HBASE-18116 Replication source in-memory accounting should not include bulk transfer hfiles

Signed-off-by: Andrew Purtell <apurtell@apache.org>
This commit is contained in:
Xu Cang 2018-05-31 20:00:04 -07:00 committed by Andrew Purtell
parent 0968668283
commit a11701ecc5
4 changed files with 45 additions and 17 deletions

View File

@ -598,6 +598,7 @@ public class ReplicationSource implements ReplicationSourceInterface {
}
@Override
//offsets totalBufferUsed by deducting shipped batchSize.
public void postShipEdits(List<Entry> entries, int batchSize) {
if (throttler.isEnabled()) {
throttler.addPushSize(batchSize);

View File

@ -123,6 +123,18 @@ public class ReplicationSourceShipper extends Thread {
protected void postFinish() {
}
/**
* get batchEntry size excludes bulk load file sizes.
* Uses ReplicationSourceWALReader's static method.
*/
private int getBatchEntrySizeExcludeBulkLoad(WALEntryBatch entryBatch) {
int totalSize = 0;
for(Entry entry : entryBatch.getWalEntries()) {
totalSize += entryReader.getEntrySizeExcludeBulkLoad(entry);
}
return totalSize;
}
/**
* Do the shipping logic
*/
@ -139,6 +151,7 @@ public class ReplicationSourceShipper extends Thread {
return;
}
int currentSize = (int) entryBatch.getHeapSize();
int sizeExcludeBulkLoad = getBatchEntrySizeExcludeBulkLoad(entryBatch);
while (isActive()) {
try {
try {
@ -175,7 +188,11 @@ public class ReplicationSourceShipper extends Thread {
// Log and clean up WAL logs
updateLogPosition(entryBatch);
source.postShipEdits(entries, currentSize);
//offsets totalBufferUsed by deducting shipped batchSize (excludes bulk load size)
//this sizeExcludeBulkLoad has to use same calculation that when calling
//acquireBufferQuota() in ReplicatinoSourceWALReader because they maintain
//same variable: totalBufferUsed
source.postShipEdits(entries, sizeExcludeBulkLoad);
// FIXME check relationship between wal group and overall
source.getSourceMetrics().shipBatch(entryBatch.getNbOperations(), currentSize,
entryBatch.getNbHFiles());

View File

@ -36,6 +36,7 @@ import org.apache.hadoop.hbase.util.Pair;
import org.apache.hadoop.hbase.util.Threads;
import org.apache.hadoop.hbase.wal.WAL.Entry;
import org.apache.hadoop.hbase.wal.WALEdit;
import org.apache.hadoop.hbase.wal.WALKey;
import org.apache.yetus.audience.InterfaceAudience;
import org.apache.yetus.audience.InterfaceStability;
import org.slf4j.Logger;
@ -168,10 +169,12 @@ class ReplicationSourceWALReader extends Thread {
if (edit == null || edit.isEmpty()) {
return false;
}
long entrySize = getEntrySize(entry);
long entrySize = getEntrySizeIncludeBulkLoad(entry);
long entrySizeExcludeBulkLoad = getEntrySizeExcludeBulkLoad(entry);
batch.addEntry(entry);
updateBatchStats(batch, entry, entrySize);
boolean totalBufferTooLarge = acquireBufferQuota(entrySize);
boolean totalBufferTooLarge = acquireBufferQuota(entrySizeExcludeBulkLoad);
// Stop if too many entries or too big
return totalBufferTooLarge || batch.getHeapSize() >= replicationBatchSizeCapacity ||
batch.getNbEntries() >= replicationBatchCountCapacity;
@ -296,11 +299,20 @@ class ReplicationSourceWALReader extends Thread {
return entryBatchQueue.take();
}
private long getEntrySize(Entry entry) {
private long getEntrySizeIncludeBulkLoad(Entry entry) {
WALEdit edit = entry.getEdit();
return edit.heapSize() + calculateTotalSizeOfStoreFiles(edit);
WALKey key = entry.getKey();
return edit.heapSize() + sizeOfStoreFilesIncludeBulkLoad(edit) +
key.estimatedSerializedSizeOf();
}
public static long getEntrySizeExcludeBulkLoad(Entry entry) {
WALEdit edit = entry.getEdit();
WALKey key = entry.getKey();
return edit.heapSize() + key.estimatedSerializedSizeOf();
}
private void updateBatchStats(WALEntryBatch batch, Entry entry, long entrySize) {
WALEdit edit = entry.getEdit();
batch.incrementHeapSize(entrySize);
@ -353,7 +365,7 @@ class ReplicationSourceWALReader extends Thread {
* @param edit edit to count row keys from
* @return the total size of the store files
*/
private int calculateTotalSizeOfStoreFiles(WALEdit edit) {
private int sizeOfStoreFilesIncludeBulkLoad(WALEdit edit) {
List<Cell> cells = edit.getCells();
int totalStoreFilesSize = 0;

View File

@ -64,6 +64,8 @@ public class TestGlobalThrottler {
HBaseClassTestRule.forClass(TestGlobalThrottler.class);
private static final Logger LOG = LoggerFactory.getLogger(TestGlobalThrottler.class);
private static final int REPLICATION_SOURCE_QUOTA = 200;
private static int numOfPeer = 0;
private static Configuration conf1;
private static Configuration conf2;
@ -84,7 +86,7 @@ public class TestGlobalThrottler {
conf1.set(HConstants.ZOOKEEPER_ZNODE_PARENT, "/1");
conf1.setLong("replication.source.sleepforretries", 100);
// Each WAL is about 120 bytes
conf1.setInt(HConstants.REPLICATION_SOURCE_TOTAL_BUFFER_KEY, 200);
conf1.setInt(HConstants.REPLICATION_SOURCE_TOTAL_BUFFER_KEY, REPLICATION_SOURCE_QUOTA);
conf1.setLong("replication.source.per.peer.node.bandwidth", 100L);
utility1 = new HBaseTestingUtility(conf1);
@ -109,6 +111,7 @@ public class TestGlobalThrottler {
admin1.addPeer("peer1", rpc, null);
admin1.addPeer("peer2", rpc, null);
admin1.addPeer("peer3", rpc, null);
numOfPeer = admin1.getPeersCount();
}
@AfterClass
@ -140,7 +143,10 @@ public class TestGlobalThrottler {
if (size > 0) {
testQuotaNonZero = true;
}
if (size > 600) {
//the reason here doing "numOfPeer + 1" is because by using method addEntryToBatch(), even the
// batch size (after added last entry) exceeds quota, it still keeps the last one in the batch
// so total used buffer size can be one "replication.total.buffer.quota" larger than expected
if (size > REPLICATION_SOURCE_QUOTA * (numOfPeer + 1)) {
// We read logs first then check throttler, so if the buffer quota limiter doesn't
// take effect, it will push many logs and exceed the quota.
testQuotaPass = false;
@ -181,13 +187,5 @@ public class TestGlobalThrottler {
Assert.assertTrue(testQuotaNonZero);
}
private List<Integer> getRowNumbers(List<Cell> cells) {
List<Integer> listOfRowNumbers = new ArrayList<>(cells.size());
for (Cell c : cells) {
listOfRowNumbers.add(Integer.parseInt(Bytes
.toString(c.getRowArray(), c.getRowOffset() + ROW.length,
c.getRowLength() - ROW.length)));
}
return listOfRowNumbers;
}
}