Many resources are needed to download a project. Please understand that we have to compensate our server costs. Thank you in advance. Project price only 1 $
You can buy this project and download/modify it how often you want.
/*
* Copyright (c) 2014-2015 VMware, Inc. All Rights Reserved.
*
* Licensed 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 com.vmware.xenon.services.common;
import java.net.URI;
import java.util.ArrayList;
import java.util.Collections;
import java.util.EnumSet;
import java.util.HashMap;
import java.util.List;
import java.util.Map;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.ThreadLocalRandom;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicInteger;
import com.vmware.xenon.common.Operation;
import com.vmware.xenon.common.Operation.CompletionHandler;
import com.vmware.xenon.common.RequestRouter.Route.RouteDocumentation;
import com.vmware.xenon.common.ServiceClient;
import com.vmware.xenon.common.ServiceDocument;
import com.vmware.xenon.common.ServiceHost.ServiceHostState;
import com.vmware.xenon.common.ServiceStatUtils;
import com.vmware.xenon.common.ServiceStats.ServiceStat;
import com.vmware.xenon.common.ServiceStats.TimeSeriesStats.AggregationType;
import com.vmware.xenon.common.StatefulService;
import com.vmware.xenon.common.UriUtils;
import com.vmware.xenon.common.Utils;
import com.vmware.xenon.services.common.NodeState.NodeOption;
import com.vmware.xenon.services.common.NodeState.NodeStatus;
/**
* Service for maintaining a list of nodes through the use of a membership gossip layer. New nodes
* are added to a group through POST
*/
public class NodeGroupService extends StatefulService {
public static final String STAT_NAME_JOIN_RETRY_COUNT = "joinRetryCount";
public static final String PROPERTY_NAME_PEER_REQUEST_TIMEOUT_MICROS = Utils.PROPERTY_NAME_PREFIX
+ "NodeGroupService.peerRequestTimeoutMicros";
public static final long PEER_REQUEST_TIMEOUT_MICROS = Long.getLong(
PROPERTY_NAME_PEER_REQUEST_TIMEOUT_MICROS,
ServiceHostState.DEFAULT_OPERATION_TIMEOUT_MICROS / 3);
private enum NodeGroupChange {
PEER_ADDED, PEER_STATUS_CHANGE, SELF_CHANGE
}
public static class JoinPeerRequest {
public static final String KIND = Utils.buildKind(JoinPeerRequest.class);
public static JoinPeerRequest create(URI peerToJoin, Integer quorum) {
JoinPeerRequest r = new JoinPeerRequest();
r.memberGroupReference = peerToJoin;
r.membershipQuorum = quorum;
r.kind = KIND;
return r;
}
/**
* Member of the group we wish to join through
*/
public URI memberGroupReference;
/**
* Optional node join options. If specified the node state representing the local node
* will be updated with these options. Further, these options determine join behavior.
*/
public EnumSet localNodeOptions;
/**
* See {@link NodeState#membershipQuorum}
*/
public Integer membershipQuorum;
/**
* See {@link NodeState#locationQuorum}
*/
public Integer locationQuorum;
public String kind;
}
public static class UpdateQuorumRequest {
public static final String KIND = Utils.buildKind(UpdateQuorumRequest.class);
public static UpdateQuorumRequest create(boolean isGroupUpdate) {
UpdateQuorumRequest r = new UpdateQuorumRequest();
r.isGroupUpdate = isGroupUpdate;
r.kind = KIND;
return r;
}
public UpdateQuorumRequest setMembershipQuorum(int count) {
this.membershipQuorum = count;
return this;
}
public UpdateQuorumRequest setLocationQuorum(int count) {
this.locationQuorum = count;
return this;
}
public boolean isGroupUpdate;
public Integer membershipQuorum;
public Integer locationQuorum;
public String kind;
}
public static class NodeGroupConfig {
public static final long DEFAULT_NODE_REMOVAL_DELAY_MICROS = TimeUnit.MINUTES.toMicros(5);
public long nodeRemovalDelayMicros = DEFAULT_NODE_REMOVAL_DELAY_MICROS;
/**
* Number of maintenance intervals after last update to group membership before owner
* selection and replication requests should be processed.
*/
public long stableGroupMaintenanceIntervalCount = 5;
/**
* Timeout for gossip requests to peers, in microseconds. The default is smaller than the operation timeout
* so we have the chance to mark a non responsive peer as unavailable, and retry pending operations
* before they expire.
*/
public long peerRequestTimeoutMicros = PEER_REQUEST_TIMEOUT_MICROS;
}
public static class NodeGroupState extends ServiceDocument {
/**
* The node group configuration
*/
public NodeGroupConfig config;
/**
* The map of peer nodes, updated through random probing of a limited number of peers and
* two way state merges
*/
public Map nodes = new ConcurrentHashMap<>();
/**
* The maximum value among all reported times from the peers. If one peer has significant
* time drift compared to others, this value will appears in the future or past, compared to local time.
* This value is updated during gossip and is considered a "global" field that settles to the same
* value across all peers, when gossip ahs converged
*/
public long membershipUpdateTimeMicros;
/**
* The local membership update time, as observed by each node. This value is only updated
* by the local node and not merged with other peer reported values. It is used to determine
* node group stability, a heuristic used in
* {@link NodeGroupUtils#isMembershipSettled(com.vmware.xenon.common.ServiceHost, long, NodeGroupState)}
*/
public long localMembershipUpdateTimeMicros;
}
public static final int MIN_PEER_GOSSIP_COUNT = 10;
public static final String STAT_NAME_RESTARTING_SERVICES_COUNT = "restartingServicesCount";
public static final String STAT_NAME_RESTARTING_SERVICES_FAILURE_COUNT = "restartingServicesFailureCount";
public static final String STAT_NAME_PREFIX_GOSSIP_PATCH_DURATION = "GossipPatchDurationMicros";
private URI uri;
private URI publicUri;
private NodeGroupState cachedState;
public NodeGroupService() {
super(NodeGroupState.class);
super.toggleOption(ServiceOption.CORE, true);
super.toggleOption(ServiceOption.INSTRUMENTATION, true);
super.toggleOption(ServiceOption.PERIODIC_MAINTENANCE, true);
super.toggleOption(ServiceOption.CONCURRENT_GET_HANDLING, true);
}
@Override
public URI getUri() {
if (this.uri == null) {
this.uri = super.getUri();
}
return this.uri;
}
private URI getPublicUri() {
if (this.publicUri == null) {
this.publicUri = UriUtils.buildPublicUri(getHost(), getSelfLink());
}
return this.publicUri;
}
@Override
public void handleStart(Operation startPost) {
NodeGroupState initState = null;
if (startPost.hasBody()) {
initState = startPost.getBody(NodeGroupState.class);
} else {
initState = new NodeGroupState();
}
initState.documentOwner = getHost().getId();
if (initState.config == null) {
initState.config = new NodeGroupConfig();
}
NodeState self = initState.nodes.get(getHost().getId());
self = buildLocalNodeState(self);
if (!validateNodeOptions(startPost, self.options)) {
return;
}
initState.nodes.put(self.id, self);
this.cachedState = Utils.clone(initState);
startPost.setBody(this.cachedState).complete();
}
@Override
public void handleGet(Operation get) {
get.setBody(this.cachedState).complete();
}
@Override
public void handlePatch(Operation patch) {
NodeGroupState body = getStateFromBody(patch);
if (body == null) {
patch.fail(new IllegalArgumentException("body of type NodeGroupState is required"));
return;
}
patch.setContentType(Operation.MEDIA_TYPE_APPLICATION_JSON);
NodeGroupState localState = getState(patch);
if (localState == null || localState.nodes == null) {
this.cachedState = localState;
logWarning("Invalid local state");
patch.fail(Operation.STATUS_CODE_FAILURE_THRESHOLD);
return;
}
if (body.config == null && body.nodes.isEmpty()) {
UpdateQuorumRequest bd = patch.getBody(UpdateQuorumRequest.class);
if (UpdateQuorumRequest.KIND.equals(bd.kind)) {
handleUpdateQuorumPatch(patch, localState);
return;
}
patch.fail(new IllegalArgumentException("nodes or config are required"));
return;
}
if (body.config != null && body.nodes.isEmpty()) {
localState.config = body.config;
this.cachedState = Utils.clone(localState);
patch.complete();
return;
}
EnumSet changes = EnumSet.noneOf(NodeGroupChange.class);
mergeRemoteAndLocalMembership(
localState,
body,
changes);
patch.setNotificationDisabled(changes.isEmpty());
localState.documentOwner = getHost().getId();
NodeState localNodeState = localState.nodes.get(getHost().getId());
localNodeState.groupReference = getPublicUri();
this.cachedState = localState;
patch.setBodyNoCloning(this.cachedState).complete();
if (!isAvailable()) {
boolean isAvailable = NodeGroupUtils.isNodeGroupAvailable(getHost(), localState);
setAvailable(isAvailable);
}
if (localNodeState.status == NodeStatus.AVAILABLE) {
return;
}
localNodeState.status = NodeStatus.AVAILABLE;
sendAvailableSelfPatch(localNodeState);
}
private void handleUpdateQuorumPatch(Operation patch,
NodeGroupState localState) {
UpdateQuorumRequest bd = patch.getBody(UpdateQuorumRequest.class);
NodeState self = localState.nodes.get(getHost().getId());
if (bd.membershipQuorum != null) {
self.membershipQuorum = Math.max(1, bd.membershipQuorum);
}
if (bd.locationQuorum != null) {
self.locationQuorum = Math.max(1, bd.locationQuorum);
}
self.documentVersion++;
self.documentUpdateTimeMicros = Utils.getNowMicrosUtc();
localState.membershipUpdateTimeMicros = self.documentUpdateTimeMicros;
localState.localMembershipUpdateTimeMicros = self.documentUpdateTimeMicros;
this.cachedState = Utils.clone(localState);
if (!bd.isGroupUpdate) {
patch.complete();
return;
}
// TODO use a three phase consensus algorithm to update quorum similar
// to the steady state replication consensus.
// Issue N requests to update quorum to all member of the group. If they
// do not all succeed the request, then the operation fails and some peers
// will be left with a quorum level different than the others. That is
// acceptable. The replication logic, can reject a peer if its quorum level
// is not set at the same level as the owner. The client of this request can
// also retry...
bd.isGroupUpdate = false;
int failureThreshold = (localState.nodes.size() - 1) / 2;
AtomicInteger pending = new AtomicInteger(localState.nodes.size());
AtomicInteger failures = new AtomicInteger();
CompletionHandler c = (o, e) -> {
if (e != null) {
logWarning("Node %s failed quorum update: %s", o.getUri(), e.toString());
failures.incrementAndGet();
}
int p = pending.decrementAndGet();
if (p != 0) {
return;
}
if (failures.get() > failureThreshold) {
patch.fail(new IllegalStateException("Majority of nodes failed request"));
} else {
if (bd.membershipQuorum != null) {
logInfo("Quorum updated across all peers to %d", bd.membershipQuorum);
}
if (bd.locationQuorum != null) {
logInfo("Location quorum updated across all peers to %d", bd.locationQuorum);
}
patch.complete();
}
};
for (NodeState node : localState.nodes.values()) {
if (!NodeState.isAvailable(node, getHost().getId(), true)) {
c.handle(null, null);
continue;
}
if (bd.membershipQuorum != null) {
node.membershipQuorum = bd.membershipQuorum;
}
if (bd.locationQuorum != null) {
node.locationQuorum = bd.locationQuorum;
}
Operation p = Operation
.createPatch(node.groupReference)
.setBody(bd)
.setCompletion(c);
sendRequest(p);
}
}
/**
* Handles a POST to either join this service to a group using a peer existing member
* (JoinPeerRequest as the body) or add a new local monitor service to track the state of a
* remote peer
*
* @param post
*/
@RouteDocumentation(
description = "Join a peer to this node group",
requestBodyType = JoinPeerRequest.class)
@Override
public void handlePost(Operation post) {
if (!post.hasBody()) {
post.fail(new IllegalArgumentException("body is required"));
return;
}
NodeGroupState localState = getState(post);
if (localState == null || localState.nodes == null) {
logWarning("invalid local state");
post.fail(Operation.STATUS_CODE_BAD_REQUEST);
return;
}
JoinPeerRequest joinBody = post.getBody(JoinPeerRequest.class);
if (joinBody != null && joinBody.memberGroupReference != null) {
// set a short join operation timeout so that join retries will occur in any environment
long joinTimeOutMicrosUtc = Utils.fromNowMicrosUtc(
Math.max(TimeUnit.SECONDS.toMicros(1),
getHost().getOperationTimeoutMicros() / 10));
handleJoinPost(joinBody, post, joinTimeOutMicrosUtc, getState(post), null);
return;
}
NodeState body = post.getBody(NodeState.class);
if (body.id == null) {
post.fail(new IllegalArgumentException("id is required"));
return;
}
boolean isLocalNode = body.id.equals(getHost().getId());
if (body.groupReference == null) {
post.fail(new IllegalArgumentException("groupReference is required"));
return;
}
if (isLocalNode) {
// this is a node instance representing us
buildLocalNodeState(body);
} else {
body.documentSelfLink = UriUtils.buildUriPath(getSelfLink(), body.id);
}
localState.nodes.put(body.id, body);
this.cachedState = Utils.clone(localState);
post.setBodyNoCloning(this.cachedState).complete();
}
private void handleJoinPost(JoinPeerRequest joinBody,
Operation joinOp,
long expirationMicros,
NodeGroupState localState,
NodeGroupState remotePeerState) {
if (UriUtils.isHostEqual(getHost(), joinBody.memberGroupReference)) {
logInfo("Skipping self join");
// we tried joining ourself, abort;
joinOp.complete();
return;
}
NodeState self = localState.nodes.get(getHost().getId());
if (joinOp != null) {
self.documentUpdateTimeMicros = Utils.getNowMicrosUtc();
self.documentVersion++;
if (joinBody.membershipQuorum != null) {
if (!joinBody.membershipQuorum.equals(self.membershipQuorum)) {
logInfo("Quorum changed from %d to %d", self.membershipQuorum,
joinBody.membershipQuorum);
}
self.membershipQuorum = joinBody.membershipQuorum;
}
if (joinBody.locationQuorum != null) {
if (!joinBody.locationQuorum.equals(self.locationQuorum)) {
logInfo("Location quorum changed from %d to %d", self.locationQuorum,
joinBody.locationQuorum);
}
self.locationQuorum = joinBody.locationQuorum;
}
if (joinBody.localNodeOptions != null) {
if (!validateNodeOptions(joinOp, joinBody.localNodeOptions)) {
return;
}
self.options = joinBody.localNodeOptions;
}
localState.membershipUpdateTimeMicros = self.documentUpdateTimeMicros;
this.cachedState = Utils.clone(localState);
// complete the join POST, continue with state merge
joinOp.complete();
}
// this method is two pass
// First pass get the remote peer state
// Second pass, insert self
if (remotePeerState == null) {
// Pass 1, get existing member state
sendRequest(Operation.createGet(joinBody.memberGroupReference)
.setCompletion(
(o, e) -> {
if (e != null) {
handleJoinFailure(e, joinBody, localState,
expirationMicros);
return;
}
NodeGroupState remoteState = getStateFromBody(o);
handleJoinPost(joinBody, null, expirationMicros,
localState, remoteState);
}));
return;
}
// Pass 2, merge remote group state with ours, send self to peer
sendRequest(Operation.createPatch(getUri()).setBody(remotePeerState));
logInfo("Sending POST to %s to insert self: %s",
joinBody.memberGroupReference, Utils.toJson(self));
Operation insertSelfToPeer = Operation
.createPost(joinBody.memberGroupReference)
.setBody(self)
.setCompletion(
(o, e) -> {
if (e != null) {
logSevere("Insert POST to %s failed", o.getUri());
}
// we will restart services to synchronize with peers on the next
// maintenance interval with a stable group membership
});
sendRequest(insertSelfToPeer);
}
private void handleJoinFailure(Throwable e, JoinPeerRequest joinBody,
NodeGroupState localState,
long expirationMicros) {
if (Utils.beforeNow(expirationMicros)) {
logSevere("Failure joining peer %s due to %s, attempt expired, will not retry",
joinBody.memberGroupReference, e.toString());
return;
}
// avoid rescheduling if the host is in the process of stopping
if (getHost().isStopping()) {
return;
}
getHost().scheduleCore(() -> {
logWarning("Retrying GET to %s, due to %s",
joinBody.memberGroupReference,
e.toString());
handleJoinPost(joinBody, null, expirationMicros, localState, null);
adjustStat(STAT_NAME_JOIN_RETRY_COUNT, 1);
}, getHost().getMaintenanceIntervalMicros(), TimeUnit.MICROSECONDS);
}
private boolean validateNodeOptions(Operation joinOp, EnumSet options) {
if (options.isEmpty()) {
joinOp.fail(new IllegalArgumentException("at least one option must be specified"));
return false;
}
if (options.contains(NodeOption.OBSERVER) && options.contains(NodeOption.PEER)) {
joinOp.fail(new IllegalArgumentException(
String.format("%s and %s are mutually exclusive",
NodeOption.OBSERVER, NodeOption.PEER)));
return false;
}
return true;
}
private void sendAvailableSelfPatch(NodeState local) {
// mark self as available by issuing self PATCH
NodeGroupState body = new NodeGroupState();
body.config = null;
body.documentOwner = getHost().getId();
body.documentSelfLink = UriUtils.buildUriPath(getSelfLink(), body.documentOwner);
local.status = NodeStatus.AVAILABLE;
body.nodes.put(local.id, local);
sendRequest(Operation.createPatch(getUri()).setBody(
body));
}
private NodeState buildLocalNodeState(NodeState body) {
if (body == null) {
body = new NodeState();
}
body.id = getHost().getId();
body.status = NodeStatus.SYNCHRONIZING;
Integer q = Integer.getInteger(NodeState.PROPERTY_NAME_MEMBERSHIP_QUORUM);
if (q != null) {
body.membershipQuorum = q;
} else {
// Initialize default quorum based on service host peerHosts argument
int total = getHost().getInitialPeerHosts().size() + 1;
int quorum = (total / 2) + 1;
body.membershipQuorum = Math.max(1, quorum);
}
Integer lq = Integer.getInteger(NodeState.PROPERTY_NAME_LOCATION_QUORUM);
if (lq != null) {
body.locationQuorum = lq;
} else {
body.locationQuorum = 1;
}
if (getHost().getLocation() != null) {
logInfo("Setting node %s location to %s, location quorum is %d",
body.id, getHost().getLocation(), body.locationQuorum);
if (body.customProperties == null) {
body.customProperties = new HashMap<>();
}
body.customProperties.put(NodeState.PROPERTY_NAME_LOCATION,
getHost().getLocation());
}
body.groupReference = getPublicUri();
body.documentSelfLink = UriUtils.buildUriPath(getSelfLink(), body.id);
body.documentKind = Utils.buildKind(NodeState.class);
body.documentUpdateTimeMicros = Utils.getNowMicrosUtc();
return body;
}
@Override
public void handleMaintenance(Operation maint) {
NodeGroupState localState = this.cachedState;
if (localState == null || localState.nodes == null) {
maint.complete();
return;
}
if (localState.nodes.size() <= 1) {
if (!isAvailable()) {
// self patch at least once, so we update availability
sendRequest(Operation.createPatch(getUri())
.setBodyNoCloning(localState)
.setCompletion((o, e) -> {
maint.complete();
}));
} else {
maint.complete();
}
return;
}
if (getHost().isStopping()) {
maint.complete();
return;
}
// probe a fixed, random selection of our peers, giving them our view of the group and
// getting back theirs
// probe log 10 of peers (exclude self)
int peersToProbe = (int) Math.log10(localState.nodes.size() - 1);
// probe at least N peers
peersToProbe = Math.max(peersToProbe, MIN_PEER_GOSSIP_COUNT);
// probe at most total number of peers
peersToProbe = Math.min(localState.nodes.size() - 1, peersToProbe);
AtomicInteger remaining = new AtomicInteger(peersToProbe);
List randomizedPeers = shuffleGroupMembers(localState);
NodeState localNode = localState.nodes.get(getHost().getId());
localNode.documentUpdateTimeMicros = Utils.getNowMicrosUtc();
localNode.groupReference = UriUtils.buildPublicUri(getHost(), getSelfLink());
localState.documentOwner = getHost().getId();
NodeGroupState patchBody = new NodeGroupState();
patchBody.documentOwner = getHost().getId();
patchBody.documentUpdateTimeMicros = localNode.documentUpdateTimeMicros;
int probeCount = 0;
for (NodeState peer : randomizedPeers) {
if (peer == null) {
continue;
}
if (peer.id.equals(getHost().getId())) {
continue;
}
NodeState remotePeer = peer;
URI peerUri = peer.groupReference;
// send a gossip PATCH to the peer, with our state
// perform a health check to N randomly selected peers
// 1) We issue a PATCH to a peer, with the body set to our view of the group
// 2a) if the peer is healthy, they will merge our state with theirs and return
// the merged state in the response. We will then update our state and mark the
// peer AVAILABLE. We just update peer node, we don't currently merge their state
// 2b) if the PATCH failed, we mark the PEER it UNAVAILABLE
long sendTimeMicros = Utils.getSystemNowMicrosUtc();
CompletionHandler ch = (o, e) -> handleGossipPatchCompletion(sendTimeMicros, maint, o,
e, localState,
patchBody,
remaining, remotePeer);
Operation patch = Operation
.createPatch(peerUri)
.setRetryCount(0)
.setConnectionTag(ServiceClient.CONNECTION_TAG_GOSSIP)
.setExpiration(
Utils.fromNowMicrosUtc(
localState.config.peerRequestTimeoutMicros))
.forceRemote()
.setCompletion(ch);
if (peer.groupReference.equals(localNode.groupReference)) {
// If we just detected this is a peer node that used to listen on our address,
// but its obviously no longer around, mark it as REPLACED and do not send PATCH
if (peer.status != NodeStatus.REPLACED) {
peer.status = NodeStatus.REPLACED;
peer.documentUpdateTimeMicros = Utils.getNowMicrosUtc();
peer.documentVersion++;
}
ch.handle(null, null);
} else {
patch.setBodyNoCloning(localState)
.setContentType(Operation.MEDIA_TYPE_APPLICATION_KRYO_OCTET_STREAM)
.sendWith(this);
}
// only probe N peers
if (++probeCount >= peersToProbe) {
break;
}
}
if (probeCount == 0) {
maint.complete();
}
}
public void handleGossipPatchCompletion(long sendTimeMicros, Operation maint, Operation patch,
Throwable e,
NodeGroupState localState, NodeGroupState patchBody, AtomicInteger remaining,
NodeState remotePeer) {
try {
if (patch == null) {
return;
}
updateGossipPatchStat(sendTimeMicros, remotePeer);
long updateTime = localState.membershipUpdateTimeMicros;
if (e != null) {
updateTime = remotePeer.status != NodeStatus.UNAVAILABLE ? Utils.getNowMicrosUtc()
: updateTime;
if (remotePeer.status != NodeStatus.UNAVAILABLE) {
logWarning("Sending patch to peer %s failed with %s; marking as UNAVAILABLE",
remotePeer.id, Utils.toString(e));
remotePeer.documentUpdateTimeMicros = Utils.getNowMicrosUtc();
remotePeer.documentVersion++;
}
remotePeer.status = NodeStatus.UNAVAILABLE;
} else {
NodeGroupState peerState = getStateFromBody(patch);
if (peerState.documentOwner.equals(remotePeer.id)) {
NodeState remotePeerStateFromRsp = peerState.nodes.get(remotePeer.id);
if (remotePeerStateFromRsp.documentVersion > remotePeer.documentVersion) {
remotePeer = remotePeerStateFromRsp;
}
} else if (remotePeer.status != NodeStatus.REPLACED) {
logWarning("Peer address %s has changed to id %s from %s",
patch.getUri(),
peerState.documentOwner,
remotePeer.id);
remotePeer.status = NodeStatus.REPLACED;
remotePeer.documentVersion++;
}
updateTime = Math.max(updateTime, peerState.membershipUpdateTimeMicros);
}
synchronized (patchBody) {
patchBody.nodes.put(remotePeer.id, remotePeer);
patchBody.membershipUpdateTimeMicros = Math.max(updateTime,
patchBody.membershipUpdateTimeMicros);
}
} finally {
int r = remaining.decrementAndGet();
if (r <= 0) {
// to merge updated state, issue a self PATCH. It contains NodeState entries for every
// peer node we just talked to
sendRequest(Operation.createPatch(getUri())
.setBodyNoCloning(patchBody));
maint.complete();
}
}
}
private void updateGossipPatchStat(long sendTimeMicros, NodeState remotePeer) {
long patchCompletionTime = Utils.getSystemNowMicrosUtc();
String statName = remotePeer.id + STAT_NAME_PREFIX_GOSSIP_PATCH_DURATION;
ServiceStat st = ServiceStatUtils.getOrCreateHourlyTimeSeriesHistogramStat(this, statName, EnumSet.of(AggregationType.AVG));
setStat(st, patchCompletionTime - sendTimeMicros);
}
/**
* Merges current node group state with state that came through a PATCH.
*
* PATCH requests are sent from
*
* 1) local service to itself, after it has communicated with a peer, during maintenance.
*
* 2) A remote peer when its probing this local service, during its maintenance cycle
*
* The key invariants that should not be violated, guaranteeing forward evolution of state even
* if nodes only talk to a small portion of their peers:
*
* - When a status changes, the change is accepted if the remote version is higher
*
* - A local node is the only node that can change its own node entry status, for a PATCH that it
* receives.
*
* - A node should never increment the version of a node entry, for other nodes, unless that node
* entry is marked UNAVAILABLE
*
* - When a status changes during gossip version must be incremented - Versions always move forward
*/
private void mergeRemoteAndLocalMembership(
NodeGroupState localState,
NodeGroupState remotePeerState,
EnumSet changes) {
if (localState == null) {
return;
}
boolean isSelfPatch = remotePeerState.documentOwner.equals(getHost().getId());
long now = Utils.getNowMicrosUtc();
NodeState selfEntry = localState.nodes.get(getHost().getId());
for (NodeState remoteEntry : remotePeerState.nodes.values()) {
NodeState currentEntry = localState.nodes.get(remoteEntry.id);
boolean isLocalNode = remoteEntry.id.equals(getHost().getId());
if (!isSelfPatch && isLocalNode) {
if (remoteEntry.status != currentEntry.status) {
logWarning("Peer %s is reporting us as %s, current status: %s",
remotePeerState.documentOwner, remoteEntry.status, currentEntry.status);
if (remoteEntry.documentVersion > currentEntry.documentVersion) {
// increment local version to re-enforce we are alive and well
currentEntry.documentVersion = remoteEntry.documentVersion;
currentEntry.documentUpdateTimeMicros = now;
changes.add(NodeGroupChange.SELF_CHANGE);
}
}
// local instance of node group service is the only one that can update its own
// status
continue;
}
if (currentEntry == null) {
boolean hasExpired = remoteEntry.documentExpirationTimeMicros > 0
&& remoteEntry.documentExpirationTimeMicros < now;
if (hasExpired || NodeState.isUnAvailable(remoteEntry, null)) {
continue;
}
if (selfEntry.groupReference.equals(remoteEntry.groupReference)) {
logWarning("Local address %s has changed to id %s from %s", remoteEntry.groupReference, getHost().getId(), remoteEntry.id);
changes.add(NodeGroupChange.SELF_CHANGE);
continue;
}
if (!isLocalNode) {
logInfo("Adding new peer %s (%s), status %s", remoteEntry.id,
remoteEntry.groupReference, remoteEntry.status);
}
// we found a new peer, through the gossip PATCH. Add to our state
localState.nodes.put(remoteEntry.id, remoteEntry);
changes.add(NodeGroupChange.PEER_ADDED);
continue;
}
boolean needsUpdate = currentEntry.status != remoteEntry.status
|| currentEntry.membershipQuorum != remoteEntry.membershipQuorum;
if (needsUpdate) {
changes.add(NodeGroupChange.PEER_STATUS_CHANGE);
}
if (isSelfPatch && isLocalNode && needsUpdate) {
// we sent a self PATCH to update our status. Move our version forward;
currentEntry.documentVersion = Math.max(remoteEntry.documentVersion,
currentEntry.documentVersion) + 1;
currentEntry.documentUpdateTimeMicros = Math.max(
remoteEntry.documentUpdateTimeMicros,
now);
currentEntry.status = remoteEntry.status;
currentEntry.options = remoteEntry.options;
continue;
}
// versions move forward only, ignore stale nodes
if (remoteEntry.documentVersion < currentEntry.documentVersion) {
logInfo("v:%d - q:%d, v:%d - q:%d , %s - %s (local:%s %d)",
currentEntry.documentVersion,
currentEntry.membershipQuorum,
remoteEntry.documentVersion, remoteEntry.membershipQuorum,
currentEntry.id,
remotePeerState.documentOwner,
getHost().getId(),
selfEntry.documentVersion);
continue;
}
if (remoteEntry.documentVersion == currentEntry.documentVersion && needsUpdate) {
// pick update with most recent time, even if that is prone to drift and jitter
// between nodes
if (remoteEntry.documentUpdateTimeMicros < currentEntry.documentUpdateTimeMicros) {
logWarning(
"Ignoring update for %s from peer %s. Local status: %s, remote status: %s",
remoteEntry.id, remotePeerState.documentOwner, currentEntry.status,
remoteEntry.status);
continue;
}
}
if (remoteEntry.status == NodeStatus.UNAVAILABLE
&& currentEntry.documentExpirationTimeMicros == 0
&& remoteEntry.documentExpirationTimeMicros == 0) {
remoteEntry.documentExpirationTimeMicros = Utils.fromNowMicrosUtc(
localState.config.nodeRemovalDelayMicros);
logInfo("Set expiration at %d for unavailable node %s(%s)",
remoteEntry.documentExpirationTimeMicros,
remoteEntry.id,
remoteEntry.groupReference);
changes.add(NodeGroupChange.PEER_STATUS_CHANGE);
needsUpdate = true;
}
if (remoteEntry.status == NodeStatus.UNAVAILABLE
&& currentEntry.status == NodeStatus.UNAVAILABLE
&& needsUpdate) {
// nodes increment their own entry version, except, if they are unavailable
remoteEntry.documentVersion++;
}
localState.nodes.put(remoteEntry.id, remoteEntry);
}
List missingNodes = new ArrayList<>();
for (NodeState l : localState.nodes.values()) {
NodeState r = remotePeerState.nodes.get(l.id);
if (!NodeState.isUnAvailable(l, null) || l.id.equals(getHost().getId())) {
continue;
}
long expirationMicros = l.documentExpirationTimeMicros;
if (r != null) {
expirationMicros = Math.max(l.documentExpirationTimeMicros,
r.documentExpirationTimeMicros);
}
if (expirationMicros > 0 && now > expirationMicros) {
changes.add(NodeGroupChange.PEER_STATUS_CHANGE);
logInfo("Removing expired unavailable node %s(%s)", l.id, l.groupReference);
missingNodes.add(l.id);
}
}
for (String id : missingNodes) {
localState.nodes.remove(id);
}
boolean isModified = !changes.isEmpty();
localState.membershipUpdateTimeMicros = Math.max(
remotePeerState.membershipUpdateTimeMicros,
isModified ? now : localState.membershipUpdateTimeMicros);
if (isModified) {
logInfo("State updated, merge with %s, self %s, %d",
remotePeerState.documentOwner,
localState.documentOwner,
localState.membershipUpdateTimeMicros);
localState.localMembershipUpdateTimeMicros = now;
}
}
private List shuffleGroupMembers(NodeGroupState localState) {
List peers = new ArrayList<>(localState.nodes.values());
Collections.shuffle(peers, ThreadLocalRandom.current());
return peers;
}
private NodeGroupState getStateFromBody(Operation o) {
if (!o.hasBody()) {
return new NodeGroupState();
}
NodeGroupState rsp = o.getBody(NodeGroupState.class);
if (rsp != null) {
if (rsp.nodes == null) {
rsp.nodes = new HashMap<>();
}
}
return rsp;
}
}
