org.jgroups.protocols.FD Maven / Gradle / Ivy
package org.jgroups.protocols;
import org.jgroups.*;
import org.jgroups.annotations.*;
import org.jgroups.stack.Protocol;
import org.jgroups.util.BoundedList;
import org.jgroups.util.MessageBatch;
import org.jgroups.util.TimeScheduler;
import org.jgroups.util.Util;
import java.io.DataInput;
import java.io.DataOutput;
import java.util.*;
import java.util.concurrent.Future;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.locks.Lock;
import java.util.concurrent.locks.ReentrantLock;
/**
* Failure detection based on simple heartbeat protocol. Regularly polls members
* for liveness. Multicasts SUSPECT messages when a member is not reachable. The
* simple algorithms works as follows: the membership is known and ordered. Each
* HB protocol periodically sends an 'are-you-alive' message to its *neighbor*.
* A neighbor is the next in rank in the membership list, which is recomputed
* upon a view change. When a response hasn't been received for n milliseconds
* and m tries, the corresponding member is suspected (and eventually excluded
* if faulty).
*
* FD starts when it detects (in a view change notification) that there are at
* least 2 members in the group. It stops running when the membership drops
* below 2.
*
* When a message is received from the monitored neighbor member, it causes the
* pinger thread to 'skip' sending the next are-you-alive message. Thus, traffic
* is reduced.
*
* @author Bela Ban
*/
@MBean(description="Failure detection based on simple heartbeat protocol")
public class FD extends Protocol {
/* ----------------------------------------- Properties -------------------------------------------------- */
@Property(description="Timeout to suspect a node P if neither a heartbeat nor data were received from P.")
protected long timeout=3000;
@Property(description="Number of times to send an are-you-alive message")
protected int max_tries=5;
@Property(description="Treat messages received from members as heartbeats. Note that this means we're updating " +
"a value in a hashmap every time a message is passing up the stack through FD, which is costly.")
boolean msg_counts_as_heartbeat=true;
/* --------------------------------------------- JMX ------------------------------------------------------ */
protected int num_heartbeats;
protected int num_suspect_events;
protected final BoundedList suspect_history=new BoundedList<>(20);
/* --------------------------------------------- Fields ------------------------------------------------------ */
protected Address local_addr;
protected volatile long last_ack=System.nanoTime();
protected final AtomicInteger num_tries=new AtomicInteger(0);
protected final Lock lock=new ReentrantLock();
@GuardedBy("lock")
protected volatile Address ping_dest;
@GuardedBy("lock")
protected final List members=new ArrayList<>();
/** Members from which we select ping_dest. Copy of {@link #members} minus the suspected members */
@GuardedBy("lock")
protected final List pingable_mbrs=new ArrayList<>();
protected TimeScheduler timer;
// task that performs the actual monitoring for failure detection
@GuardedBy("lock")
protected Future monitor_future=null;
/** Transmits SUSPECT message until view change or UNSUSPECT is received */
protected final Broadcaster bcast_task=new Broadcaster();
@ManagedAttribute(description="Member address")
public String getLocalAddress() {return local_addr != null? local_addr.toString() : "null";}
@ManagedAttribute(description="List of cluster members")
public String getMembers() {return members != null? members.toString() : "null";}
@ManagedAttribute(description="List of pingable members of a cluster")
public String getPingableMembers() {return pingable_mbrs != null? pingable_mbrs.toString() : "null";}
@ManagedAttribute(description="Ping destination")
public String getPingDest() {return ping_dest != null? ping_dest.toString() : "null";}
@ManagedAttribute(description="Number of heartbeats sent")
public int getNumberOfHeartbeatsSent() {return num_heartbeats;}
@ManagedAttribute(description="Number of suspect events received")
public int getNumSuspectEventsGenerated() {return num_suspect_events;}
public long getTimeout() {return timeout;}
public void setTimeout(long timeout) {this.timeout=timeout;}
public int getMaxTries() {return max_tries;}
public void setMaxTries(int max_tries) {this.max_tries=max_tries;}
public int getCurrentNumTries() {return num_tries.get();}
@ManagedOperation(description="Print suspect history")
public String printSuspectHistory() {
StringBuilder sb=new StringBuilder();
for(String addr: suspect_history)
sb.append(addr).append("\n");
return sb.toString();
}
public void resetStats() {
num_heartbeats=num_suspect_events=0;
suspect_history.clear();
}
public void init() throws Exception {
timer=getTransport().getTimer();
if(timer == null)
throw new Exception("timer cannot be retrieved");
}
public void stop() {
lock.lock();
try {
ping_dest=null;
stopMonitor();
}
finally {
lock.unlock();
}
}
protected Address getPingDest(List mbrs) {
Address tmp, retval=null;
if(mbrs == null || mbrs.size() < 2 || local_addr == null)
return null;
for(int i=0; i < mbrs.size(); i++) {
tmp=mbrs.get(i);
if(local_addr.equals(tmp)) {
if(i + 1 >= mbrs.size())
retval=mbrs.get(0);
else
retval=mbrs.get(i + 1);
break;
}
}
return retval;
}
@ManagedOperation(description="Stops checking for crashed members")
public void stopFailureDetection() {
stopMonitor();
}
@ManagedOperation(description="Resumes checking for crashed members")
public void startFailureDetection() {
startMonitor();
}
/** Requires lock to held by caller */
@GuardedBy("lock")
protected void startMonitor() {
if(monitor_future == null || monitor_future.isDone()) {
last_ack=System.nanoTime(); // start from scratch
monitor_future=timer.scheduleWithFixedDelay(new Monitor(), timeout, timeout, TimeUnit.MILLISECONDS);
num_tries.set(1);
}
}
/** Requires lock to be held by caller */
@GuardedBy("lock")
protected void stopMonitor() {
if(monitor_future != null) {
monitor_future.cancel(true);
monitor_future=null;
}
}
@ManagedAttribute(description="Whether the failure detection monitor is running",writable=false)
public boolean isMonitorRunning() {return monitor_future != null && !monitor_future.isDone();}
public Object up(Event evt) {
switch(evt.getType()) {
case Event.MSG:
Message msg=(Message)evt.getArg();
FdHeader hdr=(FdHeader)msg.getHeader(this.id);
if(hdr == null) {
if(msg_counts_as_heartbeat)
updateTimestamp(msg.getSrc());
break; // message did not originate from FD layer, just pass up
}
switch(hdr.type) {
case FdHeader.HEARTBEAT: // heartbeat request; send heartbeat ack
Address hb_sender=msg.getSrc();
log.trace("%s: received are-you-alive from %s, sending response", local_addr, hb_sender);
sendHeartbeatResponse(hb_sender);
break; // don't pass up !
case FdHeader.HEARTBEAT_ACK: // heartbeat ack
updateTimestamp(hdr.from);
break;
case FdHeader.SUSPECT:
if(hdr.mbrs == null)
return null;
log.trace("%s: received suspect message: %s", local_addr, hdr);
for(Address mbr: hdr.mbrs) {
if(local_addr != null && mbr.equals(local_addr)) {
log.warn("%s: I was suspected by %s; ignoring the SUSPECT message and sending back a HEARTBEAT_ACK",
local_addr, msg.src());
sendHeartbeatResponse(msg.getSrc());
continue;
}
lock.lock();
try {
computePingDest(mbr);
}
finally {
lock.unlock();
}
up_prot.up(new Event(Event.SUSPECT, mbr));
down_prot.down(new Event(Event.SUSPECT, mbr));
}
break;
case FdHeader.UNSUSPECT:
if(hdr.mbrs == null)
return null;
log.trace("%s: received unsuspect message: %s", local_addr, hdr);
for(Address tmp: hdr.mbrs)
unsuspect(tmp);
break;
}
return null;
}
return up_prot.up(evt); // pass up to the layer above us
}
public void up(MessageBatch batch) {
Collection msgs=batch.getMatchingMessages(id, true);
boolean updated=false;
if(msgs != null) {
for(Message msg: msgs) {
FdHeader hdr=(FdHeader)msg.getHeader(id); // header is not null at this point
if(hdr.type == FdHeader.HEARTBEAT_ACK)
updated=true;
else
up(new Event(Event.MSG, msg)); // SUSPECT and HEARTBEAT
}
}
if(updated || (msg_counts_as_heartbeat && batch.sender() != null))
updateTimestamp(batch.sender());
if(!batch.isEmpty())
up_prot.up(batch);
}
public Object down(Event evt) {
switch(evt.getType()) {
case Event.VIEW_CHANGE:
Object retval=down_prot.down(evt);
View view=(View)evt.getArg();
lock.lock();
try {
members.clear();
members.addAll(view.getMembers());
bcast_task.adjustSuspectedMembers(members);
computePingDest(null);
if(view.size() <= 1)
stopMonitor();
else if(!isMonitorRunning())
startMonitor();
}
finally {
lock.unlock();
}
return retval;
case Event.UNSUSPECT:
FdHeader hdr=new FdHeader(FdHeader.UNSUSPECT);
hdr.mbrs=new ArrayList<>();
hdr.mbrs.add((Address)evt.getArg());
hdr.from=local_addr;
Message unsuspect_msg=new Message().setFlag(Message.Flag.INTERNAL).putHeader(id, hdr);
log.trace("%s: broadcasting UNSUSPECT message (mbrs=%s)", local_addr, hdr.mbrs);
down_prot.down(new Event(Event.MSG, unsuspect_msg));
break;
case Event.SET_LOCAL_ADDRESS:
local_addr=(Address)evt.getArg();
break;
}
return down_prot.down(evt);
}
protected void sendHeartbeatResponse(Address dest) {
Message hb_ack=new Message(dest).setFlag(Message.Flag.INTERNAL);
FdHeader tmp_hdr=new FdHeader(FdHeader.HEARTBEAT_ACK);
tmp_hdr.from=local_addr;
hb_ack.putHeader(this.id, tmp_hdr);
down_prot.down(new Event(Event.MSG, hb_ack));
}
@GuardedBy("lock")
protected void unsuspect(Address mbr) {
lock.lock();
try {
bcast_task.removeSuspectedMember(mbr);
computePingDest(null);
}
finally {
lock.unlock();
}
}
protected void updateTimestamp(Address sender) {
if(sender != null && sender.equals(ping_dest)) {
last_ack=System.nanoTime();
num_tries.set(1);
}
}
/**
* Computes pingable_mbrs (based on the current membership and the suspected members) and ping_dest
* @param remove The member to be removed from pingable_mbrs
*/
@GuardedBy("lock")
protected void computePingDest(Address remove) {
if(remove != null)
pingable_mbrs.remove(remove);
else {
pingable_mbrs.clear();
pingable_mbrs.addAll(members);
pingable_mbrs.removeAll(bcast_task.getSuspectedMembers());
}
Address old_ping_dest=ping_dest;
ping_dest=getPingDest(pingable_mbrs);
if(Util.different(old_ping_dest, ping_dest)) {
num_tries.set(1);
last_ack=System.nanoTime();
}
}
public static class FdHeader extends Header {
public static final byte HEARTBEAT = 0;
public static final byte HEARTBEAT_ACK = 1;
public static final byte SUSPECT = 2;
public static final byte UNSUSPECT = 3;
protected byte type=HEARTBEAT;
protected Collection mbrs;
protected Address from; // member who detected that suspected_mbr has failed
public FdHeader() {
}
public FdHeader(byte type) {
this.type=type;
}
public FdHeader(byte type, Collection mbrs, Address from) {
this(type);
this.mbrs=mbrs;
this.from=from;
}
public String toString() {
switch(type) {
case HEARTBEAT:
return "heartbeat";
case HEARTBEAT_ACK:
return "heartbeat ack";
case SUSPECT:
return "SUSPECT (suspected_mbrs=" + mbrs + "), from=" + from;
case UNSUSPECT:
return "UNSUSPECT (mbrs=" + mbrs + "), from=" + from;
default:
return "unknown type (" + type + ")";
}
}
public int size() {
int retval=Global.BYTE_SIZE; // type
retval+=Util.size(mbrs);
retval+=Util.size(from);
return retval;
}
public void writeTo(DataOutput out) throws Exception {
out.writeByte(type);
Util.writeAddresses(mbrs, out);
Util.writeAddress(from, out);
}
public void readFrom(DataInput in) throws Exception {
type=in.readByte();
mbrs=(Collection)Util.readAddresses(in, ArrayList.class);
from=Util.readAddress(in);
}
}
/** Task which periodically checks of the last_ack from ping_dest exceeded timeout and - if yes - broadcasts
* a SUSPECT message */
protected class Monitor implements Runnable {
public void run() {
Address dest=ping_dest;
if(dest == null) {
log.trace("%s: ping_dest is null, skipping timeout check: members=%s, pingable_mbrs=%s",
local_addr, members, pingable_mbrs);
return;
}
// 1. send heartbeat request
Message hb_req=new Message(dest).setFlag(Message.Flag.INTERNAL).putHeader(id, new FdHeader(FdHeader.HEARTBEAT));
log.trace("%s: sending are-you-alive msg to %s", local_addr, dest);
down_prot.down(new Event(Event.MSG, hb_req));
num_heartbeats++;
// 2. If the time of the last heartbeat is > timeout and max_tries heartbeat messages have not been
// received, then broadcast a SUSPECT message. Will be handled by coordinator, which may install
// a new view
// time in msecs we haven't heard from ping_dest
long not_heard_from=TimeUnit.MILLISECONDS.convert(System.nanoTime() - last_ack, TimeUnit.NANOSECONDS);
// quick & dirty fix: increase timeout by 500 ms to allow for latency (bela June 27 2003)
if(not_heard_from > timeout + 500) { // no heartbeat ack for more than timeout msecs
int tmp_tries=num_tries.get();
if(tmp_tries >= max_tries) {
if(!dest.equals(ping_dest)) // ping_dest was changed meanwhile...
return;
log.debug("%s: received no heartbeat from %s for %d times (%d milliseconds), suspecting it",
local_addr, dest, tmp_tries, tmp_tries * timeout);
// broadcast a SUSPECT message to all members - loop until unsuspect or view change is received
bcast_task.addSuspectedMember(dest);
num_tries.set(1);
if(stats) {
num_suspect_events++;
suspect_history.add(String.format("%s: %s", new Date(), dest));
}
}
else {
log.debug("%s: heartbeat missing from %s (number=%d)", local_addr, dest, tmp_tries);
num_tries.incrementAndGet();
}
}
}
public String toString() {
return FD.class.getSimpleName() + ": Monitor (timeout=" + timeout + "ms)";
}
}
/**
* Task that periodically broadcasts a list of suspected members to the group. Goal is not to lose
* a SUSPECT message: since these are bcast unreliably, they might get dropped. The BroadcastTask makes
* sure they are retransmitted until a view has been received which doesn't contain the suspected members
* any longer. Then the task terminates.
*/
protected final class Broadcaster {
protected final List suspected_mbrs=new ArrayList<>(7);
protected final Lock bcast_lock=new ReentrantLock();
@GuardedBy("bcast_lock")
protected Future bcast_future=null;
@GuardedBy("bcast_lock")
protected BroadcastTask task;
protected List getSuspectedMembers() {
return suspected_mbrs;
}
/**
* Starts a new task, or - if already running - adds the argument to the running task.
* @param suspect
*/
protected void startBroadcastTask(Address suspect) {
bcast_lock.lock();
try {
if(bcast_future == null || bcast_future.isDone()) {
task=new BroadcastTask(suspected_mbrs);
task.addSuspectedMember(suspect);
bcast_future=timer.scheduleWithFixedDelay(task,
0, // run immediately the first time
timeout, // then every timeout milliseconds, until cancelled
TimeUnit.MILLISECONDS);
}
else {
task.addSuspectedMember(suspect);
}
}
finally {
bcast_lock.unlock();
}
}
protected void stopBroadcastTask() {
bcast_lock.lock();
try {
if(bcast_future != null) {
bcast_future.cancel(true);
bcast_future=null;
task=null;
}
}
finally {
bcast_lock.unlock();
}
}
/** Adds a suspected member. Starts the task if not yet running */
protected void addSuspectedMember(Address mbr) {
if(mbr == null) return;
if(!members.contains(mbr)) return;
synchronized(suspected_mbrs) {
if(!suspected_mbrs.contains(mbr)) {
suspected_mbrs.add(mbr);
startBroadcastTask(mbr);
}
}
}
void removeSuspectedMember(Address suspected_mbr) {
if(suspected_mbr == null) return;
synchronized(suspected_mbrs) {
suspected_mbrs.remove(suspected_mbr);
if(suspected_mbrs.isEmpty())
stopBroadcastTask();
}
}
/** Removes all elements from suspected_mbrs that are not in the new membership */
void adjustSuspectedMembers(List new_mbrship) {
if(new_mbrship == null || new_mbrship.isEmpty()) return;
synchronized(suspected_mbrs) {
suspected_mbrs.retainAll(new_mbrship);
if(suspected_mbrs.isEmpty())
stopBroadcastTask();
}
}
}
protected final class BroadcastTask implements Runnable {
protected final List suspected_members=new ArrayList<>();
BroadcastTask(List suspected_members) {
this.suspected_members.addAll(suspected_members);
}
public void stop() {
suspected_members.clear();
}
public void run() {
FD.FdHeader hdr;
synchronized(suspected_members) {
if(suspected_members.isEmpty()) {
stop();
return;
}
hdr=new FdHeader(FdHeader.SUSPECT);
hdr.mbrs=new ArrayList<>(suspected_members);
hdr.from=local_addr;
}
Message suspect_msg=new Message().setFlag(Message.Flag.INTERNAL).putHeader(id, hdr);
log.trace("%s: broadcasting SUSPECT message (suspects=%s)", local_addr, suspected_members);
down_prot.down(new Event(Event.MSG, suspect_msg));
}
public void addSuspectedMember(Address suspect) {
if(suspect != null && !suspected_members.contains(suspect)) {
suspected_members.add(suspect);
}
}
public String toString() {
return "BroadcastTask (" + suspected_members.size() + " suspected mbrs)";
}
}
}