com.fasterxml.clustermate.service.cluster.ClusterPeerImpl Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of clustermate-service Show documentation
Show all versions of clustermate-service Show documentation
Building blocks for ClusterMate-based services and servers.
package com.fasterxml.clustermate.service.cluster;
import java.io.IOException;
import java.io.InputStream;
import java.util.Iterator;
import java.util.List;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicBoolean;
import java.util.concurrent.atomic.AtomicInteger;
import org.skife.config.TimeSpan;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import com.fasterxml.storemate.shared.*;
import com.fasterxml.storemate.shared.util.IOUtil;
import com.fasterxml.storemate.store.*;
import com.fasterxml.storemate.store.state.NodeStateStore;
import com.fasterxml.storemate.store.util.BoundedInputStream;
import com.fasterxml.clustermate.api.*;
import com.fasterxml.clustermate.service.SharedServiceStuff;
import com.fasterxml.clustermate.service.state.ActiveNodeState;
import com.fasterxml.clustermate.service.store.StoredEntry;
import com.fasterxml.clustermate.service.store.StoredEntryConverter;
import com.fasterxml.clustermate.service.sync.*;
import com.fasterxml.clustermate.service.util.StoreUtil;
public class ClusterPeerImpl>
extends ClusterPeer
implements com.fasterxml.storemate.shared.StartAndStoppable
{
/**
* If access to peer's sync/list fails, wait for this duration before
* trying again.
*/
private final static long SLEEP_FOR_SYNCLIST_ERRORS_MSECS = 10000L;
private final static long SLEEP_FOR_SYNCPULL_ERRORS_MSECS = 3000L;
// no point trying to sleep for trivial time
private final static long MINIMAL_SLEEP_MSECS = 10L;
/**
* If synclist is empty and server does not instruct us, simply sleep for
* 1 second (to try avoid congestion)
*/
private final static long SLEEP_FOR_EMPTY_SYNCLIST_MSECS = 1000L;
// no real hurry; use 20 seconds to account for GC, congestion etc
private final static TimeSpan TIMEOUT_FOR_SYNCLIST = new TimeSpan(10L, TimeUnit.SECONDS);
/**
* Lowish timeout for "bye bye" message, so it won't block shutdown
*/
private final static TimeSpan TIMEOUT_FOR_BYEBYE = new TimeSpan(250L, TimeUnit.MILLISECONDS);
/**
* We will limit maximum estimate response size to some reasonable
* limit: starting with 250 megs. The idea is to use big enough sizes
* for efficient bulk transfer; but small enough not to cause timeouts
* during normal operation.
*/
private final static long MAX_TOTAL_PAYLOAD = 250 * 1000 * 1000;
/**
* During fetching of items to sync, let's cap number of failures to some
* number; this should make it easier to recover from cases where peer
* shuts down during individual sync operation (after sync list received
* but before all entries are fetched)
*/
private final static int MAX_SYNC_FAILURES = 8;
/**
* Also, let's limit maximum individual calls per sync-pull fetch portion,
* to avoid excessive calls.
*/
private final int MAX_FETCH_TRIES = 20;
private final static Logger LOG = LoggerFactory.getLogger(ClusterPeer.class);
/*
/**********************************************************************
/* Configuration, general helpers
/**********************************************************************
*/
protected final SharedServiceStuff _stuff;
/**
* This object is necessary to support "virtual time" for test cases.
*/
protected final TimeMaster _timeMaster;
/*
/**********************************************************************
/* Operation state of the peer object
/**********************************************************************
*/
/**
* Synchronization thread if (and only if) this peer shares part of keyspace
* with the local node; otherwise null.
* Note that threads may be started and stopped based on changes to cluster
* configuration.
*/
protected Thread _syncThread;
/**
* Flag used to request termination of the sync thread.
*/
protected AtomicBoolean _running = new AtomicBoolean(false);
/**
* Let's keep track of number of failures (as per caught exceptions); mostly
* so that tests can verify passing, but also potentially for monitoring.
*/
protected AtomicInteger _failCount = new AtomicInteger(0);
/*
/**********************************************************************
/* Helper objects for entry handling
/**********************************************************************
*/
/**
* And to store fetched missing entities, we need the store
*/
protected final StorableStore _entryStore;
/**
* Need to construct metadata nuggets with this factory
*/
protected final StoredEntryConverter _entryConverter;
/*
/**********************************************************************
/* Helper objects for sync handling
/**********************************************************************
*/
protected final ClusterViewByServerUpdatable _cluster;
/**
* Object used to access Node State information, needed to construct
* view of the cluster.
*/
protected final ClusterStatusAccessor _statusAccessor;
/**
* Helper object used for doing HTTP requests
*/
protected final SyncListAccessor _syncListAccessor;
/**
* Persistent data store in which we store information regarding
* synchronization.
*/
protected final NodeStateStore _stateStore;
/**
* Hash code of contents of the last cluster view we received from
* the peer. Used as optimization: cluster view only piggy-backed
* on list response if hash differs.
*/
protected long _lastClusterHash;
/*
/**********************************************************************
/* Local information for peer (which for us is external but...)
/**********************************************************************
*/
/**
* Synchronization state of this peer
*/
protected ActiveNodeState _syncState;
/*
/**********************************************************************
/* Life-cycle
/**********************************************************************
*/
public ClusterPeerImpl(SharedServiceStuff stuff, ClusterViewByServerUpdatable cluster,
NodeStateStore stateStore, StorableStore entryStore,
ActiveNodeState state,
ClusterStatusAccessor accessor)
{
super();
_cluster = cluster;
_stuff = stuff;
_syncListAccessor = new SyncListAccessor(stuff);
_syncState = state;
_stateStore = stateStore;
_entryStore = entryStore;
_timeMaster = stuff.getTimeMaster();
_entryConverter = stuff.getEntryConverter();
_statusAccessor = accessor;
}
@Override
public void start() {
startSyncing();
}
/**
* If we get notified about a shutdown in future, we can take some
* preparatory steps, to make eventual shutdown simpler, safer
* and quicker.
*/
@Override
public void prepareForStop() {
_stop(false);
}
/**
* Method called when the system is shutting down.
*/
@Override
public void stop() {
_stop(true);
}
protected void _stop(boolean forced)
{
// stopSyncing():
Thread t;
synchronized (this) {
_running.set(false);
t = _syncThread;
if (t != null) {
_syncThread = null;
LOG.info("Stop requested for sync thread for peer at {}", _syncState.getAddress());
}
}
if (t != null) {
// t.notify();
t.interrupt();
}
_syncListAccessor.stop();
}
/*
/**********************************************************************
/* Actual synchronization task
/**********************************************************************
*/
/**
* Method that can be called to ensure that there is a synchronization
* thread running to sync between the local node and this peer
*
* @return True if a new sync thread was started; false if there already
* was a thread
*/
public boolean startSyncing()
{
Thread t;
synchronized (this) {
t = _syncThread;
if (t != null) { // sanity check
return false;
}
_running.set(true);
_syncThread = t = new Thread(new Runnable() {
@Override
public void run() {
syncLoop();
}
});
_syncThread.setDaemon(true);
_syncThread.setName("NodeSync-"+_syncState.getAddress());
}
t.start();
return true;
}
/*
/**********************************************************************
/* State access
/**********************************************************************
*/
@Override
public int getFailCount() { return _failCount.get(); }
@Override
public void resetFailCount() { _failCount.set(0); }
@Override
public long getSyncedUpTo() {
return _syncState.getSyncedUpTo();
}
/*
/**********************************************************************
/* Public API
/**********************************************************************
*/
@Override
public IpAndPort getAddress() {
return _syncState.getAddress();
}
@Override
public KeyRange getActiveRange() {
return _syncState.getRangeActive();
}
@Override
public KeyRange getTotalRange() {
return _syncState.totalRange();
}
/**
* Accessor for getting key range that is shared between the local node
* and this peer; for non-overlapping nodes this may be an empty range.
*/
@Override
public KeyRange getSyncRange() {
return _syncState.getRangeSync();
}
/*
/**********************************************************************
/* Extended accessors
/**********************************************************************
*/
public ActiveNodeState getSyncState() {
return _syncState;
}
public boolean isDisabled() {
return _syncState.isDisabled();
}
/*
/**********************************************************************
/* Background synchronization processing
/**********************************************************************
*/
/**
* Main synchronization loop
*/
protected void syncLoop()
{
LOG.info("Starting sync thread for peer at {}", _syncState.getAddress());
// For testing (and only testing!), let's add little bit of
// virtual sleep (TimeMaster will block threads) before starting
// the loop; this to stabilize situation
if (_stuff.isRunningTests()) {
try {
_timeMaster.sleep(1L);
} catch (InterruptedException e) { }
}
/* At high level, we have two kinds of tasks, depending on whether
* there is any overlap:
*
* 1. If ranges overlap, we need to do proper sync list/pull handling
* 2. If no overlap, we just need to keep an eye towards changes, to
* try to keep whole cluster view up to date (since clients need it)
*/
while (_running.get()) {
try {
if (hasOverlap(_cluster.getLocalState(), _syncState)) {
doRealSync();
} else {
doMinimalSync();
}
} catch (InterruptedException e) {
if (_running.get()) {
LOG.warn("syncLoop() interrupted without clearing '_running' flag; ignoring");
}
} catch (Exception e) {
LOG.warn("Uncaught processing exception during syncLoop(): ({}) {}",
e.getClass().getName(), e.getMessage());
if (_running.get()) {
// Ignore failures during shutdown, so only increase here
_failCount.addAndGet(1);
try {
_timeMaster.sleep(SLEEP_FOR_SYNCPULL_ERRORS_MSECS);
} catch (InterruptedException e2) { }
}
}
}
if (_stuff.isRunningTests()) {
LOG.info("Stopped sync thread for peer at {} -- testing, all done!", _syncState.getAddress());
return;
}
// And send the byebye message if peer is NOT (known to be) disabled
if (_syncState.isDisabled()) {
LOG.info("Stopped sync thread for peer at {}: is disabled, no need to send bye-bye", _syncState.getAddress());
return;
}
LOG.info("Stopped sync thread for peer at {}: let's send bye-bye", _syncState.getAddress());
long start = System.currentTimeMillis();
_syncListAccessor.sendStatusUpdate(_cluster, TIMEOUT_FOR_BYEBYE,
_syncState.getAddress(), ClusterMateConstants.STATE_INACTIVE);
LOG.info("Bye-bye message to {} sent in {} msec", _syncState.getAddress(), System.currentTimeMillis()-start);
}
protected void doRealSync() throws Exception
{
/* Sequence for each iterations consists of:
*
* 1. Fetch list of newly inserted/deleted entries from peer (sync/list)
* 2a. Find subset of entries unknown to this node, if any
* 2b. Fetch unknown entries, possibly with multiple requests
*
* and we will also add bit of sleep between requests, depending on how many
* entries we get in step 1.
*/
long listTime = _timeMaster.currentTimeMillis();
SyncListResponse syncResp = _fetchSyncList();
if (!_running.get()) { // short-circuit during shutdown
return;
}
if (syncResp == null) { // only for hard errors
_timeMaster.sleep(SLEEP_FOR_SYNCLIST_ERRORS_MSECS);
return;
}
// First things first:
if (syncResp.clusterStatus != null) {
_cluster.updateWith(syncResp.clusterStatus);
_lastClusterHash = syncResp.clusterHash;
} else {
// This is fine, as long as hashes match
if (syncResp.clusterHash != _lastClusterHash) {
LOG.warn("Did not get cluster status from {} even though hashes differ 0x{} (old) vs 0x{} (response)",
_syncState.getAddress(),
Long.toHexString(_lastClusterHash), Long.toHexString(syncResp.clusterHash));
}
}
final long lastSeenTimestamp = syncResp.lastSeen();
// Sanity check; should never happen
if (lastSeenTimestamp <= 0L) {
LOG.error("Invalid lastSeen timestamp value {} for SyncList, from {}",
lastSeenTimestamp, _syncState.getAddress());
// should we sleep bit extra? Probably good to avoid flooding logs, if we end up here
Thread.sleep(100L);
}
// comment out or remove for production; left here during testing:
//long diff = (listTime - syncResp.lastSeen()) >> 10; // in seconds
//LOG.warn("Received syncList with {} responses; last timestamp {} secs ago", syncResp.size(), diff);
List newEntries = syncResp.entries;
int insertedEntryCount = newEntries.size();
if (insertedEntryCount == 0) { // nothing to update
// may still need to update timestamp?
_updatePersistentState(listTime, lastSeenTimestamp);
// Ok: maybe server instructed us as to how long to sleep?
long sleepMsecs = syncResp.clientWait;
if (sleepMsecs < MINIMAL_SLEEP_MSECS) { // if not, use some lowish default amount
sleepMsecs = SLEEP_FOR_EMPTY_SYNCLIST_MSECS;
}
// long timeSpent = _timeMaster.currentTimeMillis() - listTime;
_timeMaster.sleep(sleepMsecs);
return;
}
// Ok, we got something, good.
// First: handle tombstones we may be getting:
@SuppressWarnings("unused")
int tombstoneCount = _handleTombstones(newEntries);
// then filter out entries that we already have:
_filterSeen(newEntries);
if (!_running.get()) { // short-circuit during shutdown
return;
}
if (newEntries.isEmpty()) { // nope: just update state then
/*
long msecs = syncResp.lastSeen() - _syncState.syncedUpTo;
if (!_stuff.isRunningTests()) {
LOG.warn("No unseen entries out of {} entries: timestamp = {} (+{} sec)",
new Object[] { insertedEntryCount, syncResp.lastSeen(), String.format("%.1f", msecs/1000.0)});
}
*/
_updatePersistentState(listTime, lastSeenTimestamp);
} else { // yes: need to do batch updates
// but can at least update syncUpTo to first entry, right?
int newCount = newEntries.size();
AtomicInteger rounds = new AtomicInteger(0);
long lastProcessed = _fetchMissing(newEntries, rounds);
int fetched = newCount - newEntries.size();
double secs = (_timeMaster.currentTimeMillis() - listTime) / 1000.0;
String timeDesc = String.format("%.2f", secs);
LOG.info("Fetched {}/{} missing entries from {} in {} seconds ({} rounds)",
new Object[] { fetched, newCount, getAddress(), timeDesc, rounds.get()});
_updatePersistentState(listTime, lastProcessed);
}
// And then sleep a bit, before doing next round of syncing
long msecsBehind = (_timeMaster.currentTimeMillis() - _syncState.getSyncedUpTo());
long delay = _calculateSleepBetweenSync(insertedEntryCount, msecsBehind);
if (delay > 0L) {
// only bother informing if above 50 msec sleep
if (delay >= 50L) {
double secsBehind = delay / 1000.0;
LOG.info("With {} listed entries, {} seconds behind, will do {} msec sleep",
new Object[] { insertedEntryCount, String.format("%.2f", secsBehind), delay});
}
_timeMaster.sleep(delay);
}
}
/**
* Method called when there is no key range overlap, and at most we want to
* synchronize cluster view occasionally.
*/
protected void doMinimalSync() throws Exception
{
LOG.info("doMinimalSync(): let's just... Sleep for a bit (TBD)");
Thread.sleep(30 * 1000L);
// !!! TODO: do something!
}
/*
/**********************************************************************
/* Internal methods, cluster state updates
/**********************************************************************
*/
/**
* Method called to indicate that the node should (or should not) be
* disabled.
*/
public void markDisabled(long timestamp, boolean isDisabled)
{
if (timestamp <= 0L) { // optional
timestamp = _syncState.getDisabledUpdated();
}
ActiveNodeState state = _syncState.withDisabled(timestamp, isDisabled);
if (state != _syncState) {
_syncState = state;
try {
_stateStore.upsertEntry(state.getAddress(), state);
} catch (Exception e) {
LOG.error("Failed to update node state (disabled to {}) for {}. Problem ({}): {}",
isDisabled, _syncState, e.getClass().getName(), e.getMessage());
}
}
}
/*
/**********************************************************************
/* Internal methods, other
/**********************************************************************
*/
/**
* Helper method called to update persistent state, based on sync list
* information.
*
* @param syncStartTime Timestamp of when sync attempt was done
* @param lastSeen
*/
private void _updatePersistentState(long syncStartTime, long lastSeen)
{
ActiveNodeState orig = _syncState;
_syncState = _syncState.withLastSyncAttempt(syncStartTime);
if (lastSeen > _syncState.getSyncedUpTo()) {
_syncState = _syncState.withSyncedUpTo(lastSeen);
}
if (_syncState != orig) {
//LOG.warn("Saving sync state ({}) (args: {}, {}): lastStartTime {}, lastSeen {}", orig.address, syncStartTime, lastSeen, _syncState.lastSyncAttempt, _syncState.syncedUpTo);
try {
_stateStore.upsertEntry(_syncState.getAddress(), _syncState);
} catch (Exception e) {
LOG.error("Failed to update node state for {}. Problem ({}): {}",
_syncState, e.getClass().getName(), e.getMessage());
}
}
}
private SyncListResponse _fetchSyncList() throws InterruptedException
{
try {
return _syncListAccessor.fetchSyncList(_cluster,
TIMEOUT_FOR_SYNCLIST, _syncState, _lastClusterHash);
} catch (InterruptedException e) {
// no point in complaining if we are being shut down:
if (_running.get()) {
LOG.warn("Failed to fetch syncList from {} ({}): {}",
new Object[] { _syncState.getAddress(), e.getClass().getName(), e.getMessage()});
}
}
return null;
}
/**
* Helper method called to handle removal of entries, by handling
* tombstones received and converting existing non-deleted local
* entries into tombstones.
*
* @return Number of tombstone entries found on the list
*/
protected int _handleTombstones(List entries)
throws IOException, StoreException
{
int count = 0;
Iterator it = entries.iterator();
while (it.hasNext()) {
SyncListResponseEntry entry = it.next();
// Tombstone: if we have an entry, convert to a tombstone.
/* 06-Jul-2012, tatu: But if we don't have one, should we create one?
* Could think of it either way; but for now, let's not waste time and space
*/
if (entry.deleted()) {
++count;
it.remove();
_entryStore.softDelete(StoreOperationSource.SYNC, null, entry.key, true, true);
}
}
return count;
}
protected void _filterSeen(List entries)
throws IOException, StoreException
{
Iterator it = entries.iterator();
while (it.hasNext()) {
SyncListResponseEntry remoteEntry = it.next();
/* 11-Jun-2013, tatu: Although tombstones have been handled and removed,
* need to pay attention here, since conflict resolution may be
* necessary.
*/
Storable localEntry = _entryStore.findEntry(StoreOperationSource.SYNC, null, remoteEntry.key);
if (localEntry != null) {
// Do we have an actual conflict? If so, needs resolution as per:
if (StoreUtil.needToPullRemoteToResolve(localEntry.getLastModified(), localEntry.getContentHash(),
remoteEntry.insertionTime, remoteEntry.hash)) {
continue;
}
it.remove();
}
}
}
/**
* Helper method that handles actual fetching of missing entries, to synchronize
* content.
*
* @param missingEntries Entries to try to fetch
* @rounds Integer to update with number of rounds done to sync things completely
*
* @return Timestamp to use as the new 'syncedUpTo' value
*/
private long _fetchMissing(List missingEntries, AtomicInteger rounds)
throws InterruptedException
{
// initially create as big batches as possible
int maxToFetch = missingEntries.size();
int tries = 0;
int fails = 0;
long syncedUpTo = 0L;
do {
++tries;
final long startTime = _timeMaster.currentTimeMillis();
AtomicInteger payloadSize = new AtomicInteger(0);
SyncPullRequest req = _buildSyncPullRequest(missingEntries, maxToFetch, payloadSize);
final int expCount = req.size();
if (expCount == 0) { // sanity check, shouldn't happen but...
throw new IllegalStateException("Internal error: empty syncPullRequest list ("+missingEntries.size()+" missing entries)");
}
rounds.addAndGet(1);
AtomicInteger status = new AtomicInteger(0);
InputStream in = null;
try {
in = _syncListAccessor.readSyncPullResponse(req, TIMEOUT_FOR_SYNCLIST,
getAddress(), status, payloadSize.get());
// } catch (org.apache.http.conn.HttpHostConnectException e) { // if using Apache HC
} catch (java.net.ConnectException e) {
++fails;
LOG.warn("Failed to connect server "+getAddress()+" to fetch missing entries", e);
_timeMaster.sleep(SLEEP_FOR_SYNCPULL_ERRORS_MSECS);
} catch (Exception e) {
LOG.warn("Problem trying to make syncPull call to fetch "+expCount+" entries: ("
+e.getClass().getName() + ") " + e.getMessage(), e);
++fails;
_timeMaster.sleep(SLEEP_FOR_SYNCPULL_ERRORS_MSECS);
}
if (in == null) {
LOG.warn("Problem trying to fetch {} entries, received status code of {}",
expCount, status.get());
_timeMaster.sleep(SLEEP_FOR_SYNCPULL_ERRORS_MSECS);
++fails;
continue;
}
Iterator it = missingEntries.iterator();
int count = 0;
int headerLength = 0;
long payloadLength = 0;
final PullProblems probs = new PullProblems();
try {
// let's see if we can correlate entries nicely
headerLength = -1;
payloadLength = -1;
for (; it.hasNext(); ++count, it.remove()) {
SyncListResponseEntry reqEntry = it.next();
headerLength = SyncPullResponse.readHeaderLength(in);
// Service will indicate end-of-response with marker length
if (headerLength == SyncHandler.LENGTH_EOF) {
break;
}
// sanity check:
if (count == expCount) {
++probs.other;
LOG.warn("Server returned more than expected {} entries; ignoring rest!", expCount);
break;
}
// missing header? Unexpected, but not illegal
if (headerLength == 0) {
if (probs.missing++ == 0) {
LOG.warn("Missing entry {}/{} (from {}), id {}: expired? (will only report first)",
new Object[] { count, expCount, _syncState.getAddress(), reqEntry.key});
}
continue;
}
byte[] headerBytes = new byte[headerLength];
int len = IOUtil.readFully(in, headerBytes);
if (len < headerLength) {
throw new IOException("Unexpected end-of-input: got "+len+" bytes; needed "+headerLength);
}
SyncPullEntry header = _syncListAccessor.decodePullEntry(headerBytes);
payloadLength = header.storageSize;
// and then create the actual entry:
_pullEntry(reqEntry, header, in, probs);
syncedUpTo = reqEntry.insertionTime;
}
if (count < expCount) {
// let's consider 0 entries to be an error, to prevent infinite loops
if (count == 0) {
LOG.warn("Server returned NO entries, when requested "+expCount);
++fails;
}
LOG.warn("Server returned fewer entries than requested for sync pull: {} vs {} (in {} msecs)",
new Object[] { count, expCount, (_timeMaster.currentTimeMillis() - startTime)});
}
if (probs.hasIssues()) {
LOG.warn("Problems with pull request to {}: {}", _syncState.getAddress(), probs);
}
} catch (Exception e) {
LOG.warn("Problem trying to fetch syncPull entry {}/{} (header-length: {}, length: {}): ({}) {}",
new Object[] { count+1, expCount, headerLength, payloadLength, e.getClass().getName(), e.getMessage() } );
_timeMaster.sleep(SLEEP_FOR_SYNCPULL_ERRORS_MSECS);
++fails;
} finally {
if (in != null) {
try { in.close(); } catch (Exception e) { // shouldn't really happen
LOG.warn("Failed to close HTTP stream: {}", e.getMessage());
}
}
}
} while (fails < MAX_SYNC_FAILURES && !missingEntries.isEmpty() && tries < MAX_FETCH_TRIES);
if (fails > 0) {
_failCount.addAndGet(fails);
}
return syncedUpTo;
}
private SyncPullRequest _buildSyncPullRequest(List missingEntries,
int maxEntries, AtomicInteger expectedPayloadSize)
{
SyncPullRequest req = new SyncPullRequest();
Iterator it = missingEntries.iterator();
SyncListResponseEntry entry = it.next();
req.addEntry(entry.key);
long expSize = entry.size;
while (it.hasNext() && req.size() < maxEntries) {
entry = it.next();
expSize += entry.size;
if (expSize > MAX_TOTAL_PAYLOAD) {
expSize -= entry.size;
break;
}
req.addEntry(entry.key);
}
expectedPayloadSize.set((int) expSize);
return req;
}
/**
* Helper method called to figure out how long to sleep before doing next syncList call.
* Note that sleep times are rather arbitrary: we hope to be able to better
* tune these in future.
*
* @param listedCount number of 'newly inserted' entries that were returned
* @param Number of second that we are "behind" current time (note: due to grace period,
* will never be zero, but more like a minute or so at minimum)
*/
private long _calculateSleepBetweenSync(int listedCount, long msecsBehind)
{
// if we are behind by more than 3 minutes, or get >75% "full" response, no delay:
final int FULL_ENOUGH = _stuff.getServiceConfig().cfgMaxEntriesPerSyncList * 3 / 4;
if ((msecsBehind >= (3 * 60 * 1000)) || (listedCount >= FULL_ENOUGH)) {
return 0L;
}
// otherwise moderate delay; bit longer for shorter lists
if (listedCount < 5) {
return 300L;
}
if (listedCount <= 10) {
return 200L;
}
if (listedCount < 40) {
return 100L;
}
return 50L;
}
/**
* Method that does the heavy lifting of pulling a single synchronized entry,
* if and as necessary.
*/
private void _pullEntry(SyncListResponseEntry reqEntry, SyncPullEntry header,
InputStream in, PullProblems probs)
throws IOException
{
final StorableKey key = header.key;
/* first things first: either read things in memory (for inline inclusion),
* or pipe into a file.
*/
long expSize = header.storageSize;
// Sanity check: although rare, deletion could have occurred after we got
// the initial sync list, so:
if (header.isDeleted) {
_entryStore.softDelete(StoreOperationSource.SYNC, null, key, true, true);
return;
}
StorableCreationResult result;
StorableCreationMetadata stdMetadata = new StorableCreationMetadata(header.compression,
header.checksum, header.checksumForCompressed);
stdMetadata.uncompressedSize = header.size;
stdMetadata.storageSize = header.storageSize;
// 16-Apr-2014, tatu: Need to remember to set replica flag now
stdMetadata.replicated = true;
/* 25-Apr-2014, As per [#32], we need to compensate time-to-live settings so that
* it is not reset; rather, it stay as close to remaining TTL as possible.
*
* Note that this means that "maxTTLSecs" IS modified, and "minTTLSecs" NOT, since
* former is measured from creation and latter (if used) from last-access.
*/
ByteContainer customMetadata = _entryConverter.createMetadata(_timeMaster.currentTimeMillis(),
header.lastAccessMethod, header.minTTLSecs, header.maxTTLSecs);
// although not 100% required, we can simplify handling of smallest entries
if (expSize <= _stuff.getServiceConfig().storeConfig.maxInlinedStorageSize) { // inlineable
ByteContainer data;
if (expSize == 0) {
data = ByteContainer.emptyContainer();
} else {
byte[] bytes = new byte[(int) expSize];
int len = IOUtil.readFully(in, bytes);
if (len < expSize) {
throw new IOException("Unexpected end-of-input: got "+len+" bytes; needed "+expSize);
}
data = ByteContainer.simple(bytes);
}
// 19-Sep-2013, tatu: May need to upsert, when resolving conflicts
result = _entryStore.upsertConditionally(StoreOperationSource.SYNC, null, key, data,
stdMetadata, customMetadata, true,
new ConflictOverwriteChecker(reqEntry.insertionTime));
} else {
/* 21-Sep-2012, tatu: Important -- we must ensure that store only reads
* bytes that belong to the entry payload. The easiest way is by adding
* a wrapper stream that ensures this...
*/
BoundedInputStream bin = new BoundedInputStream(in, stdMetadata.storageSize, false);
// 19-Sep-2013, tatu: May need to upsert, when resolving conflicts
result = _entryStore.upsertConditionally(StoreOperationSource.SYNC, null, key, bin,
stdMetadata, customMetadata, true,
new ConflictOverwriteChecker(reqEntry.insertionTime));
if (result.succeeded() && !bin.isCompletelyRead()) { // error or warning?
Storable entry = result.getNewEntry();
long ssize = (entry == null) ? -1L : entry.getStorageLength();
++probs.other;
LOG.warn("Problems with sync-pull for '{}': read {} bytes, should have read {} more; entry storageSize: {}",
new Object[] { header.key, bin.bytesRead(), bin.bytesLeft(), ssize });
}
}
// should we care whether this was redundant or not?
if (!result.succeeded()) {
if (probs.redundant++ == 0) {
if (result.getPreviousEntry() != null) {
// most likely ok: already had the entry
LOG.info("Redundant sync-pull for '{}' (from {}): entry already existed locally (will only report first)",
header.key, _syncState.getAddress());
} else {
// should this add to 'failCount'? For now, don't
LOG.warn("Failed sync-pull for '{}' (from {}): no old entry. Strange! (will only report first)",
header.key, _syncState.getAddress());
}
}
}
}
protected final boolean hasOverlap(NodeState state1, NodeState state2)
{
return state1.totalRange().overlapsWith(state2.totalRange());
}
private static class PullProblems {
public int redundant = 0;
public int missing = 0;
public int other = 0;
public boolean hasIssues() {
return (redundant > 0) || (missing > 0) || (other > 0);
}
@Override
public String toString() {
return new StringBuilder(60)
.append(redundant).append(" redundant, ")
.append(missing).append(" missing entries and ")
.append(other).append(" other problems")
.toString();
}
}
}