HDFS-14476. lock too long when fix inconsistent blocks between disk and in-memory. Contributed by Sean Chow.

This commit is contained in:
Wei-Chiu Chuang 2019-08-19 08:38:06 -07:00
parent b3c14d4132
commit 8b802d6b44
1 changed files with 14 additions and 0 deletions

View File

@ -68,6 +68,7 @@ public class DirectoryScanner implements Runnable {
+ " starting at %s with interval of %dms";
private static final String START_MESSAGE_WITH_THROTTLE = START_MESSAGE
+ " and throttle limit of %dms/s";
private static final int RECONCILE_BLOCKS_BATCH_SIZE = 1000;
private final FsDatasetSpi<?> dataset;
private final ExecutorService reportCompileThreadPool;
@ -373,7 +374,11 @@ public class DirectoryScanner implements Runnable {
*/
@VisibleForTesting
public void reconcile() throws IOException {
LOG.debug("reconcile start DirectoryScanning");
scan();
// HDFS-14476: run checkAndUpadte with batch to avoid holding the lock too
// long
int loopCount = 0;
for (Entry<String, LinkedList<ScanInfo>> entry : diffs.entrySet()) {
String bpid = entry.getKey();
LinkedList<ScanInfo> diff = entry.getValue();
@ -381,6 +386,15 @@ public class DirectoryScanner implements Runnable {
for (ScanInfo info : diff) {
dataset.checkAndUpdate(bpid, info);
}
if (loopCount % RECONCILE_BLOCKS_BATCH_SIZE == 0) {
try {
Thread.sleep(2000);
} catch (InterruptedException e) {
// do nothing
}
}
loopCount++;
}
if (!retainDiffs) clear();
}