = Distribution.serializeColumnTypes(columnTypes);
    runIgnoringConstraintViolation(() -> dao.insertDistribution(name, types, bucketCount));
    Distribution distribution = dao.getDistribution(name);
    if (distribution == null) {
        throw new TrinoException(RAPTOR_ERROR, "Distribution does not exist after insert");
    }
    return distribution;
}