View Javadoc

1   /**
2    * Licensed to the Apache Software Foundation (ASF) under one
3    * or more contributor license agreements.  See the NOTICE file
4    * distributed with this work for additional information
5    * regarding copyright ownership.  The ASF licenses this file
6    * to you under the Apache License, Version 2.0 (the
7    * "License"); you may not use this file except in compliance
8    * with the License.  You may obtain a copy of the License at
9    *
10   *     http://www.apache.org/licenses/LICENSE-2.0
11   *
12   * Unless required by applicable law or agreed to in writing, software
13   * distributed under the License is distributed on an "AS IS" BASIS,
14   * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15   * See the License for the specific language governing permissions and
16   * limitations under the License.
17   */
18  package org.apache.hadoop.hbase.regionserver.snapshot;
19  
20  import java.io.IOException;
21  import java.util.ArrayList;
22  import java.util.Collection;
23  import java.util.List;
24  import java.util.concurrent.Callable;
25  import java.util.concurrent.ExecutionException;
26  import java.util.concurrent.ExecutorCompletionService;
27  import java.util.concurrent.Future;
28  import java.util.concurrent.LinkedBlockingQueue;
29  import java.util.concurrent.ThreadPoolExecutor;
30  import java.util.concurrent.TimeUnit;
31  
32  import org.apache.commons.logging.Log;
33  import org.apache.commons.logging.LogFactory;
34  import org.apache.hadoop.classification.InterfaceAudience;
35  import org.apache.hadoop.classification.InterfaceStability;
36  import org.apache.hadoop.conf.Configuration;
37  import org.apache.hadoop.hbase.DaemonThreadFactory;
38  import org.apache.hadoop.hbase.TableName;
39  import org.apache.hadoop.hbase.errorhandling.ForeignException;
40  import org.apache.hadoop.hbase.errorhandling.ForeignExceptionDispatcher;
41  import org.apache.hadoop.hbase.master.snapshot.MasterSnapshotVerifier;
42  import org.apache.hadoop.hbase.master.snapshot.SnapshotManager;
43  import org.apache.hadoop.hbase.procedure.ProcedureMember;
44  import org.apache.hadoop.hbase.procedure.ProcedureMemberRpcs;
45  import org.apache.hadoop.hbase.procedure.RegionServerProcedureManager;
46  import org.apache.hadoop.hbase.procedure.Subprocedure;
47  import org.apache.hadoop.hbase.procedure.SubprocedureFactory;
48  import org.apache.hadoop.hbase.procedure.ZKProcedureMemberRpcs;
49  import org.apache.hadoop.hbase.protobuf.generated.HBaseProtos.SnapshotDescription;
50  import org.apache.hadoop.hbase.regionserver.HRegion;
51  import org.apache.hadoop.hbase.regionserver.HRegionServer;
52  import org.apache.hadoop.hbase.regionserver.RegionServerServices;
53  import org.apache.hadoop.hbase.snapshot.SnapshotCreationException;
54  import org.apache.hadoop.hbase.zookeeper.ZooKeeperWatcher;
55  import org.apache.zookeeper.KeeperException;
56  
57  import com.google.protobuf.InvalidProtocolBufferException;
58  
59  /**
60   * This manager class handles the work dealing with snapshots for a {@link HRegionServer}.
61   * <p>
62   * This provides the mechanism necessary to kick off a online snapshot specific
63   * {@link Subprocedure} that is responsible for the regions being served by this region server.
64   * If any failures occur with the subprocedure, the RegionSeverSnapshotManager's subprocedure
65   * handler, {@link ProcedureMember}, notifies the master's ProcedureCoordinator to abort all
66   * others.
67   * <p>
68   * On startup, requires {@link #start()} to be called.
69   * <p>
70   * On shutdown, requires {@link #stop(boolean)} to be called
71   */
72  @InterfaceAudience.Private
73  @InterfaceStability.Unstable
74  public class RegionServerSnapshotManager extends RegionServerProcedureManager {
75    private static final Log LOG = LogFactory.getLog(RegionServerSnapshotManager.class);
76  
77    /** Maximum number of snapshot region tasks that can run concurrently */
78    private static final String CONCURENT_SNAPSHOT_TASKS_KEY = "hbase.snapshot.region.concurrentTasks";
79    private static final int DEFAULT_CONCURRENT_SNAPSHOT_TASKS = 3;
80  
81    /** Conf key for number of request threads to start snapshots on regionservers */
82    public static final String SNAPSHOT_REQUEST_THREADS_KEY = "hbase.snapshot.region.pool.threads";
83    /** # of threads for snapshotting regions on the rs. */
84    public static final int SNAPSHOT_REQUEST_THREADS_DEFAULT = 10;
85  
86    /** Conf key for max time to keep threads in snapshot request pool waiting */
87    public static final String SNAPSHOT_TIMEOUT_MILLIS_KEY = "hbase.snapshot.region.timeout";
88    /** Keep threads alive in request pool for max of 60 seconds */
89    public static final long SNAPSHOT_TIMEOUT_MILLIS_DEFAULT = 60000;
90  
91    /** Conf key for millis between checks to see if snapshot completed or if there are errors*/
92    public static final String SNAPSHOT_REQUEST_WAKE_MILLIS_KEY = "hbase.snapshot.region.wakefrequency";
93    /** Default amount of time to check for errors while regions finish snapshotting */
94    private static final long SNAPSHOT_REQUEST_WAKE_MILLIS_DEFAULT = 500;
95  
96    private RegionServerServices rss;
97    private ProcedureMemberRpcs memberRpcs;
98    private ProcedureMember member;
99  
100   /**
101    * Exposed for testing.
102    * @param conf HBase configuration.
103    * @param parent parent running the snapshot handler
104    * @param memberRpc use specified memberRpc instance
105    * @param procMember use specified ProcedureMember
106    */
107    RegionServerSnapshotManager(Configuration conf, HRegionServer parent,
108       ProcedureMemberRpcs memberRpc, ProcedureMember procMember) {
109     this.rss = parent;
110     this.memberRpcs = memberRpc;
111     this.member = procMember;
112   }
113 
114   public RegionServerSnapshotManager() {}
115 
116   /**
117    * Start accepting snapshot requests.
118    */
119   @Override
120   public void start() {
121     LOG.debug("Start Snapshot Manager " + rss.getServerName().toString());
122     this.memberRpcs.start(rss.getServerName().toString(), member);
123   }
124 
125   /**
126    * Close <tt>this</tt> and all running snapshot tasks
127    * @param force forcefully stop all running tasks
128    * @throws IOException
129    */
130   @Override
131   public void stop(boolean force) throws IOException {
132     String mode = force ? "abruptly" : "gracefully";
133     LOG.info("Stopping RegionServerSnapshotManager " + mode + ".");
134 
135     try {
136       this.member.close();
137     } finally {
138       this.memberRpcs.close();
139     }
140   }
141 
142   /**
143    * If in a running state, creates the specified subprocedure for handling an online snapshot.
144    *
145    * Because this gets the local list of regions to snapshot and not the set the master had,
146    * there is a possibility of a race where regions may be missed.  This detected by the master in
147    * the snapshot verification step.
148    *
149    * @param snapshot
150    * @return Subprocedure to submit to the ProcedureMemeber.
151    */
152   public Subprocedure buildSubprocedure(SnapshotDescription snapshot) {
153 
154     // don't run a snapshot if the parent is stop(ping)
155     if (rss.isStopping() || rss.isStopped()) {
156       throw new IllegalStateException("Can't start snapshot on RS: " + rss.getServerName()
157           + ", because stopping/stopped!");
158     }
159 
160     // check to see if this server is hosting any regions for the snapshots
161     // check to see if we have regions for the snapshot
162     List<HRegion> involvedRegions;
163     try {
164       involvedRegions = getRegionsToSnapshot(snapshot);
165     } catch (IOException e1) {
166       throw new IllegalStateException("Failed to figure out if we should handle a snapshot - "
167           + "something has gone awry with the online regions.", e1);
168     }
169 
170     // We need to run the subprocedure even if we have no relevant regions.  The coordinator
171     // expects participation in the procedure and without sending message the snapshot attempt
172     // will hang and fail.
173 
174     LOG.debug("Launching subprocedure for snapshot " + snapshot.getName() + " from table "
175         + snapshot.getTable());
176     ForeignExceptionDispatcher exnDispatcher = new ForeignExceptionDispatcher(snapshot.getName());
177     Configuration conf = rss.getConfiguration();
178     long timeoutMillis = conf.getLong(SNAPSHOT_TIMEOUT_MILLIS_KEY,
179         SNAPSHOT_TIMEOUT_MILLIS_DEFAULT);
180     long wakeMillis = conf.getLong(SNAPSHOT_REQUEST_WAKE_MILLIS_KEY,
181         SNAPSHOT_REQUEST_WAKE_MILLIS_DEFAULT);
182 
183     switch (snapshot.getType()) {
184     case FLUSH:
185       SnapshotSubprocedurePool taskManager =
186         new SnapshotSubprocedurePool(rss.getServerName().toString(), conf);
187       return new FlushSnapshotSubprocedure(member, exnDispatcher, wakeMillis,
188           timeoutMillis, involvedRegions, snapshot, taskManager);
189     default:
190       throw new UnsupportedOperationException("Unrecognized snapshot type:" + snapshot.getType());
191     }
192   }
193 
194   /**
195    * Determine if the snapshot should be handled on this server
196    *
197    * NOTE: This is racy -- the master expects a list of regionservers.
198    * This means if a region moves somewhere between the calls we'll miss some regions.
199    * For example, a region move during a snapshot could result in a region to be skipped or done
200    * twice.  This is manageable because the {@link MasterSnapshotVerifier} will double check the
201    * region lists after the online portion of the snapshot completes and will explicitly fail the
202    * snapshot.
203    *
204    * @param snapshot
205    * @return the list of online regions. Empty list is returned if no regions are responsible for
206    *         the given snapshot.
207    * @throws IOException
208    */
209   private List<HRegion> getRegionsToSnapshot(SnapshotDescription snapshot) throws IOException {
210     return rss.getOnlineRegions(TableName.valueOf(snapshot.getTable()));
211   }
212 
213   /**
214    * Build the actual snapshot runner that will do all the 'hard' work
215    */
216   public class SnapshotSubprocedureBuilder implements SubprocedureFactory {
217 
218     @Override
219     public Subprocedure buildSubprocedure(String name, byte[] data) {
220       try {
221         // unwrap the snapshot information
222         SnapshotDescription snapshot = SnapshotDescription.parseFrom(data);
223         return RegionServerSnapshotManager.this.buildSubprocedure(snapshot);
224       } catch (InvalidProtocolBufferException e) {
225         throw new IllegalArgumentException("Could not read snapshot information from request.");
226       }
227     }
228 
229   }
230 
231   /**
232    * We use the SnapshotSubprocedurePool, a class specific thread pool instead of
233    * {@link org.apache.hadoop.hbase.executor.ExecutorService}.
234    *
235    * It uses a {@link java.util.concurrent.ExecutorCompletionService} which provides queuing of
236    * completed tasks which lets us efficiently cancel pending tasks upon the earliest operation
237    * failures.
238    *
239    * HBase's ExecutorService (different from {@link java.util.concurrent.ExecutorService}) isn't
240    * really built for coordinated tasks where multiple threads as part of one larger task.  In
241    * RS's the HBase Executor services are only used for open and close and not other threadpooled
242    * operations such as compactions and replication  sinks.
243    */
244   static class SnapshotSubprocedurePool {
245     private final ExecutorCompletionService<Void> taskPool;
246     private final ThreadPoolExecutor executor;
247     private volatile boolean stopped;
248     private final List<Future<Void>> futures = new ArrayList<Future<Void>>();
249     private final String name;
250 
251     SnapshotSubprocedurePool(String name, Configuration conf) {
252       // configure the executor service
253       long keepAlive = conf.getLong(
254         RegionServerSnapshotManager.SNAPSHOT_TIMEOUT_MILLIS_KEY,
255         RegionServerSnapshotManager.SNAPSHOT_TIMEOUT_MILLIS_DEFAULT);
256       int threads = conf.getInt(CONCURENT_SNAPSHOT_TASKS_KEY, DEFAULT_CONCURRENT_SNAPSHOT_TASKS);
257       this.name = name;
258       executor = new ThreadPoolExecutor(1, threads, keepAlive, TimeUnit.MILLISECONDS,
259           new LinkedBlockingQueue<Runnable>(), new DaemonThreadFactory("rs("
260               + name + ")-snapshot-pool"));
261       taskPool = new ExecutorCompletionService<Void>(executor);
262     }
263 
264     boolean hasTasks() {
265       return futures.size() != 0;
266     }
267 
268     /**
269      * Submit a task to the pool.
270      *
271      * NOTE: all must be submitted before you can safely {@link #waitForOutstandingTasks()}. This
272      * version does not support issuing tasks from multiple concurrent table snapshots requests.
273      */
274     void submitTask(final Callable<Void> task) {
275       Future<Void> f = this.taskPool.submit(task);
276       futures.add(f);
277     }
278 
279     /**
280      * Wait for all of the currently outstanding tasks submitted via {@link #submitTask(Callable)}.
281      * This *must* be called after all tasks are submitted via submitTask.
282      *
283      * @return <tt>true</tt> on success, <tt>false</tt> otherwise
284      * @throws InterruptedException
285      * @throws SnapshotCreationException if the snapshot failed while we were waiting
286      */
287     boolean waitForOutstandingTasks() throws ForeignException, InterruptedException {
288       LOG.debug("Waiting for local region snapshots to finish.");
289 
290       int sz = futures.size();
291       try {
292         // Using the completion service to process the futures that finish first first.
293         for (int i = 0; i < sz; i++) {
294           Future<Void> f = taskPool.take();
295           f.get();
296           if (!futures.remove(f)) {
297             LOG.warn("unexpected future" + f);
298           }
299           LOG.debug("Completed " + (i+1) + "/" + sz +  " local region snapshots.");
300         }
301         LOG.debug("Completed " + sz +  " local region snapshots.");
302         return true;
303       } catch (InterruptedException e) {
304         LOG.warn("Got InterruptedException in SnapshotSubprocedurePool", e);
305         if (!stopped) {
306           Thread.currentThread().interrupt();
307           throw new ForeignException("SnapshotSubprocedurePool", e);
308         }
309         // we are stopped so we can just exit.
310       } catch (ExecutionException e) {
311         if (e.getCause() instanceof ForeignException) {
312           LOG.warn("Rethrowing ForeignException from SnapshotSubprocedurePool", e);
313           throw (ForeignException)e.getCause();
314         }
315         LOG.warn("Got Exception in SnapshotSubprocedurePool", e);
316         throw new ForeignException(name, e.getCause());
317       } finally {
318         cancelTasks();
319       }
320       return false;
321     }
322 
323     /**
324      * This attempts to cancel out all pending and in progress tasks (interruptions issues)
325      * @throws InterruptedException
326      */
327     void cancelTasks() throws InterruptedException {
328       Collection<Future<Void>> tasks = futures;
329       LOG.debug("cancelling " + tasks.size() + " tasks for snapshot " + name);
330       for (Future<Void> f: tasks) {
331         // TODO Ideally we'd interrupt hbase threads when we cancel.  However it seems that there
332         // are places in the HBase code where row/region locks are taken and not released in a
333         // finally block.  Thus we cancel without interrupting.  Cancellations will be slower to
334         // complete but we won't suffer from unreleased locks due to poor code discipline.
335         f.cancel(false);
336       }
337 
338       // evict remaining tasks and futures from taskPool.
339       while (!futures.isEmpty()) {
340         // block to remove cancelled futures;
341         LOG.warn("Removing cancelled elements from taskPool");
342         futures.remove(taskPool.take());
343       }
344       stop();
345     }
346 
347     /**
348      * Abruptly shutdown the thread pool.  Call when exiting a region server.
349      */
350     void stop() {
351       if (this.stopped) return;
352 
353       this.stopped = true;
354       this.executor.shutdownNow();
355     }
356   }
357 
358   /**
359    * Create a default snapshot handler - uses a zookeeper based member controller.
360    * @param rss region server running the handler
361    * @throws KeeperException if the zookeeper cluster cannot be reached
362    */
363   @Override
364   public void initialize(RegionServerServices rss) throws KeeperException {
365     this.rss = rss;
366     ZooKeeperWatcher zkw = rss.getZooKeeper();
367     this.memberRpcs = new ZKProcedureMemberRpcs(zkw,
368         SnapshotManager.ONLINE_SNAPSHOT_CONTROLLER_DESCRIPTION);
369 
370     // read in the snapshot request configuration properties
371     Configuration conf = rss.getConfiguration();
372     long keepAlive = conf.getLong(SNAPSHOT_TIMEOUT_MILLIS_KEY, SNAPSHOT_TIMEOUT_MILLIS_DEFAULT);
373     int opThreads = conf.getInt(SNAPSHOT_REQUEST_THREADS_KEY, SNAPSHOT_REQUEST_THREADS_DEFAULT);
374 
375     // create the actual snapshot procedure member
376     ThreadPoolExecutor pool = ProcedureMember.defaultPool(rss.getServerName().toString(),
377       opThreads, keepAlive);
378     this.member = new ProcedureMember(memberRpcs, pool, new SnapshotSubprocedureBuilder());
379   }
380 
381   @Override
382   public String getProcedureSignature() {
383     return SnapshotManager.ONLINE_SNAPSHOT_CONTROLLER_DESCRIPTION;
384   }
385 
386 }