Examples with AffinityTopologyVersion org.apache.ignite.internal.processors.affinity.AffinityTopologyVersion used on opensource projects

Example 31 with AffinityTopologyVersion

use of org.apache.ignite.internal.processors.affinity.AffinityTopologyVersion in project ignite by apache.

the class GridDhtPartitionDemander method addAssignments.

/**
 * Initiates new rebalance process from given {@code assignments}.
 * If previous rebalance is not finished method cancels it.
 * In case of delayed rebalance method schedules new with configured delay.
 *
 * @param assignments Assignments.
 * @param force {@code True} if dummy reassign.
 * @param rebalanceId Rebalance id.
 * @param next Runnable responsible for cache rebalancing start.
 * @param forcedRebFut External future for forced rebalance.
 * @return Rebalancing runnable.
 */
Runnable addAssignments(final GridDhtPreloaderAssignments assignments, boolean force, long rebalanceId, final Runnable next, @Nullable final GridCompoundFuture<Boolean, Boolean> forcedRebFut) {
    if (log.isDebugEnabled())
        log.debug("Adding partition assignments: " + assignments);
    assert force == (forcedRebFut != null);
    long delay = grp.config().getRebalanceDelay();
    if ((delay == 0 || force) && assignments != null) {
        final RebalanceFuture oldFut = rebalanceFut;
        final RebalanceFuture fut = new RebalanceFuture(grp, assignments, log, rebalanceId);
        if (!oldFut.isInitial())
            oldFut.cancel();
        else
            fut.listen(f -> oldFut.onDone(f.result()));
        if (forcedRebFut != null)
            forcedRebFut.add(fut);
        rebalanceFut = fut;
        for (final GridCacheContext cctx : grp.caches()) {
            if (cctx.statisticsEnabled()) {
                final CacheMetricsImpl metrics = cctx.cache().metrics0();
                metrics.clearRebalanceCounters();
                metrics.startRebalance(0);
                rebalanceFut.listen(f -> metrics.clearRebalanceCounters());
            }
        }
        fut.sendRebalanceStartedEvent();
        if (assignments.cancelled()) {
            // Pending exchange.
            if (log.isDebugEnabled())
                log.debug("Rebalancing skipped due to cancelled assignments.");
            fut.onDone(false);
            fut.sendRebalanceFinishedEvent();
            return null;
        }
        if (assignments.isEmpty()) {
            // Nothing to rebalance.
            if (log.isDebugEnabled())
                log.debug("Rebalancing skipped due to empty assignments.");
            fut.onDone(true);
            ((GridFutureAdapter) grp.preloader().syncFuture()).onDone();
            fut.sendRebalanceFinishedEvent();
            return null;
        }
        return () -> {
            if (next != null)
                fut.listen(f -> {
                    try {
                        if (// Not cancelled.
                        f.get())
                            // Starts next cache rebalancing (according to the order).
                            next.run();
                    } catch (IgniteCheckedException e) {
                        if (log.isDebugEnabled())
                            log.debug(e.getMessage());
                    }
                });
            requestPartitions(fut, assignments);
        };
    } else if (delay > 0) {
        for (GridCacheContext cctx : grp.caches()) {
            if (cctx.statisticsEnabled()) {
                final CacheMetricsImpl metrics = cctx.cache().metrics0();
                metrics.startRebalance(delay);
            }
        }
        GridTimeoutObject obj = lastTimeoutObj.get();
        if (obj != null)
            ctx.time().removeTimeoutObject(obj);
        final GridDhtPartitionsExchangeFuture exchFut = lastExchangeFut;
        assert exchFut != null : "Delaying rebalance process without topology event.";
        obj = new GridTimeoutObjectAdapter(delay) {

            @Override
            public void onTimeout() {
                exchFut.listen(new CI1<IgniteInternalFuture<AffinityTopologyVersion>>() {

                    @Override
                    public void apply(IgniteInternalFuture<AffinityTopologyVersion> f) {
                        ctx.exchange().forceRebalance(exchFut.exchangeId());
                    }
                });
            }
        };
        lastTimeoutObj.set(obj);
        ctx.time().addTimeoutObject(obj);
    }
    return null;
}

Example 32 with AffinityTopologyVersion

use of org.apache.ignite.internal.processors.affinity.AffinityTopologyVersion in project ignite by apache.

the class GridDhtPartitionDemander method handleSupplyMessage.

/**
 * Handles supply message from {@code nodeId} with specified {@code topicId}.
 *
 * Supply message contains entries to populate rebalancing partitions.
 *
 * There is a cyclic process:
 * Populate rebalancing partitions with entries from Supply message.
 * If not all partitions specified in {@link #rebalanceFut} were rebalanced or marked as missed
 * send new Demand message to request next batch of entries.
 *
 * @param topicId Topic id.
 * @param nodeId Node id.
 * @param supply Supply message.
 */
public void handleSupplyMessage(int topicId, final UUID nodeId, final GridDhtPartitionSupplyMessage supply) {
    AffinityTopologyVersion topVer = supply.topologyVersion();
    final RebalanceFuture fut = rebalanceFut;
    ClusterNode node = ctx.node(nodeId);
    if (node == null)
        return;
    if (// Topology already changed (for the future that supply message based on).
    topologyChanged(fut))
        return;
    if (!fut.isActual(supply.rebalanceId())) {
        // Supple message based on another future.
        return;
    }
    if (log.isDebugEnabled())
        log.debug("Received supply message [grp=" + grp.cacheOrGroupName() + ", msg=" + supply + ']');
    // Check whether there were class loading errors on unmarshal
    if (supply.classError() != null) {
        U.warn(log, "Rebalancing from node cancelled [grp=" + grp.cacheOrGroupName() + ", node=" + nodeId + "]. Class got undeployed during preloading: " + supply.classError());
        fut.cancel(nodeId);
        return;
    }
    final GridDhtPartitionTopology top = grp.topology();
    if (grp.sharedGroup()) {
        for (GridCacheContext cctx : grp.caches()) {
            if (cctx.statisticsEnabled()) {
                long keysCnt = supply.keysForCache(cctx.cacheId());
                if (keysCnt != -1)
                    cctx.cache().metrics0().onRebalancingKeysCountEstimateReceived(keysCnt);
                // Can not be calculated per cache.
                cctx.cache().metrics0().onRebalanceBatchReceived(supply.messageSize());
            }
        }
    } else {
        GridCacheContext cctx = grp.singleCacheContext();
        if (cctx.statisticsEnabled()) {
            if (supply.estimatedKeysCount() != -1)
                cctx.cache().metrics0().onRebalancingKeysCountEstimateReceived(supply.estimatedKeysCount());
            cctx.cache().metrics0().onRebalanceBatchReceived(supply.messageSize());
        }
    }
    try {
        AffinityAssignment aff = grp.affinity().cachedAffinity(topVer);
        GridCacheContext cctx = grp.sharedGroup() ? null : grp.singleCacheContext();
        // Preload.
        for (Map.Entry<Integer, CacheEntryInfoCollection> e : supply.infos().entrySet()) {
            int p = e.getKey();
            if (aff.get(p).contains(ctx.localNode())) {
                GridDhtLocalPartition part = top.localPartition(p, topVer, true);
                assert part != null;
                boolean last = supply.last().containsKey(p);
                if (part.state() == MOVING) {
                    boolean reserved = part.reserve();
                    assert reserved : "Failed to reserve partition [igniteInstanceName=" + ctx.igniteInstanceName() + ", grp=" + grp.cacheOrGroupName() + ", part=" + part + ']';
                    part.lock();
                    try {
                        // Loop through all received entries and try to preload them.
                        for (GridCacheEntryInfo entry : e.getValue().infos()) {
                            if (!preloadEntry(node, p, entry, topVer)) {
                                if (log.isDebugEnabled())
                                    log.debug("Got entries for invalid partition during " + "preloading (will skip) [p=" + p + ", entry=" + entry + ']');
                                break;
                            }
                            if (grp.sharedGroup() && (cctx == null || cctx.cacheId() != entry.cacheId()))
                                cctx = ctx.cacheContext(entry.cacheId());
                            if (cctx != null && cctx.statisticsEnabled())
                                cctx.cache().metrics0().onRebalanceKeyReceived();
                        }
                        // then we take ownership.
                        if (last) {
                            top.own(part);
                            fut.partitionDone(nodeId, p);
                            if (log.isDebugEnabled())
                                log.debug("Finished rebalancing partition: " + part);
                        }
                    } finally {
                        part.unlock();
                        part.release();
                    }
                } else {
                    if (last)
                        fut.partitionDone(nodeId, p);
                    if (log.isDebugEnabled())
                        log.debug("Skipping rebalancing partition (state is not MOVING): " + part);
                }
            } else {
                fut.partitionDone(nodeId, p);
                if (log.isDebugEnabled())
                    log.debug("Skipping rebalancing partition (it does not belong on current node): " + p);
            }
        }
        // Only request partitions based on latest topology version.
        for (Integer miss : supply.missed()) {
            if (aff.get(miss).contains(ctx.localNode()))
                fut.partitionMissed(nodeId, miss);
        }
        for (Integer miss : supply.missed()) fut.partitionDone(nodeId, miss);
        GridDhtPartitionDemandMessage d = new GridDhtPartitionDemandMessage(supply.rebalanceId(), supply.topologyVersion(), grp.groupId());
        d.timeout(grp.config().getRebalanceTimeout());
        d.topic(rebalanceTopics.get(topicId));
        if (!topologyChanged(fut) && !fut.isDone()) {
            // Send demand message.
            try {
                ctx.io().sendOrderedMessage(node, rebalanceTopics.get(topicId), d.convertIfNeeded(node.version()), grp.ioPolicy(), grp.config().getRebalanceTimeout());
            } catch (ClusterTopologyCheckedException e) {
                if (log.isDebugEnabled()) {
                    log.debug("Node left during rebalancing [grp=" + grp.cacheOrGroupName() + ", node=" + node.id() + ", msg=" + e.getMessage() + ']');
                }
            }
        }
    } catch (IgniteSpiException | IgniteCheckedException e) {
        LT.error(log, e, "Error during rebalancing [grp=" + grp.cacheOrGroupName() + ", srcNode=" + node.id() + ", err=" + e + ']');
    }
}

Example 33 with AffinityTopologyVersion

use of org.apache.ignite.internal.processors.affinity.AffinityTopologyVersion in project ignite by apache.

the class IgniteClusterActivateDeactivateTest method startNodesAndBlockStatusChange.

/**
 * @param srvs Number of servers.
 * @param clients Number of clients.
 * @param stateChangeFrom Index of node initiating changes.
 * @param initiallyActive If {@code true} start cluster in active state (otherwise in inactive).
 * @param blockMsgNodes Nodes whcis block exchange messages.
 * @return State change future.
 * @throws Exception If failed.
 */
private IgniteInternalFuture<?> startNodesAndBlockStatusChange(int srvs, int clients, final int stateChangeFrom, final boolean initiallyActive, int... blockMsgNodes) throws Exception {
    active = initiallyActive;
    testSpi = true;
    startWithCaches1(srvs, clients);
    int minorVer = 1;
    if (initiallyActive && persistenceEnabled()) {
        ignite(0).cluster().active(true);
        minorVer++;
    }
    if (blockMsgNodes.length == 0)
        blockMsgNodes = new int[] { 1 };
    final AffinityTopologyVersion STATE_CHANGE_TOP_VER = new AffinityTopologyVersion(srvs + clients, minorVer);
    List<TestRecordingCommunicationSpi> spis = new ArrayList<>();
    for (int idx : blockMsgNodes) {
        TestRecordingCommunicationSpi spi = TestRecordingCommunicationSpi.spi(ignite(idx));
        spis.add(spi);
        blockExchangeSingleMessage(spi, STATE_CHANGE_TOP_VER);
    }
    IgniteInternalFuture<?> stateChangeFut = GridTestUtils.runAsync(() -> ignite(stateChangeFrom).cluster().active(!initiallyActive));
    for (TestRecordingCommunicationSpi spi : spis) spi.waitForBlocked();
    U.sleep(500);
    assertFalse(stateChangeFut.isDone());
    return stateChangeFut;
}

Example 34 with AffinityTopologyVersion

use of org.apache.ignite.internal.processors.affinity.AffinityTopologyVersion in project ignite by apache.

the class GridCommonAbstractTest method awaitPartitionMapExchange.

/**
 * @param waitEvicts If {@code true} will wait for evictions finished.
 * @param waitNode2PartUpdate If {@code true} will wait for nodes node2part info update finished.
 * @param nodes Optional nodes. If {@code null} method will wait for all nodes, for non null collection nodes will
 *      be filtered
 * @param printPartState If {@code true} will print partition state if evictions not happened.
 * @throws InterruptedException If interrupted.
 */
@SuppressWarnings("BusyWait")
protected void awaitPartitionMapExchange(boolean waitEvicts, boolean waitNode2PartUpdate, @Nullable Collection<ClusterNode> nodes, boolean printPartState) throws InterruptedException {
    long timeout = getPartitionMapExchangeTimeout();
    long startTime = -1;
    Set<String> names = new HashSet<>();
    Ignite crd = null;
    for (Ignite g : G.allGrids()) {
        ClusterNode node = g.cluster().localNode();
        if (crd == null || node.order() < crd.cluster().localNode().order()) {
            crd = g;
            if (node.order() == 1)
                break;
        }
    }
    if (crd == null)
        return;
    AffinityTopologyVersion waitTopVer = ((IgniteKernal) crd).context().discovery().topologyVersionEx();
    if (waitTopVer.topologyVersion() <= 0)
        waitTopVer = new AffinityTopologyVersion(1, 0);
    for (Ignite g : G.allGrids()) {
        if (nodes != null && !nodes.contains(g.cluster().localNode()))
            continue;
        IgniteKernal g0 = (IgniteKernal) g;
        names.add(g0.configuration().getIgniteInstanceName());
        if (startTime != -1) {
            if (startTime != g0.context().discovery().gridStartTime())
                fail("Found nodes from different clusters, probable some test does not stop nodes " + "[allNodes=" + names + ']');
        } else
            startTime = g0.context().discovery().gridStartTime();
        if (g.cluster().localNode().isDaemon())
            continue;
        IgniteInternalFuture<?> exchFut = g0.context().cache().context().exchange().affinityReadyFuture(waitTopVer);
        if (exchFut != null && !exchFut.isDone()) {
            try {
                exchFut.get(timeout);
            } catch (IgniteCheckedException e) {
                log.error("Failed to wait for exchange [topVer=" + waitTopVer + ", node=" + g0.name() + ']', e);
            }
        }
        for (IgniteCacheProxy<?, ?> c : g0.context().cache().jcaches()) {
            CacheConfiguration cfg = c.context().config();
            if (cfg == null)
                continue;
            if (cfg.getCacheMode() != LOCAL && cfg.getRebalanceMode() != NONE && g.cluster().nodes().size() > 1) {
                AffinityFunction aff = cfg.getAffinity();
                GridDhtCacheAdapter<?, ?> dht = dht(c);
                GridDhtPartitionTopology top = dht.topology();
                for (int p = 0; p < aff.partitions(); p++) {
                    long start = 0;
                    for (int i = 0; ; i++) {
                        boolean match = false;
                        GridCachePartitionExchangeManager<?, ?> exchMgr = dht.context().shared().exchange();
                        AffinityTopologyVersion readyVer = exchMgr.readyAffinityVersion();
                        // Otherwise, there may be an assertion when printing top.readyTopologyVersion().
                        try {
                            IgniteInternalFuture<?> fut = exchMgr.affinityReadyFuture(readyVer);
                            if (fut != null)
                                fut.get();
                        } catch (IgniteCheckedException e) {
                            throw new IgniteException(e);
                        }
                        if (readyVer.topologyVersion() > 0 && c.context().started()) {
                            // Must map on updated version of topology.
                            Collection<ClusterNode> affNodes = dht.context().affinity().assignment(readyVer).idealAssignment().get(p);
                            int affNodesCnt = affNodes.size();
                            GridDhtTopologyFuture topFut = top.topologyVersionFuture();
                            Collection<ClusterNode> owners = (topFut != null && topFut.isDone()) ? top.owners(p, AffinityTopologyVersion.NONE) : Collections.<ClusterNode>emptyList();
                            int ownerNodesCnt = owners.size();
                            GridDhtLocalPartition loc = top.localPartition(p, readyVer, false);
                            if (affNodesCnt != ownerNodesCnt || !affNodes.containsAll(owners) || (waitEvicts && loc != null && loc.state() != GridDhtPartitionState.OWNING)) {
                                LT.warn(log(), "Waiting for topology map update [" + "igniteInstanceName=" + g.name() + ", cache=" + cfg.getName() + ", cacheId=" + dht.context().cacheId() + ", topVer=" + top.readyTopologyVersion() + ", p=" + p + ", affNodesCnt=" + affNodesCnt + ", ownersCnt=" + ownerNodesCnt + ", affNodes=" + F.nodeIds(affNodes) + ", owners=" + F.nodeIds(owners) + ", topFut=" + topFut + ", locNode=" + g.cluster().localNode() + ']');
                            } else
                                match = true;
                        } else {
                            LT.warn(log(), "Waiting for topology map update [" + "igniteInstanceName=" + g.name() + ", cache=" + cfg.getName() + ", cacheId=" + dht.context().cacheId() + ", topVer=" + top.readyTopologyVersion() + ", started=" + dht.context().started() + ", p=" + p + ", readVer=" + readyVer + ", locNode=" + g.cluster().localNode() + ']');
                        }
                        if (!match) {
                            if (i == 0)
                                start = System.currentTimeMillis();
                            if (System.currentTimeMillis() - start > timeout) {
                                U.dumpThreads(log);
                                if (printPartState)
                                    printPartitionState(c);
                                throw new IgniteException("Timeout of waiting for topology map update [" + "igniteInstanceName=" + g.name() + ", cache=" + cfg.getName() + ", cacheId=" + dht.context().cacheId() + ", topVer=" + top.readyTopologyVersion() + ", p=" + p + ", readVer=" + readyVer + ", locNode=" + g.cluster().localNode() + ']');
                            }
                            // Busy wait.
                            Thread.sleep(20);
                            continue;
                        }
                        if (i > 0)
                            log().warning("Finished waiting for topology map update [igniteInstanceName=" + g.name() + ", p=" + p + ", duration=" + (System.currentTimeMillis() - start) + "ms]");
                        break;
                    }
                }
                if (waitNode2PartUpdate) {
                    long start = System.currentTimeMillis();
                    boolean failed = true;
                    while (failed) {
                        failed = false;
                        for (GridDhtPartitionMap pMap : top.partitionMap(true).values()) {
                            if (failed)
                                break;
                            for (Map.Entry entry : pMap.entrySet()) {
                                if (System.currentTimeMillis() - start > timeout) {
                                    U.dumpThreads(log);
                                    throw new IgniteException("Timeout of waiting for partition state update [" + "igniteInstanceName=" + g.name() + ", cache=" + cfg.getName() + ", cacheId=" + dht.context().cacheId() + ", topVer=" + top.readyTopologyVersion() + ", locNode=" + g.cluster().localNode() + ']');
                                }
                                if (entry.getValue() != GridDhtPartitionState.OWNING) {
                                    LT.warn(log(), "Waiting for correct partition state part=" + entry.getKey() + ", should be OWNING [state=" + entry.getValue() + "], node=" + g.name() + ", cache=" + c.getName());
                                    // Busy wait.
                                    Thread.sleep(200);
                                    failed = true;
                                    break;
                                }
                            }
                        }
                    }
                }
            }
        }
    }
    log.info("awaitPartitionMapExchange finished");
}

Example 35 with AffinityTopologyVersion

use of org.apache.ignite.internal.processors.affinity.AffinityTopologyVersion in project ignite by apache.

the class GridAffinityAssignmentJolBenchmark method measure.

/**
 * @param aff Aff.
 * @param parts Parts.
 * @param nodeCnt Node count.
 * @param backups Backups.
 */
private static void measure(RendezvousAffinityFunction aff, int parts, int nodeCnt, int backups) throws Exception {
    List<ClusterNode> nodes = new ArrayList<>();
    for (int i = 0; i < nodeCnt; i++) {
        ClusterNode node = node(i);
        nodes.add(node);
    }
    AffinityFunctionContext ctx = new GridAffinityFunctionContextImpl(nodes, new ArrayList<>(), new DiscoveryEvent(), new AffinityTopologyVersion(), backups);
    List<List<ClusterNode>> assignment = aff.assignPartitions(ctx);
    setOptimization(false);
    GridAffinityAssignmentV2 ga = new GridAffinityAssignmentV2(new AffinityTopologyVersion(1, 0), assignment, new ArrayList<>());
    System.gc();
    long totalSize = GraphLayout.parseInstance(ga).totalSize();
    System.out.println("Optimized, parts " + parts + " nodeCount " + nodeCnt + " backups " + backups + " " + totalSize);
    setOptimization(true);
    GridAffinityAssignmentV2 ga2 = new GridAffinityAssignmentV2(new AffinityTopologyVersion(1, 0), assignment, new ArrayList<>());
    System.gc();
    long totalSize2 = GraphLayout.parseInstance(ga2).totalSize();
    System.out.println("Deoptimized, parts " + parts + " nodeCount " + nodeCnt + " backups " + backups + " " + totalSize2);
    if (totalSize > totalSize2)
        throw new Exception("Optimized AffinityAssignment size " + totalSize + " is more than deoptimized " + totalSize2);
}