use of org.apache.accumulo.server.master.state.MergeInfo in project accumulo by apache.
the class Master method getMergeInfo.
public MergeInfo getMergeInfo(Table.ID tableId) {
synchronized (mergeLock) {
try {
String path = ZooUtil.getRoot(getInstance().getInstanceID()) + Constants.ZTABLES + "/" + tableId + "/merge";
if (!ZooReaderWriter.getInstance().exists(path))
return new MergeInfo();
byte[] data = ZooReaderWriter.getInstance().getData(path, new Stat());
DataInputBuffer in = new DataInputBuffer();
in.reset(data, data.length);
MergeInfo info = new MergeInfo();
info.readFields(in);
return info;
} catch (KeeperException.NoNodeException ex) {
log.info("Error reading merge state, it probably just finished");
return new MergeInfo();
} catch (Exception ex) {
log.warn("Unexpected error reading merge state", ex);
return new MergeInfo();
}
}
}
use of org.apache.accumulo.server.master.state.MergeInfo in project accumulo by apache.
the class TabletStateChangeIteratorIT method test.
@Test
public void test() throws AccumuloException, AccumuloSecurityException, TableExistsException, TableNotFoundException {
String[] tables = getUniqueNames(4);
final String t1 = tables[0];
final String t2 = tables[1];
final String t3 = tables[2];
final String cloned = tables[3];
// create some metadata
createTable(t1, true);
createTable(t2, false);
createTable(t3, true);
// examine a clone of the metadata table, so we can manipulate it
cloneMetadataTable(cloned);
State state = new State();
while (findTabletsNeedingAttention(cloned, state) > 0) {
UtilWaitThread.sleep(500);
}
assertEquals("No tables should need attention", 0, findTabletsNeedingAttention(cloned, state));
// test the assigned case (no location)
removeLocation(cloned, t3);
assertEquals("Should have two tablets without a loc", 2, findTabletsNeedingAttention(cloned, state));
// test the cases where the assignment is to a dead tserver
getConnector().tableOperations().delete(cloned);
cloneMetadataTable(cloned);
reassignLocation(cloned, t3);
assertEquals("Should have one tablet that needs to be unassigned", 1, findTabletsNeedingAttention(cloned, state));
// test the cases where there is ongoing merges
state = new State() {
@Override
public Collection<MergeInfo> merges() {
Table.ID tableIdToModify = Table.ID.of(getConnector().tableOperations().tableIdMap().get(t3));
return Collections.singletonList(new MergeInfo(new KeyExtent(tableIdToModify, null, null), MergeInfo.Operation.MERGE));
}
};
assertEquals("Should have 2 tablets that need to be chopped or unassigned", 1, findTabletsNeedingAttention(cloned, state));
// test the bad tablet location state case (inconsistent metadata)
state = new State();
cloneMetadataTable(cloned);
addDuplicateLocation(cloned, t3);
assertEquals("Should have 1 tablet that needs a metadata repair", 1, findTabletsNeedingAttention(cloned, state));
// clean up
dropTables(t1, t2, t3, cloned);
}
use of org.apache.accumulo.server.master.state.MergeInfo in project accumulo by apache.
the class TabletGroupWatcher method run.
@Override
public void run() {
Thread.currentThread().setName("Watching " + store.name());
int[] oldCounts = new int[TabletState.values().length];
EventCoordinator.Listener eventListener = this.master.nextEvent.getListener();
WalStateManager wals = new WalStateManager(master.getInstance(), ZooReaderWriter.getInstance());
while (this.master.stillMaster()) {
// slow things down a little, otherwise we spam the logs when there are many wake-up events
sleepUninterruptibly(100, TimeUnit.MILLISECONDS);
masterState = master.getMasterState();
int totalUnloaded = 0;
int unloaded = 0;
ClosableIterator<TabletLocationState> iter = null;
try {
Map<Table.ID, MergeStats> mergeStatsCache = new HashMap<>();
Map<Table.ID, MergeStats> currentMerges = new HashMap<>();
for (MergeInfo merge : master.merges()) {
if (merge.getExtent() != null) {
currentMerges.put(merge.getExtent().getTableId(), new MergeStats(merge));
}
}
// Get the current status for the current list of tservers
SortedMap<TServerInstance, TabletServerStatus> currentTServers = new TreeMap<>();
for (TServerInstance entry : this.master.tserverSet.getCurrentServers()) {
currentTServers.put(entry, this.master.tserverStatus.get(entry));
}
if (currentTServers.size() == 0) {
eventListener.waitForEvents(Master.TIME_TO_WAIT_BETWEEN_SCANS);
synchronized (this) {
lastScanServers = ImmutableSortedSet.of();
}
continue;
}
// Don't move tablets to servers that are shutting down
SortedMap<TServerInstance, TabletServerStatus> destinations = new TreeMap<>(currentTServers);
destinations.keySet().removeAll(this.master.serversToShutdown);
List<Assignment> assignments = new ArrayList<>();
List<Assignment> assigned = new ArrayList<>();
List<TabletLocationState> assignedToDeadServers = new ArrayList<>();
List<TabletLocationState> suspendedToGoneServers = new ArrayList<>();
Map<KeyExtent, TServerInstance> unassigned = new HashMap<>();
Map<TServerInstance, List<Path>> logsForDeadServers = new TreeMap<>();
MasterState masterState = master.getMasterState();
int[] counts = new int[TabletState.values().length];
stats.begin();
// Walk through the tablets in our store, and work tablets
// towards their goal
iter = store.iterator();
while (iter.hasNext()) {
TabletLocationState tls = iter.next();
if (tls == null) {
continue;
}
Master.log.debug("{} location State: {}", store.name(), tls);
// ignore entries for tables that do not exist in zookeeper
if (TableManager.getInstance().getTableState(tls.extent.getTableId()) == null)
continue;
if (Master.log.isTraceEnabled())
Master.log.trace("{} walogs {}", tls, tls.walogs.size());
// Don't overwhelm the tablet servers with work
if (unassigned.size() + unloaded > Master.MAX_TSERVER_WORK_CHUNK * currentTServers.size()) {
flushChanges(destinations, assignments, assigned, assignedToDeadServers, logsForDeadServers, suspendedToGoneServers, unassigned);
assignments.clear();
assigned.clear();
assignedToDeadServers.clear();
suspendedToGoneServers.clear();
unassigned.clear();
unloaded = 0;
eventListener.waitForEvents(Master.TIME_TO_WAIT_BETWEEN_SCANS);
}
Table.ID tableId = tls.extent.getTableId();
TableConfiguration tableConf = this.master.getConfigurationFactory().getTableConfiguration(tableId);
MergeStats mergeStats = mergeStatsCache.get(tableId);
if (mergeStats == null) {
mergeStats = currentMerges.get(tableId);
if (mergeStats == null) {
mergeStats = new MergeStats(new MergeInfo());
}
mergeStatsCache.put(tableId, mergeStats);
}
TabletGoalState goal = this.master.getGoalState(tls, mergeStats.getMergeInfo());
TServerInstance server = tls.getServer();
TabletState state = tls.getState(currentTServers.keySet());
if (Master.log.isTraceEnabled()) {
Master.log.trace("Goal state {} current {} for {}", goal, state, tls.extent);
}
stats.update(tableId, state);
mergeStats.update(tls.extent, state, tls.chopped, !tls.walogs.isEmpty());
sendChopRequest(mergeStats.getMergeInfo(), state, tls);
sendSplitRequest(mergeStats.getMergeInfo(), state, tls);
// Always follow through with assignments
if (state == TabletState.ASSIGNED) {
goal = TabletGoalState.HOSTED;
}
// if we are shutting down all the tabletservers, we have to do it in order
if (goal == TabletGoalState.SUSPENDED && state == TabletState.HOSTED) {
if (this.master.serversToShutdown.equals(currentTServers.keySet())) {
if (dependentWatcher != null && dependentWatcher.assignedOrHosted() > 0) {
goal = TabletGoalState.HOSTED;
}
}
}
if (goal == TabletGoalState.HOSTED) {
if (state != TabletState.HOSTED && !tls.walogs.isEmpty()) {
if (this.master.recoveryManager.recoverLogs(tls.extent, tls.walogs))
continue;
}
switch(state) {
case HOSTED:
if (server.equals(this.master.migrations.get(tls.extent)))
this.master.migrations.remove(tls.extent);
break;
case ASSIGNED_TO_DEAD_SERVER:
assignedToDeadServers.add(tls);
if (server.equals(this.master.migrations.get(tls.extent)))
this.master.migrations.remove(tls.extent);
TServerInstance tserver = tls.futureOrCurrent();
if (!logsForDeadServers.containsKey(tserver)) {
logsForDeadServers.put(tserver, wals.getWalsInUse(tserver));
}
break;
case SUSPENDED:
if (master.getSteadyTime() - tls.suspend.suspensionTime < tableConf.getTimeInMillis(Property.TABLE_SUSPEND_DURATION)) {
// Tablet is suspended. See if its tablet server is back.
TServerInstance returnInstance = null;
Iterator<TServerInstance> find = destinations.tailMap(new TServerInstance(tls.suspend.server, " ")).keySet().iterator();
if (find.hasNext()) {
TServerInstance found = find.next();
if (found.getLocation().equals(tls.suspend.server)) {
returnInstance = found;
}
}
// Old tablet server is back. Return this tablet to its previous owner.
if (returnInstance != null) {
assignments.add(new Assignment(tls.extent, returnInstance));
} else {
// leave suspended, don't ask for a new assignment.
}
} else {
// Treat as unassigned, ask for a new assignment.
unassigned.put(tls.extent, server);
}
break;
case UNASSIGNED:
// maybe it's a finishing migration
TServerInstance dest = this.master.migrations.get(tls.extent);
if (dest != null) {
// if destination is still good, assign it
if (destinations.keySet().contains(dest)) {
assignments.add(new Assignment(tls.extent, dest));
} else {
// get rid of this migration
this.master.migrations.remove(tls.extent);
unassigned.put(tls.extent, server);
}
} else {
unassigned.put(tls.extent, server);
}
break;
case ASSIGNED:
// Send another reminder
assigned.add(new Assignment(tls.extent, tls.future));
break;
}
} else {
switch(state) {
case SUSPENDED:
// Request a move to UNASSIGNED, so as to allow balancing to continue.
suspendedToGoneServers.add(tls);
cancelOfflineTableMigrations(tls);
break;
case UNASSIGNED:
cancelOfflineTableMigrations(tls);
break;
case ASSIGNED_TO_DEAD_SERVER:
assignedToDeadServers.add(tls);
if (!logsForDeadServers.containsKey(tls.futureOrCurrent())) {
logsForDeadServers.put(tls.futureOrCurrent(), wals.getWalsInUse(tls.futureOrCurrent()));
}
break;
case HOSTED:
TServerConnection conn = this.master.tserverSet.getConnection(server);
if (conn != null) {
conn.unloadTablet(this.master.masterLock, tls.extent, goal.howUnload(), master.getSteadyTime());
unloaded++;
totalUnloaded++;
} else {
Master.log.warn("Could not connect to server {}", server);
}
break;
case ASSIGNED:
break;
}
}
counts[state.ordinal()]++;
}
flushChanges(destinations, assignments, assigned, assignedToDeadServers, logsForDeadServers, suspendedToGoneServers, unassigned);
// provide stats after flushing changes to avoid race conditions w/ delete table
stats.end(masterState);
// Report changes
for (TabletState state : TabletState.values()) {
int i = state.ordinal();
if (counts[i] > 0 && counts[i] != oldCounts[i]) {
this.master.nextEvent.event("[%s]: %d tablets are %s", store.name(), counts[i], state.name());
}
}
Master.log.debug(String.format("[%s]: scan time %.2f seconds", store.name(), stats.getScanTime() / 1000.));
oldCounts = counts;
if (totalUnloaded > 0) {
this.master.nextEvent.event("[%s]: %d tablets unloaded", store.name(), totalUnloaded);
}
updateMergeState(mergeStatsCache);
synchronized (this) {
lastScanServers = ImmutableSortedSet.copyOf(currentTServers.keySet());
}
if (this.master.tserverSet.getCurrentServers().equals(currentTServers.keySet())) {
Master.log.debug(String.format("[%s] sleeping for %.2f seconds", store.name(), Master.TIME_TO_WAIT_BETWEEN_SCANS / 1000.));
eventListener.waitForEvents(Master.TIME_TO_WAIT_BETWEEN_SCANS);
} else {
Master.log.info("Detected change in current tserver set, re-running state machine.");
}
} catch (Exception ex) {
Master.log.error("Error processing table state for store " + store.name(), ex);
if (ex.getCause() != null && ex.getCause() instanceof BadLocationStateException) {
repairMetadata(((BadLocationStateException) ex.getCause()).getEncodedEndRow());
} else {
sleepUninterruptibly(Master.WAIT_BETWEEN_ERRORS, TimeUnit.MILLISECONDS);
}
} finally {
if (iter != null) {
try {
iter.close();
} catch (IOException ex) {
Master.log.warn("Error closing TabletLocationState iterator: " + ex, ex);
}
}
}
}
}
use of org.apache.accumulo.server.master.state.MergeInfo in project accumulo by apache.
the class TableRangeOp method call.
@Override
public Repo<Master> call(long tid, Master env) throws Exception {
if (RootTable.ID.equals(tableId) && Operation.MERGE.equals(op)) {
log.warn("Attempt to merge tablets for {} does nothing. It is not splittable.", RootTable.NAME);
}
Text start = startRow.length == 0 ? null : new Text(startRow);
Text end = endRow.length == 0 ? null : new Text(endRow);
if (start != null && end != null)
if (start.compareTo(end) >= 0)
throw new AcceptableThriftTableOperationException(tableId.canonicalID(), null, TableOperation.MERGE, TableOperationExceptionType.BAD_RANGE, "start row must be less than end row");
env.mustBeOnline(tableId);
MergeInfo info = env.getMergeInfo(tableId);
if (info.getState() == MergeState.NONE) {
KeyExtent range = new KeyExtent(tableId, end, start);
env.setMergeState(new MergeInfo(range, op), MergeState.STARTED);
}
return new TableRangeOpWait(namespaceId, tableId);
}
use of org.apache.accumulo.server.master.state.MergeInfo in project accumulo by apache.
the class TableRangeOp method undo.
@Override
public void undo(long tid, Master env) throws Exception {
// Not sure this is a good thing to do. The Master state engine should be the one to remove it.
MergeInfo mergeInfo = env.getMergeInfo(tableId);
if (mergeInfo.getState() != MergeState.NONE)
log.info("removing merge information {}", mergeInfo);
env.clearMergeState(tableId);
Utils.unreserveNamespace(namespaceId, tid, false);
Utils.unreserveTable(tableId, tid, true);
}
Aggregations