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

Example 41 with AffinityTopologyVersion

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

the class GridDhtPartitionsExchangeFuture method finishExchangeOnCoordinator.

/**
 * @param sndResNodes Additional nodes to send finish message to.
 */
private void finishExchangeOnCoordinator(@Nullable Collection<ClusterNode> sndResNodes) {
    if (isDone() || !enterBusy())
        return;
    try {
        if (!F.isEmpty(exchangeGlobalExceptions) && dynamicCacheStartExchange() && isRollbackSupported()) {
            sendExchangeFailureMessage();
            return;
        }
        AffinityTopologyVersion resTopVer = exchCtx.events().topologyVersion();
        if (log.isInfoEnabled()) {
            log.info("finishExchangeOnCoordinator [topVer=" + initialVersion() + ", resVer=" + resTopVer + ']');
        }
        Map<Integer, CacheGroupAffinityMessage> idealAffDiff = null;
        // Reserve at least 2 threads for system operations.
        int parallelismLvl = U.availableThreadCount(cctx.kernalContext(), GridIoPolicy.SYSTEM_POOL, 2);
        if (exchCtx.mergeExchanges()) {
            synchronized (mux) {
                if (mergedJoinExchMsgs != null) {
                    for (Map.Entry<UUID, GridDhtPartitionsSingleMessage> e : mergedJoinExchMsgs.entrySet()) {
                        msgs.put(e.getKey(), e.getValue());
                        updatePartitionSingleMap(e.getKey(), e.getValue());
                    }
                }
            }
            assert exchCtx.events().hasServerJoin() || exchCtx.events().hasServerLeft();
            exchCtx.events().processEvents(this);
            if (exchCtx.events().hasServerLeft())
                idealAffDiff = cctx.affinity().onServerLeftWithExchangeMergeProtocol(this);
            else
                cctx.affinity().onServerJoinWithExchangeMergeProtocol(this, true);
            doInParallel(parallelismLvl, cctx.kernalContext().pools().getSystemExecutorService(), cctx.affinity().cacheGroups().values(), desc -> {
                if (desc.config().getCacheMode() == CacheMode.LOCAL)
                    return null;
                CacheGroupContext grp = cctx.cache().cacheGroup(desc.groupId());
                GridDhtPartitionTopology top = grp != null ? grp.topology() : cctx.exchange().clientTopology(desc.groupId(), events().discoveryCache());
                top.beforeExchange(this, true, true);
                return null;
            });
        }
        span.addLog(() -> "Affinity recalculation (crd)");
        timeBag.finishGlobalStage("Affinity recalculation (crd)");
        Map<Integer, CacheGroupAffinityMessage> joinedNodeAff = new ConcurrentHashMap<>(cctx.cache().cacheGroups().size());
        doInParallel(parallelismLvl, cctx.kernalContext().pools().getSystemExecutorService(), msgs.values(), msg -> {
            processSingleMessageOnCrdFinish(msg, joinedNodeAff);
            return null;
        });
        timeBag.finishGlobalStage("Collect update counters and create affinity messages");
        if (firstDiscoEvt.type() == EVT_DISCOVERY_CUSTOM_EVT) {
            assert firstDiscoEvt instanceof DiscoveryCustomEvent;
            if (activateCluster() || changedBaseline())
                assignPartitionsStates(null);
            DiscoveryCustomMessage discoveryCustomMessage = ((DiscoveryCustomEvent) firstDiscoEvt).customMessage();
            if (discoveryCustomMessage instanceof DynamicCacheChangeBatch) {
                if (exchActions != null) {
                    Set<String> caches = exchActions.cachesToResetLostPartitions();
                    if (!F.isEmpty(caches))
                        resetLostPartitions(caches);
                    Set<Integer> cacheGroupsToResetOwners = concat(exchActions.cacheGroupsToStart().stream().map(grp -> grp.descriptor().groupId()), exchActions.cachesToResetLostPartitions().stream().map(CU::cacheId)).collect(Collectors.toSet());
                    assignPartitionsStates(cacheGroupsToResetOwners);
                }
            } else if (discoveryCustomMessage instanceof SnapshotDiscoveryMessage && ((SnapshotDiscoveryMessage) discoveryCustomMessage).needAssignPartitions()) {
                markAffinityReassign();
                assignPartitionsStates(null);
            }
        } else if (exchCtx.events().hasServerJoin())
            assignPartitionsStates(null);
        else if (exchCtx.events().hasServerLeft())
            assignPartitionsStates(emptySet());
        // Validation should happen after resetting owners to avoid false desync reporting.
        validatePartitionsState();
        // Recalculate new affinity based on partitions availability.
        if (!exchCtx.mergeExchanges() && forceAffReassignment) {
            idealAffDiff = cctx.affinity().onCustomEventWithEnforcedAffinityReassignment(this);
            timeBag.finishGlobalStage("Ideal affinity diff calculation (enforced)");
        }
        for (CacheGroupContext grpCtx : cctx.cache().cacheGroups()) {
            if (!grpCtx.isLocal())
                grpCtx.topology().applyUpdateCounters();
        }
        timeBag.finishGlobalStage("Apply update counters");
        updateLastVersion(cctx.versions().last());
        cctx.versions().onExchange(lastVer.get().order());
        IgniteProductVersion minVer = exchCtx.events().discoveryCache().minimumNodeVersion();
        GridDhtPartitionsFullMessage msg = createPartitionsMessage(true, minVer.compareToIgnoreTimestamp(PARTIAL_COUNTERS_MAP_SINCE) >= 0);
        if (!cctx.affinity().rebalanceRequired() && !deactivateCluster())
            msg.rebalanced(true);
        if (exchCtx.mergeExchanges()) {
            assert !centralizedAff;
            msg.resultTopologyVersion(resTopVer);
            if (exchCtx.events().hasServerLeft())
                msg.idealAffinityDiff(idealAffDiff);
        } else if (forceAffReassignment)
            msg.idealAffinityDiff(idealAffDiff);
        msg.prepareMarshal(cctx);
        timeBag.finishGlobalStage("Full message preparing");
        synchronized (mux) {
            finishState = new FinishState(crd.id(), resTopVer, msg);
            state = ExchangeLocalState.DONE;
        }
        if (centralizedAff) {
            assert !exchCtx.mergeExchanges();
            IgniteInternalFuture<Map<Integer, Map<Integer, List<UUID>>>> fut = cctx.affinity().initAffinityOnNodeLeft(this);
            if (!fut.isDone())
                fut.listen(this::onAffinityInitialized);
            else
                onAffinityInitialized(fut);
        } else {
            Set<ClusterNode> nodes;
            Map<UUID, GridDhtPartitionsSingleMessage> mergedJoinExchMsgs0;
            synchronized (mux) {
                srvNodes.remove(cctx.localNode());
                nodes = new LinkedHashSet<>(srvNodes);
                mergedJoinExchMsgs0 = mergedJoinExchMsgs;
                if (mergedJoinExchMsgs != null) {
                    for (Map.Entry<UUID, GridDhtPartitionsSingleMessage> e : mergedJoinExchMsgs.entrySet()) {
                        if (e.getValue() != null) {
                            ClusterNode node = cctx.discovery().node(e.getKey());
                            if (node != null)
                                nodes.add(node);
                        }
                    }
                } else
                    mergedJoinExchMsgs0 = Collections.emptyMap();
                if (!F.isEmpty(sndResNodes))
                    nodes.addAll(sndResNodes);
            }
            if (msg.rebalanced())
                markRebalanced();
            if (!nodes.isEmpty())
                sendAllPartitions(msg, nodes, mergedJoinExchMsgs0, joinedNodeAff);
            timeBag.finishGlobalStage("Full message sending");
            discoveryLag = calculateDiscoveryLag(msgs, mergedJoinExchMsgs0);
            if (!stateChangeExchange())
                onDone(exchCtx.events().topologyVersion(), null);
            for (Map.Entry<UUID, GridDhtPartitionsSingleMessage> e : pendingSingleMsgs.entrySet()) {
                if (log.isInfoEnabled()) {
                    log.info("Process pending message on coordinator [node=" + e.getKey() + ", ver=" + initialVersion() + ", resVer=" + resTopVer + ']');
                }
                processSingleMessage(e.getKey(), e.getValue());
            }
        }
        if (stateChangeExchange()) {
            StateChangeRequest req = exchActions.stateChangeRequest();
            assert req != null : exchActions;
            boolean stateChangeErr = false;
            if (!F.isEmpty(exchangeGlobalExceptions)) {
                stateChangeErr = true;
                cctx.kernalContext().state().onStateChangeError(exchangeGlobalExceptions, req);
            } else {
                boolean hasMoving = !partsToReload.isEmpty();
                Set<Integer> waitGrps = cctx.affinity().waitGroups();
                if (!hasMoving) {
                    for (CacheGroupContext grpCtx : cctx.cache().cacheGroups()) {
                        if (waitGrps.contains(grpCtx.groupId()) && grpCtx.topology().hasMovingPartitions()) {
                            hasMoving = true;
                            break;
                        }
                    }
                }
                cctx.kernalContext().state().onExchangeFinishedOnCoordinator(this, hasMoving);
            }
            if (!cctx.kernalContext().state().clusterState().localBaselineAutoAdjustment()) {
                ClusterState state = stateChangeErr ? ClusterState.INACTIVE : req.state();
                ChangeGlobalStateFinishMessage stateFinishMsg = new ChangeGlobalStateFinishMessage(req.requestId(), state, !stateChangeErr);
                cctx.discovery().sendCustomEvent(stateFinishMsg);
            }
            timeBag.finishGlobalStage("State finish message sending");
            if (!centralizedAff)
                onDone(exchCtx.events().topologyVersion(), null);
        }
        // Try switch late affinity right now if an exchange has been completed normally.
        if (!centralizedAff && isDone() && error() == null && !cctx.kernalContext().isStopping())
            cctx.exchange().checkRebalanceState();
    } catch (IgniteCheckedException e) {
        if (reconnectOnError(e))
            onDone(new IgniteNeedReconnectException(cctx.localNode(), e));
        else
            onDone(e);
    } finally {
        leaveBusy();
    }
}

Example 42 with AffinityTopologyVersion

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

the class GridDhtPartitionsExchangeFuture method onAllReceived.

/**
 * Called only for coordinator node when all {@link GridDhtPartitionsSingleMessage}s were received
 *
 * @param sndResNodes Additional nodes to send finish message to.
 */
private void onAllReceived(@Nullable Collection<ClusterNode> sndResNodes) {
    try {
        initFut.get();
        span.addLog(() -> "Waiting for all single messages");
        timeBag.finishGlobalStage("Waiting for all single messages");
        assert crd.isLocal();
        assert partHistSuppliers.isEmpty() : partHistSuppliers;
        if (!exchCtx.mergeExchanges() && !crd.equals(events().discoveryCache().serverNodes().get(0))) {
            for (CacheGroupContext grp : cctx.cache().cacheGroups()) {
                if (grp.isLocal())
                    continue;
                // For example, dynamic cache start failed.
                if (grp.affinity().lastVersion().topologyVersion() > 0)
                    grp.topology().beforeExchange(this, !centralizedAff && !forceAffReassignment, false);
                else
                    assert exchangeLocE != null : "Affinity is not calculated for the cache group [groupName=" + grp.name() + "]";
            }
        }
        if (exchCtx.mergeExchanges()) {
            if (log.isInfoEnabled())
                log.info("Coordinator received all messages, try merge [ver=" + initialVersion() + ']');
            AffinityTopologyVersion threshold = newCrdFut != null ? newCrdFut.resultTopologyVersion() : null;
            if (threshold != null) {
                assert newCrdFut.fullMessage() == null : "There is full message in new coordinator future, but exchange was not finished using it: " + newCrdFut.fullMessage();
            }
            boolean finish = cctx.exchange().mergeExchangesOnCoordinator(this, threshold);
            timeBag.finishGlobalStage("Exchanges merge");
            if (!finish)
                return;
        }
        finishExchangeOnCoordinator(sndResNodes);
    } catch (IgniteCheckedException e) {
        if (reconnectOnError(e))
            onDone(new IgniteNeedReconnectException(cctx.localNode(), e));
        else
            onDone(e);
    }
}

Example 43 with AffinityTopologyVersion

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

the class GridDhtPartitionTopologyImpl method beforeExchange.

/**
 * {@inheritDoc}
 */
@Override
public void beforeExchange(GridDhtPartitionsExchangeFuture exchFut, boolean affReady, boolean updateMoving) throws IgniteCheckedException {
    ctx.database().checkpointReadLock();
    try {
        U.writeLock(lock);
        try {
            if (stopping)
                return;
            assert lastTopChangeVer.equals(exchFut.initialVersion()) : "Invalid topology version [topVer=" + lastTopChangeVer + ", exchId=" + exchFut.exchangeId() + ']';
            ExchangeDiscoveryEvents evts = exchFut.context().events();
            if (affReady) {
                assert grp.affinity().lastVersion().equals(evts.topologyVersion()) : "Invalid affinity version [" + "grp=" + grp.cacheOrGroupName() + ", affVer=" + grp.affinity().lastVersion() + ", evtsVer=" + evts.topologyVersion() + ']';
                lastTopChangeVer = readyTopVer = evts.topologyVersion();
                discoCache = evts.discoveryCache();
            }
            if (log.isDebugEnabled()) {
                log.debug("Partition map beforeExchange [grp=" + grp.cacheOrGroupName() + ", exchId=" + exchFut.exchangeId() + ", fullMap=" + fullMapString() + ']');
            }
            long updateSeq = this.updateSeq.incrementAndGet();
            if (exchFut.exchangeType() == ALL && !exchFut.rebalanced())
                cntrMap.clear();
            initializeFullMap(updateSeq);
            boolean grpStarted = exchFut.cacheGroupAddedOnExchange(grp.groupId(), grp.receivedFrom());
            if (evts.hasServerLeft()) {
                for (DiscoveryEvent evt : evts.events()) {
                    if (ExchangeDiscoveryEvents.serverLeftEvent(evt))
                        removeNode(evt.eventNode().id());
                }
            } else if (affReady && grpStarted && exchFut.exchangeType() == NONE) {
                assert !exchFut.context().mergeExchanges() : exchFut;
                assert node2part != null && node2part.valid() : exchFut;
                // Initialize node maps if group was started from joining client.
                final List<ClusterNode> nodes = exchFut.firstEventCache().cacheGroupAffinityNodes(grp.groupId());
                for (ClusterNode node : nodes) {
                    if (!node2part.containsKey(node.id()) && ctx.discovery().alive(node)) {
                        final GridDhtPartitionMap partMap = new GridDhtPartitionMap(node.id(), 1L, exchFut.initialVersion(), new GridPartitionStateMap(), false);
                        final AffinityAssignment aff = grp.affinity().cachedAffinity(exchFut.initialVersion());
                        for (Integer p0 : aff.primaryPartitions(node.id())) partMap.put(p0, OWNING);
                        for (Integer p0 : aff.backupPartitions(node.id())) partMap.put(p0, OWNING);
                        node2part.put(node.id(), partMap);
                    }
                }
            }
            if (grp.affinityNode()) {
                if (grpStarted || exchFut.firstEvent().type() == EVT_DISCOVERY_CUSTOM_EVT || exchFut.serverNodeDiscoveryEvent()) {
                    AffinityTopologyVersion affVer;
                    List<List<ClusterNode>> affAssignment;
                    if (affReady) {
                        affVer = evts.topologyVersion();
                        assert grp.affinity().lastVersion().equals(affVer) : "Invalid affinity [topVer=" + grp.affinity().lastVersion() + ", grp=" + grp.cacheOrGroupName() + ", affVer=" + affVer + ", fut=" + exchFut + ']';
                        affAssignment = grp.affinity().readyAssignments(affVer);
                    } else {
                        assert !exchFut.context().mergeExchanges();
                        affVer = exchFut.initialVersion();
                        affAssignment = grp.affinity().idealAssignmentRaw();
                    }
                    initPartitions(affVer, affAssignment, exchFut, updateSeq);
                }
            }
            consistencyCheck();
            if (updateMoving) {
                assert grp.affinity().lastVersion().equals(evts.topologyVersion());
                createMovingPartitions(grp.affinity().readyAffinity(evts.topologyVersion()));
            }
            if (log.isDebugEnabled()) {
                log.debug("Partition map after beforeExchange [grp=" + grp.cacheOrGroupName() + ", " + "exchId=" + exchFut.exchangeId() + ", fullMap=" + fullMapString() + ']');
            }
            if (log.isTraceEnabled()) {
                log.trace("Partition states after beforeExchange [grp=" + grp.cacheOrGroupName() + ", exchId=" + exchFut.exchangeId() + ", states=" + dumpPartitionStates() + ']');
            }
        } finally {
            lock.writeLock().unlock();
        }
    } finally {
        ctx.database().checkpointReadUnlock();
    }
}

Example 44 with AffinityTopologyVersion

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

the class GridDhtPartitionTopologyImpl method afterExchange.

/**
 * {@inheritDoc}
 */
@Override
public boolean afterExchange(GridDhtPartitionsExchangeFuture exchFut) {
    boolean changed = false;
    int partitions = grp.affinity().partitions();
    AffinityTopologyVersion topVer = exchFut.context().events().topologyVersion();
    assert grp.affinity().lastVersion().equals(topVer) : "Affinity is not initialized " + "[grp=" + grp.cacheOrGroupName() + ", topVer=" + topVer + ", affVer=" + grp.affinity().lastVersion() + ", fut=" + exchFut + ']';
    ctx.database().checkpointReadLock();
    try {
        lock.writeLock().lock();
        try {
            if (stopping)
                return false;
            assert readyTopVer.initialized() : readyTopVer;
            assert lastTopChangeVer.equals(readyTopVer);
            if (log.isDebugEnabled()) {
                log.debug("Partition map before afterExchange [grp=" + grp.cacheOrGroupName() + ", exchId=" + exchFut.exchangeId() + ", fullMap=" + fullMapString() + ']');
            }
            if (log.isTraceEnabled()) {
                log.trace("Partition states before afterExchange [grp=" + grp.cacheOrGroupName() + ", exchVer=" + exchFut.exchangeId() + ", states=" + dumpPartitionStates() + ']');
            }
            long updateSeq = this.updateSeq.incrementAndGet();
            // Skip partition updates in case of not real exchange.
            if (!ctx.localNode().isClient() && exchFut.exchangeType() == ALL) {
                for (int p = 0; p < partitions; p++) {
                    GridDhtLocalPartition locPart = localPartition0(p, topVer, false, true);
                    if (partitionLocalNode(p, topVer)) {
                        // Prepare partition to rebalance if it's not happened on full map update phase.
                        if (locPart == null || locPart.state() == RENTING || locPart.state() == EVICTED)
                            locPart = rebalancePartition(p, true, exchFut);
                        GridDhtPartitionState state = locPart.state();
                        if (state == MOVING) {
                            if (grp.rebalanceEnabled()) {
                                Collection<ClusterNode> owners = owners(p);
                                // then new exchange should be started with detecting lost partitions.
                                if (!F.isEmpty(owners)) {
                                    if (log.isDebugEnabled())
                                        log.debug("Will not own partition (there are owners to rebalance from) " + "[grp=" + grp.cacheOrGroupName() + ", p=" + p + ", owners = " + owners + ']');
                                }
                            } else
                                updateSeq = updateLocal(p, locPart.state(), updateSeq, topVer);
                        }
                    } else {
                        if (locPart != null) {
                            GridDhtPartitionState state = locPart.state();
                            if (state == MOVING) {
                                locPart.rent();
                                updateSeq = updateLocal(p, locPart.state(), updateSeq, topVer);
                                changed = true;
                                if (log.isDebugEnabled()) {
                                    log.debug("Evicting MOVING partition (it does not belong to affinity) [" + "grp=" + grp.cacheOrGroupName() + ", p=" + locPart.id() + ']');
                                }
                            }
                        }
                    }
                }
            }
            AffinityAssignment aff = grp.affinity().readyAffinity(topVer);
            if (node2part != null && node2part.valid())
                changed |= checkEvictions(updateSeq, aff);
            updateRebalanceVersion(aff.topologyVersion(), aff.assignment());
            consistencyCheck();
            if (log.isTraceEnabled()) {
                log.trace("Partition states after afterExchange [grp=" + grp.cacheOrGroupName() + ", exchVer=" + exchFut.exchangeId() + ", states=" + dumpPartitionStates() + ']');
            }
        } finally {
            lock.writeLock().unlock();
        }
    } finally {
        ctx.database().checkpointReadUnlock();
    }
    return changed;
}

Example 45 with AffinityTopologyVersion

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

the class GridClientPartitionTopology method updateTopologyVersion.

/**
 * {@inheritDoc}
 */
@Override
public void updateTopologyVersion(GridDhtTopologyFuture exchFut, DiscoCache discoCache, long updSeq, boolean stopping) throws IgniteInterruptedCheckedException {
    U.writeLock(lock);
    try {
        AffinityTopologyVersion exchTopVer = exchFut.initialVersion();
        assert exchTopVer.compareTo(topVer) > 0 : "Invalid topology version [grp=" + grpId + ", topVer=" + topVer + ", exchVer=" + exchTopVer + ", discoCacheVer=" + (this.discoCache != null ? this.discoCache.version() : "None") + ", exchDiscoCacheVer=" + discoCache.version() + ']';
        this.stopping = stopping;
        topVer = exchTopVer;
        this.discoCache = discoCache;
        updateSeq.setIfGreater(updSeq);
        topReadyFut = exchFut;
    } finally {
        lock.writeLock().unlock();
    }
}