com.fasterxml.clustermate.service.remote.RemoteClusterStateFetcher Maven / Gradle / Ivy
Show all versions of clustermate-service Show documentation
package com.fasterxml.clustermate.service.remote;
import java.io.IOException;
import java.util.*;
import java.util.concurrent.atomic.AtomicBoolean;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import com.fasterxml.storemate.shared.IpAndPort;
import com.fasterxml.storemate.shared.TimeMaster;
import com.fasterxml.clustermate.api.ClusterStatusAccessor;
import com.fasterxml.clustermate.api.KeyRange;
import com.fasterxml.clustermate.api.NodeState;
import com.fasterxml.clustermate.api.msg.ClusterStatusMessage;
import com.fasterxml.clustermate.json.ClusterMessageConverter;
import com.fasterxml.clustermate.service.*;
import com.fasterxml.clustermate.std.JdkClusterStatusAccessor;
/**
* Helper class used for accessing status of remote cluster, for purpose
* of syncing from one of peers of this instance.
*
* NOTE: lots of duplication with StoreClientBootstrapper, unfortunately.
*/
public class RemoteClusterStateFetcher
{
/**
* Not sure where this should be located; probably not here.
* But since the other definition is in VagabondClientConfigBuilder,
* which is inaccessible from here, this'll have to do for now.
*/
protected final static String[] DEFAULT_BASE_PATH = new String[] { "ts" };
/**
* Let's keep initial timeouts relatively low, since we can usually
* try to go through multiple server nodes to get response quickly.
*/
public final static long BOOTSTRAP_TIMEOUT_MSECS = 2000L;
/**
* Let's limit validity of Remote cluster view info to 15 minutes, so that
* information is reloaded every now and then.
*/
private final static long MSECS_FOR_REMOTE_RELOAD = 15 * 60 * 1000L;
private final Logger LOG = LoggerFactory.getLogger(getClass());
/*
///////////////////////////////////////////////////////////////////////
// Configuration
///////////////////////////////////////////////////////////////////////
*/
protected final SharedServiceStuff _config;
/**
* Set of server nodes used for bootstrapping; we need at least
* one to be able to locate others.
*/
protected final Set _initialEndpoints;
/**
* Object used to access Node State information, needed to construct
* view of the cluster.
*/
protected ClusterStatusAccessor _accessor;
/**
* We need bit of local state for negotiating with remote peer.
*/
protected final NodeState _localNode;
protected final AtomicBoolean _running;
/*
///////////////////////////////////////////////////////////////////////
// Construction
///////////////////////////////////////////////////////////////////////
*/
public RemoteClusterStateFetcher(SharedServiceStuff stuff,
AtomicBoolean running,
Set endpoints, NodeState local)
{
_config = stuff;
_running = running;
_initialEndpoints = endpoints;
_localNode = local;
_accessor = new JdkClusterStatusAccessor(new ClusterMessageConverter(stuff.jsonMapper()),
DEFAULT_BASE_PATH, stuff.getPathStrategy());
}
/**
* @return Whether at least one of endpoints was resolvable or not
*/
protected boolean init()
{
LOG.info("Try to resolve {} remote IP endpoints", _initialEndpoints.size());
Iterator it = _initialEndpoints.iterator();
while (it.hasNext()) {
if (!_running.get()) {
LOG.warn("Terminating remote IP endpoint resolution");
return false;
}
IpAndPort ip = it.next();
try {
ip.getIP();
} catch (Exception e) {
LOG.error("Failed to resolve end point '"+ip.toString()+"', removing. Problem: "+e.getMessage());
it.remove();
}
}
LOG.info("Completed remote IP endpoint resolution");
return true;
}
/*
///////////////////////////////////////////////////////////////////////
// Client bootstrapping
///////////////////////////////////////////////////////////////////////
*/
/**
* Method called to fetch state information from the remote cluster so that
* it is possible to determine optimal peer matching, and start making
* sync-list/pull calls.
*
* @return Fully initialized remote cluster info object, if successful; or
* null to indicate that fetch failed
*/
public RemoteCluster fetch(int maxWaitSecs) throws IOException
{
final long startTime = _config.currentTimeMillis();
final long waitUntil = (maxWaitSecs <= 0) ? Long.MAX_VALUE : (startTime + 1000 * maxWaitSecs);
// Since we'll be removing entries, need to make a local copy:
ArrayList ips = new ArrayList(_initialEndpoints);
BootstrapState bs = new BootstrapState();
long roundStart;
int round = 0;
while (!ips.isEmpty() && (roundStart = System.currentTimeMillis()) < waitUntil) {
Iterator it = ips.iterator();
while (it.hasNext()) {
IpAndPort ip = it.next();
final long requestTime = System.currentTimeMillis();
long maxTimeout = waitUntil - requestTime;
try {
ClusterStatusMessage resp = _accessor.getRemoteStatus(ip,
Math.min(maxTimeout, BOOTSTRAP_TIMEOUT_MSECS));
if (resp == null) {
continue;
}
it.remove(); // remove from bootstrap list
NodeState peer = resp.local;
bs.updateDirectState(ip, peer,
requestTime, System.currentTimeMillis(), resp.creationTime);
// Important: include LOCAL peers of REMOTE that we called; must NOT
// Cross the Lines
for (NodeState stateSec : resp.localPeers) {
bs.updateIndirectState(ip, stateSec);
}
// and we don't really care about remote peers here
} catch (IOException e) {
// only warn during first round
if (round == 0) {
LOG.warn("Failed to access Remote Status on {} ({}: {}), will retry",
ip, e.getClass().getName(), e.getMessage());
}
} catch (RuntimeException e) { // usually more severe ones, NPE etc
LOG.error("Internal error with cluster state call (IP "+ip+"): "
+"("+e.getClass().getName()+") "+e.getMessage(), e);
} catch (Exception e) {
LOG.error("Initial cluster state call (IP "+ip+") failed: ("+e.getClass().getName()+") "
+e.getMessage(), e);
}
}
++round;
// If we failed first time around, let's wait a bit...
long timeTaken = System.currentTimeMillis() - roundStart;
if (timeTaken < 1000L) { // if we had string of failures, wait a bit
try {
Thread.sleep(1000L - timeTaken);
} catch (InterruptedException e) {
throw new IOException(e.getMessage(), e);
}
}
}
return bs.finish(_config.getTimeMaster(), _localNode);
}
/**
* Helper class used for aggregating state of remote cluster during boostrapping.
*/
static class BootstrapState
{
protected final Map _nodes = new HashMap();
/*
///////////////////////////////////////////////////////////////////////
// Life-cycle
///////////////////////////////////////////////////////////////////////
*/
public BootstrapState() { }
public RemoteCluster finish(TimeMaster timeMaster, NodeState localState)
{
List overlapping = new ArrayList();
final KeyRange localRange = localState.totalRange();
for (RemoteClusterNode node : _nodes.values()) {
if (localRange.overlapsWith(node.getTotalRange())) {
overlapping.add(node);
}
}
if (overlapping.isEmpty()) {
return null;
}
// Also, need to sort in descending order of preference
Collections.sort(overlapping, new Comparator() {
@Override
public int compare(RemoteClusterNode c1, RemoteClusterNode c2) {
KeyRange r1 = c1.getTotalRange();
// First, exact match is the usual way, preferred
if (r1.equals(localRange)) {
return -1;
}
KeyRange r2 = c2.getTotalRange();
if (r2.equals(localRange)) {
return 1;
}
// If not, let's actually base it on stable clockwise-distance between
// starting points, such that we'll try to find range that starts as soon
// as possible _after_ local range
int dist1 = localRange.clockwiseDistance(r1);
int dist2 = localRange.clockwiseDistance(r2);
if (dist1 != dist2) {
return dist1 - dist2;
}
// And if this still doesn't resolve, choose one with _smaller_ range; assumption
// is that most likely cluster is growing, and smaller range indicates newer
// information
int diff = r1.getLength() - r2.getLength();
if (diff != 0) { // sort from smaller to bigger
return diff;
}
// and otherwise use simple lexicographic (~= alphabetic) ordering of endpoint IP
// Goal being a stable ordering
return c1.getAddress().compareTo(c2.getAddress());
}
});
return new RemoteCluster(timeMaster.currentTimeMillis() + MSECS_FOR_REMOTE_RELOAD, localState,
overlapping);
}
/*
///////////////////////////////////////////////////////////////////////
// State updates
///////////////////////////////////////////////////////////////////////
*/
/**
* Method called to add information directly related to node that served
* the request.
*/
public void updateDirectState(IpAndPort byNode, NodeState stateInfo,
long requestTime, long responseTime,
long clusterInfoVersion)
{
RemoteClusterNode localState = _nodes.get(byNode);
if (localState == null) { // new info
localState = new RemoteClusterNode(byNode, stateInfo.getRangeActive(), stateInfo.getRangePassive());
_addNode(byNode, localState);
}
/*boolean needInvalidate =*/ localState.updateRanges(stateInfo.getRangeActive(),
stateInfo.getRangePassive());
localState.setLastRequestSent(requestTime);
localState.setLastResponseReceived(responseTime);
localState.setLastNodeUpdateFetched(stateInfo.getLastUpdated());
}
/**
* Method called to add information obtained indirectly; i.e. "gossip".
*/
public void updateIndirectState(IpAndPort byNode, NodeState stateInfo)
{
// First: ensure references are properly resolved (eliminate "localhost" if need be):
IpAndPort ip = stateInfo.getAddress();
if (ip.isLocalReference()) {
ip = byNode.withPort(ip.getPort());
}
final long nodeInfoTimestamp = stateInfo.getLastUpdated();
// otherwise pretty simple:
RemoteClusterNode state = _nodes.get(ip);
if (state == null) { // new info
state = new RemoteClusterNode(ip, stateInfo.getRangeActive(), stateInfo.getRangePassive());
_addNode(ip, state);
} else {
// quick check to ensure info we get is newer: if not, skip
if (nodeInfoTimestamp <= state.getLastNodeUpdateFetched()) {
return;
}
}
state.setLastNodeUpdateFetched(nodeInfoTimestamp);
/*boolean needInvalidate =*/ state.updateRanges(stateInfo.getRangeActive(),
stateInfo.getRangePassive());
}
private void _addNode(IpAndPort key, RemoteClusterNode state)
{
_nodes.put(key, state);
// and, if had other things to add, they'd get initialized here as well
}
}
}