Index: hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java =================================================================== --- hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java (revision 1403958) +++ hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java (working copy) @@ -81,7 +81,6 @@ import org.apache.hadoop.hbase.HConstants.OperationStatusCode; import org.apache.hadoop.hbase.HDFSBlocksDistribution; import org.apache.hadoop.hbase.HRegionInfo; -import org.apache.hadoop.hbase.HServerInfo; import org.apache.hadoop.hbase.HTableDescriptor; import org.apache.hadoop.hbase.KeyValue; import org.apache.hadoop.hbase.NotServingRegionException; @@ -116,7 +115,7 @@ import org.apache.hadoop.hbase.ipc.RpcCallContext; import org.apache.hadoop.hbase.monitoring.MonitoredTask; import org.apache.hadoop.hbase.monitoring.TaskMonitor; -import org.apache.hadoop.hbase.protobuf.generated.ClientProtos; +import org.apache.hadoop.hbase.regionserver.MultiVersionConsistencyControl.WriteEntry; import org.apache.hadoop.hbase.regionserver.compactions.CompactionRequest; import org.apache.hadoop.hbase.regionserver.metrics.OperationMetrics; import org.apache.hadoop.hbase.regionserver.metrics.RegionMetricsStorage; @@ -4544,11 +4543,7 @@ // TODO: There's a lot of boiler plate code identical // to increment... See how to better unify that. /** - * * Perform one or more append operations on a row. - *

- * Appends performed are done under row lock but reads do not take locks out - * so this can be seen partially complete by gets and scans. * * @param append * @param lockid @@ -4558,7 +4553,6 @@ */ public Result append(Append append, Integer lockid, boolean writeToWAL) throws IOException { - // TODO: Use MVCC to make this set of appends atomic to reads byte[] row = append.getRow(); checkRow(row, "append"); boolean flush = false; @@ -4572,9 +4566,16 @@ // Lock row startRegionOperation(); this.writeRequestsCount.increment(); + WriteEntry w = null; try { Integer lid = getLock(lockid, row, true); this.updatesLock.readLock().lock(); + // wait for all prior MVCC transactions to finish - while we hold the row lock + // (so that we are guaranteed to see the latest state) + WriteEntry tmp = mvcc.beginMemstoreInsert(); + mvcc.completeMemstoreInsert(tmp); + // now start my own transaction + w = mvcc.beginMemstoreInsert(); try { long now = EnvironmentEdgeManager.currentTimeMillis(); // Process each family @@ -4637,6 +4638,7 @@ newKV.getBuffer(), newKV.getQualifierOffset(), kv.getQualifierLength()); + newKV.setMemstoreTS(w.getWriteNumber()); kvs.add(newKV); // Append update to WAL @@ -4666,7 +4668,7 @@ //Actually write to Memstore now for (Map.Entry> entry : tempMemstore.entrySet()) { Store store = entry.getKey(); - size += store.upsert(entry.getValue()); + size += store.upsert(entry.getValue(), getSmallestReadPoint()); allKVs.addAll(entry.getValue()); } size = this.addAndGetGlobalMemstoreSize(size); @@ -4679,6 +4681,9 @@ syncOrDefer(txid); // sync the transaction log outside the rowlock } } finally { + if (w != null) { + mvcc.completeMemstoreInsert(w); + } closeRegionOperation(); } @@ -4696,11 +4701,7 @@ } /** - * * Perform one or more increment operations on a row. - *

- * Increments performed are done under row lock but reads do not take locks - * out so this can be seen partially complete by gets and scans. * @param increment * @param lockid * @param writeToWAL @@ -4710,7 +4711,6 @@ public Result increment(Increment increment, Integer lockid, boolean writeToWAL) throws IOException { - // TODO: Use MVCC to make this set of increments atomic to reads byte [] row = increment.getRow(); checkRow(row, "increment"); TimeRange tr = increment.getTimeRange(); @@ -4725,9 +4725,16 @@ // Lock row startRegionOperation(); this.writeRequestsCount.increment(); + WriteEntry w = null; try { Integer lid = getLock(lockid, row, true); this.updatesLock.readLock().lock(); + // wait for all prior MVCC transactions to finish - while we hold the row lock + // (so that we are guaranteed to see the latest state) + WriteEntry tmp = mvcc.beginMemstoreInsert(); + mvcc.completeMemstoreInsert(tmp); + // now start my own transaction + w = mvcc.beginMemstoreInsert(); try { long now = EnvironmentEdgeManager.currentTimeMillis(); // Process each family @@ -4766,6 +4773,7 @@ // Append new incremented KeyValue to list KeyValue newKV = new KeyValue(row, family.getKey(), column.getKey(), now, Bytes.toBytes(amount)); + newKV.setMemstoreTS(w.getWriteNumber()); kvs.add(newKV); // Prepare WAL updates @@ -4794,7 +4802,7 @@ //Actually write to Memstore now for (Map.Entry> entry : tempMemstore.entrySet()) { Store store = entry.getKey(); - size += store.upsert(entry.getValue()); + size += store.upsert(entry.getValue(), getSmallestReadPoint()); allKVs.addAll(entry.getValue()); } size = this.addAndGetGlobalMemstoreSize(size); @@ -4807,6 +4815,9 @@ syncOrDefer(txid); // sync the transaction log outside the rowlock } } finally { + if (w != null) { + mvcc.completeMemstoreInsert(w); + } closeRegionOperation(); long after = EnvironmentEdgeManager.currentTimeMillis(); this.opMetrics.updateIncrementMetrics(increment.getFamilyMap().keySet(), after - before); Index: hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HStore.java =================================================================== --- hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HStore.java (revision 1403958) +++ hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HStore.java (working copy) @@ -2033,7 +2033,19 @@ return this.region.regionInfo; } - @Override + /** + * Used in tests. TODO: Remove + * + * Updates the value for the given row/family/qualifier. This function will always be seen as + * atomic by other readers because it only puts a single KV to memstore. Thus no read/write + * control necessary. + * @param row row to update + * @param f family to update + * @param qualifier qualifier to update + * @param newValue the new value to set into memstore + * @return memstore size delta + * @throws IOException + */ public long updateColumnValue(byte [] row, byte [] f, byte [] qualifier, long newValue) throws IOException { @@ -2054,11 +2066,10 @@ } @Override - public long upsert(Iterable kvs) throws IOException { + public long upsert(Iterable kvs, long readpoint) throws IOException { this.lock.readLock().lock(); try { - // TODO: Make this operation atomic w/ MVCC - return this.memstore.upsert(kvs); + return this.memstore.upsert(kvs, readpoint); } finally { this.lock.readLock().unlock(); } Index: hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/MemStore.java =================================================================== --- hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/MemStore.java (revision 1403958) +++ hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/MemStore.java (working copy) @@ -435,6 +435,8 @@ } /** + * Only used by tests. TODO: Remove + * * Given the specs of a column, update it, first by inserting a new record, * then removing the old one. Since there is only 1 KeyValue involved, the memstoreTS * will be set to 0, thus ensuring that they instantly appear to anyone. The underlying @@ -449,7 +451,7 @@ * @param now * @return Timestamp */ - public long updateColumnValue(byte[] row, + long updateColumnValue(byte[] row, byte[] family, byte[] qualifier, long newValue, @@ -497,7 +499,7 @@ // create or update (upsert) a new KeyValue with // 'now' and a 0 memstoreTS == immediately visible return upsert(Arrays.asList( - new KeyValue(row, family, qualifier, now, Bytes.toBytes(newValue))) + new KeyValue(row, family, qualifier, now, Bytes.toBytes(newValue))), 1L ); } finally { this.lock.readLock().unlock(); @@ -519,15 +521,15 @@ * atomically. Scans will only see each KeyValue update as atomic. * * @param kvs + * @param readpoint readpoint below which we can safely remove duplicate KVs * @return change in memstore size */ - public long upsert(Iterable kvs) { + public long upsert(Iterable kvs, long readpoint) { this.lock.readLock().lock(); try { long size = 0; for (KeyValue kv : kvs) { - kv.setMemstoreTS(0); - size += upsert(kv); + size += upsert(kv, readpoint); } return size; } finally { @@ -549,7 +551,7 @@ * @param kv * @return change in size of MemStore */ - private long upsert(KeyValue kv) { + private long upsert(KeyValue kv, long readpoint) { // Add the KeyValue to the MemStore // Use the internalAdd method here since we (a) already have a lock // and (b) cannot safely use the MSLAB here without potentially @@ -566,6 +568,7 @@ kv.getBuffer(), kv.getQualifierOffset(), kv.getQualifierLength()); SortedSet ss = kvset.tailSet(firstKv); Iterator it = ss.iterator(); + int versionsOlderThanReadpoint = 0; while ( it.hasNext() ) { KeyValue cur = it.next(); @@ -573,23 +576,23 @@ // ignore the one just put in continue; } - // if this isn't the row we are interested in, then bail - if (!kv.matchingRow(cur)) { - break; - } + // check that this is the row and column we are interested in, otherwise bail + if (kv.matchingRow(cur) && kv.matchingQualifier(cur)) { + // only remove Puts that concurrent scanners cannot possibly see + if (cur.getType() == KeyValue.Type.Put.getCode() && cur.getMemstoreTS() < readpoint) { + if (versionsOlderThanReadpoint > 1) { + // if we get here we have seen at least one version older than the readpoint, + // which means we can prove that no scanner will see this version - // if the qualifier matches and it's a put, remove it - if (kv.matchingQualifier(cur)) { - - // to be extra safe we only remove Puts that have a memstoreTS==0 - if (kv.getType() == KeyValue.Type.Put.getCode() && - kv.getMemstoreTS() == 0) { - // false means there was a change, so give us the size. - addedSize -= heapSizeChange(kv, true); - it.remove(); + // false means there was a change, so give us the size. + addedSize -= heapSizeChange(cur, true); + it.remove(); + } else { + versionsOlderThanReadpoint++; + } } } else { - // past the column, done + // past the row or column, done break; } } Index: hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/Store.java =================================================================== --- hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/Store.java (revision 1403958) +++ hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/Store.java (working copy) @@ -75,20 +75,6 @@ throws IOException; /** - * Updates the value for the given row/family/qualifier. This function will always be seen as - * atomic by other readers because it only puts a single KV to memstore. Thus no read/write - * control necessary. - * @param row row to update - * @param f family to update - * @param qualifier qualifier to update - * @param newValue the new value to set into memstore - * @return memstore size delta - * @throws IOException - */ - public long updateColumnValue(byte[] row, byte[] f, byte[] qualifier, long newValue) - throws IOException; - - /** * Adds or replaces the specified KeyValues. *

* For each KeyValue specified, if a cell with the same row, family, and qualifier exists in @@ -97,10 +83,11 @@ * This operation is atomic on each KeyValue (row/family/qualifier) but not necessarily atomic * across all of them. * @param kvs + * @param readpoint readpoint below which we can safely remove duplicate KVs * @return memstore size delta * @throws IOException */ - public long upsert(Iterable kvs) throws IOException; + public long upsert(Iterable kvs, long readpoint) throws IOException; /** * Adds a value to the memstore Index: hbase-server/src/test/java/org/apache/hadoop/hbase/regionserver/TestAtomicOperation.java =================================================================== --- hbase-server/src/test/java/org/apache/hadoop/hbase/regionserver/TestAtomicOperation.java (revision 1403958) +++ hbase-server/src/test/java/org/apache/hadoop/hbase/regionserver/TestAtomicOperation.java (working copy) @@ -108,7 +108,7 @@ public void testIncrementMultiThreads() throws IOException { LOG.info("Starting test testIncrementMultiThreads"); - initHRegion(tableName, getName(), fam1); + initHRegion(tableName, getName(), fam1, fam2); // create 100 threads, each will increment by its own quantity int numThreads = 100; @@ -135,6 +135,8 @@ } } assertICV(row, fam1, qual1, expectedTotal); + assertICV(row, fam1, qual2, expectedTotal*2); + assertICV(row, fam2, qual3, expectedTotal*3); LOG.info("testIncrementMultiThreads successfully verified that total is " + expectedTotal); } @@ -166,7 +168,9 @@ throws IOException{ HTableDescriptor htd = new HTableDescriptor(tableName); for(byte [] family : families) { - htd.addFamily(new HColumnDescriptor(family)); + HColumnDescriptor hcd = new HColumnDescriptor(family); + hcd.setMaxVersions(1); + htd.addFamily(hcd); } HRegionInfo info = new HRegionInfo(htd.getName(), null, null, false); Path path = new Path(DIR + callingMethod); @@ -184,18 +188,14 @@ public static class Incrementer extends Thread { private final HRegion region; - private final int threadNumber; private final int numIncrements; private final int amount; - private int count; public Incrementer(HRegion region, int threadNumber, int amount, int numIncrements) { this.region = region; - this.threadNumber = threadNumber; this.numIncrements = numIncrements; - this.count = 0; this.amount = amount; setDaemon(true); } @@ -206,16 +206,78 @@ try { Increment inc = new Increment(row); inc.addColumn(fam1, qual1, amount); - Result result = region.increment(inc, null, true); - // LOG.info("thread:" + threadNumber + " iter:" + i); + inc.addColumn(fam1, qual2, amount*2); + inc.addColumn(fam2, qual3, amount*3); + region.increment(inc, null, true); + + // verify: Make sure we only see completed increments + Get g = new Get(row); + Result result = region.get(g, null); + assertEquals(Bytes.toLong(result.getValue(fam1, qual1))*2, Bytes.toLong(result.getValue(fam1, qual2))); + assertEquals(Bytes.toLong(result.getValue(fam1, qual1))*3, Bytes.toLong(result.getValue(fam2, qual3))); } catch (IOException e) { e.printStackTrace(); } - count++; } } } + public void testAppendMultiThreads() throws IOException { + LOG.info("Starting test testAppendMultiThreads"); + initHRegion(tableName, getName(), fam1, fam2); + + int numThreads = 100; + int opsPerThread = 100; + AtomicOperation[] all = new AtomicOperation[numThreads]; + final byte[] val = new byte[]{1}; + + AtomicInteger failures = new AtomicInteger(0); + // create all threads + for (int i = 0; i < numThreads; i++) { + all[i] = new AtomicOperation(region, opsPerThread, null, failures) { + @Override + public void run() { + for (int i=0; i