Github user tillrohrmann commented on a diff in the pull request:
https://github.com/apache/flink/pull/2388#discussion_r75698687
--- Diff:
flink-runtime/src/main/java/org/apache/flink/runtime/clusterframework/SlotManager.java
---
@@ -0,0 +1,519 @@
+/*
+ * 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.flink.runtime.clusterframework;
+
+import com.google.common.collect.BiMap;
+import com.google.common.collect.HashBiMap;
+import org.apache.flink.annotation.VisibleForTesting;
+import org.apache.flink.runtime.clusterframework.types.AllocationID;
+import org.apache.flink.runtime.clusterframework.types.ResourceID;
+import org.apache.flink.runtime.clusterframework.types.ResourceProfile;
+import org.apache.flink.runtime.clusterframework.types.ResourceSlot;
+import org.apache.flink.runtime.clusterframework.types.SlotID;
+import org.apache.flink.runtime.rpc.resourcemanager.ResourceManagerGateway;
+import org.apache.flink.runtime.rpc.resourcemanager.SlotRequest;
+import org.apache.flink.runtime.rpc.taskexecutor.SlotReport;
+import org.apache.flink.runtime.rpc.taskexecutor.SlotStatus;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.util.HashMap;
+import java.util.LinkedHashMap;
+import java.util.Map;
+
+import static org.apache.flink.util.Preconditions.checkNotNull;
+
+/**
+ * SlotManager is responsible for receiving slot requests and do slot
allocations. It allows to request
+ * slots from registered TaskManagers and issues container allocation
requests in case of there are not
+ * enough available slots. Besides, it should sync its slot allocation
with TaskManager's heartbeat.
+ * <p>
+ * The main operation principle of SlotManager is:
+ * <ul>
+ * <li>1. All slot allocation status should be synced with TaskManager,
which is the ground truth.</li>
+ * <li>2. All slots that have registered must be tracked, either by free
pool or allocated pool.</li>
+ * <li>3. All slot requests will be handled by best efforts, there is no
guarantee that one request will be
+ * fulfilled in time or correctly allocated. Conflicts or timeout or some
special error will happen, it should
+ * be handled outside SlotManager. SlotManager will make each decision
based on the information it currently
+ * holds.</li>
+ * </ul>
+ * <b>IMPORTANT:</b> This class is <b>Not Thread-safe</b>.
+ */
+public abstract class SlotManager {
+
+ private static final Logger LOG =
LoggerFactory.getLogger(SlotManager.class);
+
+ /** Gateway to communicate with ResourceManager */
+ private final ResourceManagerGateway resourceManagerGateway;
+
+ /** All registered slots, including free and allocated slots */
+ private final Map<ResourceID, Map<SlotID, ResourceSlot>>
registeredSlots;
+
+ /** All pending slot requests, waiting available slots to fulfil */
+ private final Map<AllocationID, SlotRequest> pendingSlotRequests;
+
+ /** All free slots that can be used to be allocated */
+ private final Map<SlotID, ResourceSlot> freeSlots;
+
+ /** All allocations, we can lookup allocations either by SlotID or
AllocationID */
+ private final AllocationMap allocationMap;
+
+ public SlotManager(ResourceManagerGateway resourceManagerGateway) {
+ this.resourceManagerGateway = resourceManagerGateway;
+ this.registeredSlots = new HashMap<>();
+ this.pendingSlotRequests = new LinkedHashMap<>();
+ this.freeSlots = new HashMap<>();
+ this.allocationMap = new AllocationMap();
+ }
+
+ //
------------------------------------------------------------------------
+ // slot managements
+ //
------------------------------------------------------------------------
+
+ /**
+ * Request a slot with requirements, we may either fulfill the request
or pending it. Trigger container
+ * allocation if we don't have enough resource. If we have free slot
which can match the request, record
+ * this allocation and forward the request to TaskManager through
ResourceManager (we want this done by
+ * RPC's main thread to avoid race condition).
+ *
+ * @param request The detailed request of the slot
+ */
+ public void requestSlot(final SlotRequest request) {
+ if (isRequestDuplicated(request)) {
+ LOG.warn("Duplicated slot request, AllocationID:{}",
request.getAllocationId());
+ return;
+ }
+
+ // try to fulfil the request with current free slots
+ ResourceSlot slot = chooseSlotToUse(request, freeSlots);
+ if (slot != null) {
+ LOG.info("Assigning SlotID({}) to AllocationID({}),
JobID:{}", slot.getSlotId(),
+ request.getAllocationId(), request.getJobId());
+
+ // record this allocation in bookkeeping
+ allocationMap.addAllocation(slot.getSlotId(),
request.getAllocationId());
+
+ // remove selected slot from free pool
+ freeSlots.remove(slot.getSlotId());
+
+ // TODO: send slot request to TaskManager
+ } else {
+ LOG.info("Does not have enough slot now, pending
request and try to allocate new container, " +
+ "AllocationID:{}, JobID:{}",
request.getAllocationId(), request.getJobId());
+ allocateContainer(request.getResourceProfile());
+ pendingSlotRequests.put(request.getAllocationId(),
request);
+ }
+ }
+
+ /**
+ * Sync slot status with TaskManager's SlotReport.
+ */
+ public void updateSlotStatus(final SlotReport slotReport) {
+ for (SlotStatus slotStatus : slotReport.getSlotsStatus()) {
+ updateSlotStatus(slotStatus);
+ }
+ }
+
+ /**
+ * The slot request to TaskManager may be either failed by rpc
communication(timeout, network error, etc.)
+ * or really rejected by TaskManager. We shall retry this request by:
+ * <ul>
+ * <li>1. verify and clear all the previous allocate information for
this request
+ * <li>2. try to request slot again
+ * </ul>
+ * <p>
+ * This may cause some duplicate allocation, e.g. the slot request to
TaskManager is successful but the response
+ * is lost somehow, so we may request a slot in another TaskManager,
this causes two slots assigned to one request,
+ * but it can be taken care of by rejecting registration at JobManager.
+ *
+ * @param originalRequest The original slot request
+ * @param slotId The target SlotID
+ */
+ public void handleSlotRequestFailedAtTaskManager(final SlotRequest
originalRequest, final SlotID slotId) {
+ final AllocationID originalAllocationId =
originalRequest.getAllocationId();
+ LOG.info("Slot request failed at TaskManager, SlotID:{},
AllocationID:{}, JobID:{}",
+ slotId, originalAllocationId,
originalRequest.getJobId());
+
+ // verify the allocation info before we do anything
+ if (freeSlots.containsKey(slotId)) {
+ // this slot is currently empty, no need to de-allocate
it from our allocations
+ LOG.info("Original slot is somehow empty, retrying this
request");
+
+ // before retry, we should double check whether this
request was allocated by some other ways
+ if (!allocationMap.isAllocated(originalAllocationId)) {
+ requestSlot(originalRequest);
+ } else {
+ LOG.info("The failed request is somehow been
allocated, SlotID:{}",
+
allocationMap.getSlotID(originalAllocationId));
+ }
+ } else if (allocationMap.isAllocated(slotId)) {
+ final AllocationID currentAllocationId =
allocationMap.getAllocationID(slotId);
+
+ // check whether we have an agreement on whom this slot
belongs to
+ if (originalAllocationId.equals(currentAllocationId)) {
+ LOG.info("De-allocate this request and retry");
+
allocationMap.removeAllocation(currentAllocationId);
+
+ // put this slot back to free pool
+ ResourceSlot slot =
checkNotNull(getRegisteredSlot(slotId));
+ freeSlots.put(slotId, slot);
+
+ // retry the request
+ requestSlot(originalRequest);
+ } else {
+ // the slot is taken by someone else, no need
to de-allocate it from our allocations
+ LOG.info("Original slot is taken by someone
else, current AllocationID:{}", currentAllocationId);
+
+ // before retry, we should double check whether
this request was allocated by some other ways
+ if
(!allocationMap.isAllocated(originalAllocationId)) {
+ requestSlot(originalRequest);
+ } else {
+ LOG.info("The failed request is somehow
been allocated, SlotID:{}",
+
allocationMap.getSlotID(originalAllocationId));
+ }
+ }
+ } else {
+ LOG.error("BUG! {} is neither in free pool nor in
allocated pool", slotId);
+ }
+ }
+
+ /**
+ * TaskManager been notified failure, we should clean up all its slots.
+ *
+ * @param resourceId The ResourceID of the TaskManager
+ */
+ public void notifyTaskManagerFailure(final ResourceID resourceId) {
+ LOG.info("Resource:{} been notified failure", resourceId);
+ final Map<SlotID, ResourceSlot> slotIdsToRemove =
registeredSlots.remove(resourceId);
+ if (slotIdsToRemove == null) {
+ return;
+ }
+ for (SlotID slotId : slotIdsToRemove.keySet()) {
+ LOG.info("Removing Slot:{} upon resource failure",
slotId);
+ if (freeSlots.containsKey(slotId)) {
+ freeSlots.remove(slotId);
+ } else if (allocationMap.isAllocated(slotId)) {
+ allocationMap.removeAllocation(slotId);
+ } else {
+ LOG.error("BUG! {} is neither in free pool nor
in allocated pool", slotId);
+ }
+ }
+ }
+
+ //
------------------------------------------------------------------------
+ // internal behaviors
+ //
------------------------------------------------------------------------
+
+ /**
+ * Update slot status based on TaskManager's report. There are mainly
two situations when we receive the report:
+ * <ul>
+ * <li>1. The slot is newly registered.</li>
+ * <li>2. The slot has registered, it contains its current status.</li>
+ * </ul>
+ * <p>
+ * Regarding 1: It's fairly simple, we just record this slot's status,
and trigger schedule if sot is empty.
+ * <p>
+ * Regarding 2: It will cause some weird situation since we may have
some time-gap on how the slot's status really
+ * is. We may have some updates on the slot's allocation, but it
doesn't reflected by TaskManager's heartbeat yet,
+ * and we may make some wrong decision if we cannot guarantee we have
the exactly status about all the slots. So
--- End diff --
Typo: exact status
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
