sodonnel commented on code in PR #3963: URL: https://github.com/apache/ozone/pull/3963#discussion_r1028269342
########## hadoop-hdds/server-scm/src/main/java/org/apache/hadoop/hdds/scm/container/move/MoveManagerImpl.java: ########## @@ -0,0 +1,750 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.hadoop.hdds.scm.container.move; + +import org.apache.hadoop.hdds.protocol.DatanodeDetails; +import org.apache.hadoop.hdds.protocol.proto.HddsProtos; +import org.apache.hadoop.hdds.protocol.proto.StorageContainerDatanodeProtocolProtos; +import org.apache.hadoop.hdds.scm.ContainerPlacementStatus; +import org.apache.hadoop.hdds.scm.PlacementPolicyValidateProxy; +import org.apache.hadoop.hdds.scm.container.ContainerID; +import org.apache.hadoop.hdds.scm.container.ContainerInfo; +import org.apache.hadoop.hdds.scm.container.ContainerManager; +import org.apache.hadoop.hdds.scm.container.ContainerNotFoundException; +import org.apache.hadoop.hdds.scm.container.ContainerReplica; +import org.apache.hadoop.hdds.scm.container.ContainerReplicaNotFoundException; +import org.apache.hadoop.hdds.scm.container.common.helpers.MoveDataNodePair; +import org.apache.hadoop.hdds.scm.container.replication.ContainerReplicaCount; +import org.apache.hadoop.hdds.scm.container.replication.ContainerReplicaOp; +import org.apache.hadoop.hdds.scm.container.replication.ContainerReplicaOp.PendingOpType; +import org.apache.hadoop.hdds.scm.events.SCMEvents; +import org.apache.hadoop.hdds.scm.ha.SCMContext; +import org.apache.hadoop.hdds.scm.ha.SCMHAInvocationHandler; +import org.apache.hadoop.hdds.scm.ha.SCMRatisServer; +import org.apache.hadoop.hdds.scm.metadata.DBTransactionBuffer; +import org.apache.hadoop.hdds.scm.metadata.Replicate; +import org.apache.hadoop.hdds.scm.node.NodeManager; +import org.apache.hadoop.hdds.scm.node.NodeStatus; +import org.apache.hadoop.hdds.scm.node.states.NodeNotFoundException; +import org.apache.hadoop.hdds.scm.server.StorageContainerManager; +import org.apache.hadoop.hdds.server.events.EventPublisher; +import org.apache.hadoop.hdds.utils.db.Table; +import org.apache.hadoop.hdds.utils.db.TableIterator; +import org.apache.hadoop.ozone.protocol.commands.CommandForDatanode; +import org.apache.hadoop.ozone.protocol.commands.DeleteContainerCommand; +import org.apache.hadoop.ozone.protocol.commands.ReplicateContainerCommand; +import org.apache.ratis.protocol.exceptions.NotLeaderException; +import org.apache.ratis.util.ExitUtils; +import org.apache.ratis.util.Preconditions; +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import java.io.IOException; +import java.lang.reflect.Proxy; +import java.util.Collections; +import java.util.List; +import java.util.Map; +import java.util.Set; +import java.util.concurrent.CompletableFuture; +import java.util.concurrent.ConcurrentHashMap; +import java.util.concurrent.TimeoutException; +import java.util.stream.Collectors; + +import static org.apache.hadoop.hdds.protocol.proto.SCMRatisProtocol.RequestType.MOVE; + +/** + * An HA enabled implementation of MoveManager. + */ +public final class MoveManagerImpl implements MoveManager { + public static final Logger LOG = + LoggerFactory.getLogger(MoveManagerImpl.class); + + private final SCMContext scmContext; + private final NodeManager nodeManager; + private final ContainerManager containerManager; + private final PlacementPolicyValidateProxy placementPolicyValidateProxy; + private final EventPublisher eventPublisher; + private final StorageContainerManager scm; + private final Map<ContainerID, CompletableFuture<MoveResult>> + pendingMoveFuture = new ConcurrentHashMap<>(); + private final Map<ContainerID, MoveDataNodePair> + pendingMoveOps = new ConcurrentHashMap<>(); + private final MoveDBHandler moveDBHandler; + private volatile boolean running = false; + + public MoveManagerImpl(final StorageContainerManager scm) throws IOException { + this.nodeManager = scm.getScmNodeManager(); + this.containerManager = scm.getContainerManager(); + this.scmContext = scm.getScmContext(); + this.placementPolicyValidateProxy = scm.getPlacementPolicyValidateProxy(); + this.eventPublisher = scm.getEventQueue(); + this.scm = scm; + this.moveDBHandler = new MoveDBHandlerImpl.Builder() + .setMoveTable(scm.getScmMetadataStore().getMoveTable()) + .setRatisServer(scm.getScmHAManager().getRatisServer()) + .setDBTransactionBuffer(scm.getScmHAManager().getDBTransactionBuffer()) + .build(); + } + + /** + * get all the pending move operations . + */ + public Map<ContainerID, MoveDataNodePair> getPendingMove() { + return pendingMoveOps; + } + + /** + * completeMove a move action for a given container. + * + * @param cid Container id to which the move option is finished + */ + private void completeMove(final ContainerID cid, final MoveResult mr) { + if (pendingMoveOps.containsKey(cid)) { + boolean successful = false; + try { + successful = moveDBHandler.deleteFromDB(cid); + } catch (TimeoutException t) { + //TimeoutException here is caused by ratis request timeout. + //we do not throw TimeoutException out here since completeMove is called + //in many places. Handling this exception inside itself seems + //a better way to have lower complexity. + LOG.warn("fail to delete move from DB"); + } + + // no matter deletion is successful or not, we could complete the + // corresponding pending CompletableFuture, because : + // 1 if completeMove is called , it means move is actually completed. + // 2 a CompletableFuture here means a `move` method call from + // container balancer, and balancer will block at CompletableFuture.get(), + // so we would better let it return as soon as possible. + CompletableFuture<MoveResult> future = pendingMoveFuture.get(cid); + pendingMoveFuture.remove(cid); + if (successful) { + pendingMoveOps.remove(cid); + if (future != null && mr != null) { + // when we know the future is null, and we want to complete + // the move , then we set mr to null. + // for example , after becoming a new leader, we might need + // to complete some moves and we know these futures does not + // exist. + future.complete(mr); + } + } else { + //TODO: if unsuccessful(actually rarely), cid will not be removed + // from pendingMoveOps. we need to try to delete cid from db and + // remove it from pendingMoveOps again. we could start a background + // service to handle those failed cases. + // Note that, when scm leader switches, new leader will also try + // to delete those stale completed move from db. + if (future != null) { + future.complete(MoveResult.FAIL_CAN_NOT_RECORD_TO_DB); + } + } + } + } + + /** + * start a move for a given container. + * + * @param cid Container id + * @param src source datanode + * @param tgt target datanode + */ + private void startMove( + final ContainerID cid, final DatanodeDetails src, + final DatanodeDetails tgt, final CompletableFuture<MoveResult> ret) + throws TimeoutException, ContainerReplicaNotFoundException, + ContainerNotFoundException { + if (!pendingMoveOps.containsKey(cid)) { + MoveDataNodePair mp = new MoveDataNodePair(src, tgt); + if (moveDBHandler.addToDB(cid, mp)) { + sendReplicateCommand(cid, tgt, src); + pendingMoveOps.putIfAbsent(cid, mp); + pendingMoveFuture.putIfAbsent(cid, ret); + } else { + ret.complete(MoveResult.FAIL_CAN_NOT_RECORD_TO_DB); + } + } + } + + /** + * notify MoveManager that the current scm has become leader and ready. + */ + @Override + public void onLeaderReady() { + //discard all stale records + pendingMoveOps.clear(); + pendingMoveFuture.clear(); + + //initialize with db record + try (TableIterator<ContainerID, + ? extends Table.KeyValue<ContainerID, MoveDataNodePair>> iterator = + moveDBHandler.getMoveTable().iterator()) { + while (iterator.hasNext()) { + Table.KeyValue<ContainerID, MoveDataNodePair> kv = iterator.next(); + final ContainerID cid = kv.getKey(); + final MoveDataNodePair mp = kv.getValue(); + Preconditions.assertNotNull(cid, + "moved container id should not be null"); + Preconditions.assertNotNull(mp, + "MoveDataNodePair container id should not be null"); + pendingMoveOps.put(cid, mp); + } + } catch (IOException ex) { + ExitUtils.terminate(1, + "Resuming MoveManager failed on SCM leader change.", + ex, true, LOG); + } + + pendingMoveOps.forEach((cid, mdmp) -> { + Set<ContainerReplica> replicas; + try { + replicas = containerManager.getContainerReplicas(cid); + } catch (ContainerNotFoundException e) { + LOG.error("can not find container {} " + + "while processing replicated move", cid); + completeMove(cid, null); + return; + } + DatanodeDetails src = mdmp.getSrc(); + DatanodeDetails tgt = mdmp.getTgt(); + + boolean isSrcExist = replicas.stream() + .anyMatch(r -> r.getDatanodeDetails().equals(src)); + boolean isTgtExist = replicas.stream() + .anyMatch(r -> r.getDatanodeDetails().equals(tgt)); + if (isSrcExist) { + if (isTgtExist) { + //the former scm leader may or may not send the deletion command + //before reelection.here, we just try to send the command again. + try { + handleSuccessfulAdd(cid); + } catch (ContainerReplicaNotFoundException | + NodeNotFoundException | + ContainerNotFoundException e) { + LOG.warn("Can not handle successful Add for move : {}", e); + completeMove(cid, null); + } + } else { + // resenting replication command is ok , no matter whether there is an + // on-going replication + try { + sendReplicateCommand(cid, tgt, src); + } catch (ContainerReplicaNotFoundException | + ContainerNotFoundException e) { + LOG.warn("can not send replication command, {}", e); + completeMove(cid, null); + } + } + } else { + // if container does not exist in src datanode, no matter it exists + // in target datanode, we can not take more actions to this option, + // so just remove it through ratis + completeMove(cid, null); + } + }); + + running = true; + } + + /** + * notify MoveManager that the current scm leader steps down. + */ + @Override + public void onNotLeader() { + running = false; + } + + /** + * move a container replica from source datanode to + * target datanode. + * + * @param cid Container to move + * @param src source datanode + * @param tgt target datanode + */ + @Override + public CompletableFuture<MoveResult> move( + ContainerID cid, DatanodeDetails src, DatanodeDetails tgt) + throws ContainerNotFoundException, NodeNotFoundException, + TimeoutException, ContainerReplicaNotFoundException { + CompletableFuture<MoveResult> ret = new CompletableFuture<>(); + + if (!running) { + ret.complete(MoveResult.FAIL_LEADER_NOT_READY); + return ret; + } + + /* + * make sure the flowing conditions are met: + * 1 the given two datanodes are in healthy state + * 2 the given container exists on the given source datanode + * 3 the given container does not exist on the given target datanode + * 4 the given container is in closed state + * 5 the giver container is not taking any inflight action + * 6 the given two datanodes are in IN_SERVICE state + * 7 {Existing replicas + Target_Dn - Source_Dn} satisfies + * the placement policy + * + * move is a combination of two steps : replication and deletion. + * if the conditions above are all met, then we take a conservative + * strategy here : replication can always be executed, but the execution + * of deletion always depends on placement policy + */ + + NodeStatus currentNodeStat = nodeManager.getNodeStatus(src); + HddsProtos.NodeState healthStat = currentNodeStat.getHealth(); + HddsProtos.NodeOperationalState operationalState = + currentNodeStat.getOperationalState(); + if (healthStat != HddsProtos.NodeState.HEALTHY) { + ret.complete(MoveResult.REPLICATION_FAIL_NODE_UNHEALTHY); + return ret; + } + if (operationalState != HddsProtos.NodeOperationalState.IN_SERVICE) { + ret.complete(MoveResult.REPLICATION_FAIL_NODE_NOT_IN_SERVICE); + return ret; + } + + currentNodeStat = nodeManager.getNodeStatus(tgt); + healthStat = currentNodeStat.getHealth(); + operationalState = currentNodeStat.getOperationalState(); + if (healthStat != HddsProtos.NodeState.HEALTHY) { + ret.complete(MoveResult.REPLICATION_FAIL_NODE_UNHEALTHY); + return ret; + } + if (operationalState != HddsProtos.NodeOperationalState.IN_SERVICE) { + ret.complete(MoveResult.REPLICATION_FAIL_NODE_NOT_IN_SERVICE); + return ret; + } + + // we need to synchronize on ContainerInfo, since it is + // shared by ICR/FCR handler and this.processContainer + // TODO: use a Read lock after introducing a RW lock into ContainerInfo + ContainerInfo cif = containerManager.getContainer(cid); + final Set<ContainerReplica> currentReplicas = containerManager + .getContainerReplicas(cid); + final Set<DatanodeDetails> replicas = currentReplicas.stream() + .map(ContainerReplica::getDatanodeDetails) + .collect(Collectors.toSet()); + if (replicas.contains(tgt)) { + ret.complete(MoveResult.REPLICATION_FAIL_EXIST_IN_TARGET); + return ret; + } + if (!replicas.contains(src)) { + ret.complete(MoveResult.REPLICATION_FAIL_NOT_EXIST_IN_SOURCE); + return ret; + } + + /* + * the reason why the given container should not be taking any inflight + * action is that: if the given container is being replicated or deleted, + * the num of its replica is not deterministic, so move operation issued + * by balancer may cause a nondeterministic result, so we should drop + * this option for this time. + * */ + List<ContainerReplicaOp> pendingOps = + scm.getContainerReplicaPendingOps().getPendingOps(cid); + + if (pendingOps.stream().anyMatch(op -> + op.getOpType() == ContainerReplicaOp.PendingOpType.ADD)) { + ret.complete(MoveResult.REPLICATION_FAIL_INFLIGHT_REPLICATION); + return ret; + } + + if (pendingOps.stream().anyMatch(op -> + op.getOpType() == ContainerReplicaOp.PendingOpType.DELETE)) { + ret.complete(MoveResult.REPLICATION_FAIL_INFLIGHT_DELETION); + return ret; + } + + /* + * here, no need to see whether cid is in inflightMove, because + * these three map are all synchronized on ContainerInfo, if cid + * is in infligtMove , it must now being replicated or deleted, + * so it must be in inflightReplication or in infligthDeletion. + * thus, if we can not find cid in both of them , this cid must + * not be in inflightMove. + */ + + HddsProtos.LifeCycleState currentContainerStat = cif.getState(); + if (currentContainerStat != HddsProtos.LifeCycleState.CLOSED) { + ret.complete(MoveResult.REPLICATION_FAIL_CONTAINER_NOT_CLOSED); + return ret; + } + + // check whether {Existing replicas + Target_Dn - Source_Dn} + // satisfies current placement policy + if (!isPolicySatisfiedAfterMove(cif, src, tgt)) { + ret.complete(MoveResult.PLACEMENT_POLICY_NOT_SATISFIED); + return ret; + } + startMove(cid, src, tgt, ret); + LOG.info("receive a move request about container {} , from {} to {}", + cid, src.getUuid(), tgt.getUuid()); + return ret; + } + + /** + * notify move manager that a container op has been completed. + * + * @param cop ContainerReplicaOp + * @param containerID ContainerID for which to complete + */ + @Override + public void notifyContainerOpCompleted(ContainerReplicaOp cop, + ContainerID containerID) { + if (!running) { + return; + } + + MoveDataNodePair mdnp = pendingMoveOps.get(containerID); + if (mdnp != null) { + PendingOpType opType = cop.getOpType(); + DatanodeDetails dn = cop.getTarget(); + if (opType.equals(PendingOpType.ADD) && mdnp.getTgt().equals(dn)) { + if (pendingMoveFuture.containsKey(containerID)) { + try { + handleSuccessfulAdd(containerID); + } catch (ContainerNotFoundException | NodeNotFoundException | + ContainerReplicaNotFoundException e) { + LOG.warn("Can not handle successful Add for move : {}", e); + } + } else { + //if MoveManager is reinitialize, pendingMoveOps will be restored, + //but pendingMoveFuture not. so there will be a case that + //container is in pendingMoveOps, but not in pendingMoveFuture. + // we just complete this move for now. + // also, we can continue this move, to be discussed. + completeMove(containerID, null); + } + } else if (opType.equals(PendingOpType.DELETE) + && mdnp.getSrc().equals(dn)) { + completeMove(containerID, MoveResult.COMPLETED); + } + } + } + + /** + * notify move manager that a container op has been Expired. + * + * @param cop ContainerReplicaOp + * @param containerID ContainerID for which to complete + */ + @Override + public void notifyContainerOpExpired(ContainerReplicaOp cop, + ContainerID containerID) { + if (!running) { + return; + } + + MoveDataNodePair mdnp = pendingMoveOps.get(containerID); + if (mdnp != null) { + PendingOpType opType = cop.getOpType(); + DatanodeDetails dn = cop.getTarget(); + if (opType.equals(PendingOpType.ADD) && mdnp.getTgt().equals(dn)) { + completeMove(containerID, MoveResult.REPLICATION_FAIL_TIME_OUT); + } else if (opType.equals(PendingOpType.DELETE) && + mdnp.getSrc().equals(dn)) { + completeMove(containerID, MoveResult.DELETION_FAIL_TIME_OUT); + } + } + } + + private void handleSuccessfulAdd(final ContainerID cid) + throws ContainerNotFoundException, + ContainerReplicaNotFoundException, NodeNotFoundException { + MoveDataNodePair movePair = pendingMoveOps.get(cid); + if (movePair == null) { + return; + } + final DatanodeDetails src = movePair.getSrc(); + final Set<ContainerReplica> replicaSet = + containerManager.getContainerReplicas(cid); + if (!replicaSet.stream() + .anyMatch(r -> r.getDatanodeDetails().equals(src))) { + // if the target is present but source disappears somehow, + // we can consider move is successful. + completeMove(cid, MoveResult.COMPLETED); + return; + } + + final NodeStatus nodeStatus = nodeManager.getNodeStatus(src); + if (!nodeStatus.isInService()) { + completeMove(cid, MoveResult.DELETION_FAIL_NODE_NOT_IN_SERVICE); + return; + } + if (!nodeStatus.isHealthy()) { + completeMove(cid, MoveResult.DELETION_FAIL_NODE_UNHEALTHY); + return; + } + + final ContainerInfo cif = containerManager.getContainer(cid); Review Comment: I feel this logic needs to go via `ReplicationManager.processOverReplicatedContainer` so we are not duplicating the OverRep handling logic here. We can change this interface a bit to make it more suitable and allow for passing some "preferred delete" replica. We still need to add the handlers for Ratis and make the interface work with both of them. I am also not sure these processUnderReplicatedContainer / processOverReplicatedContainer method need to be given a ContainerHealthResult object as their input. That is what is on the underReplicated queue, but I am not sure if we could get away with just passing the containerInfo instead. -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
