@Test
public void testEntrySink() throws Exception {
Configuration conf = new Configuration();
RecoveryMode mode =
(conf.getBoolean(HConstants.DISTRIBUTED_LOG_REPLAY_KEY, false)
? RecoveryMode.LOG_REPLAY
: RecoveryMode.LOG_SPLITTING);
EntryBuffers sink = new EntryBuffers(new PipelineController(), 1 * 1024 * 1024);
for (int i = 0; i < 1000; i++) {
WAL.Entry entry = createTestLogEntry(i);
sink.appendEntry(entry);
}
assertTrue(sink.totalBuffered > 0);
long amountInChunk = sink.totalBuffered;
// Get a chunk
RegionEntryBuffer chunk = sink.getChunkToWrite();
assertEquals(chunk.heapSize(), amountInChunk);
// Make sure it got marked that a thread is "working on this"
assertTrue(sink.isRegionCurrentlyWriting(TEST_REGION));
// Insert some more entries
for (int i = 0; i < 500; i++) {
WAL.Entry entry = createTestLogEntry(i);
sink.appendEntry(entry);
}
// Asking for another chunk shouldn't work since the first one
// is still writing
assertNull(sink.getChunkToWrite());
// If we say we're done writing the first chunk, then we should be able
// to get the second
sink.doneWriting(chunk);
RegionEntryBuffer chunk2 = sink.getChunkToWrite();
assertNotNull(chunk2);
assertNotSame(chunk, chunk2);
long amountInChunk2 = sink.totalBuffered;
// The second chunk had fewer rows than the first
assertTrue(amountInChunk2 < amountInChunk);
sink.doneWriting(chunk2);
assertEquals(0, sink.totalBuffered);
}
@Override
public boolean replicate(ReplicateContext replicateContext) {
/* A note on batching in RegionReplicaReplicationEndpoint (RRRE):
*
* RRRE relies on batching from two different mechanisms. The first is the batching from
* ReplicationSource since RRRE is a ReplicationEndpoint driven by RS. RS reads from a single
* WAL file filling up a buffer of heap size "replication.source.size.capacity"(64MB) or at most
* "replication.source.nb.capacity" entries or until it sees the end of file (in live tailing).
* Then RS passes all the buffered edits in this replicate() call context. RRRE puts the edits
* to the WALSplitter.EntryBuffers which is a blocking buffer space of up to
* "hbase.region.replica.replication.buffersize" (128MB) in size. This buffer splits the edits
* based on regions.
*
* There are "hbase.region.replica.replication.writer.threads"(default 3) writer threads which
* pick largest per-region buffer and send it to the SinkWriter (see RegionReplicaOutputSink).
* The SinkWriter in this case will send the wal edits to all secondary region replicas in
* parallel via a retrying rpc call. EntryBuffers guarantees that while a buffer is
* being written to the sink, another buffer for the same region will not be made available to
* writers ensuring regions edits are not replayed out of order.
*
* The replicate() call won't return until all the buffers are sent and ack'd by the sinks so
* that the replication can assume all edits are persisted. We may be able to do a better
* pipelining between the replication thread and output sinks later if it becomes a bottleneck.
*/
while (this.isRunning()) {
try {
for (Entry entry : replicateContext.getEntries()) {
entryBuffers.appendEntry(entry);
}
outputSink.flush(); // make sure everything is flushed
ctx.getMetrics().incrLogEditsFiltered(outputSink.getSkippedEditsCounter().getAndSet(0));
return true;
} catch (InterruptedException e) {
Thread.currentThread().interrupt();
return false;
} catch (IOException e) {
LOG.warn(
"Received IOException while trying to replicate" + StringUtils.stringifyException(e));
}
}
return false;
}
