public void prepare(ColumnFamilyStore cfs) {
if (tree.partitioner() instanceof RandomPartitioner) {
// You can't beat an even tree distribution for md5
tree.init();
} else {
List<DecoratedKey> keys = new ArrayList<DecoratedKey>();
for (DecoratedKey sample : cfs.keySamples(request.range)) {
assert request.range.contains(sample.token)
: "Token " + sample.token + " is not within range " + request.range;
keys.add(sample);
}
if (keys.isEmpty()) {
// use an even tree distribution
tree.init();
} else {
int numkeys = keys.size();
Random random = new Random();
// sample the column family using random keys from the index
while (true) {
DecoratedKey dk = keys.get(random.nextInt(numkeys));
if (!tree.split(dk.token)) break;
}
}
}
logger.debug("Prepared AEService tree of size " + tree.size() + " for " + request);
ranges = tree.invalids();
}
/** Compares our trees, and triggers repairs for any ranges that mismatch. */
public void run() {
InetAddress local = FBUtilities.getLocalAddress();
// restore partitioners (in case we were serialized)
if (ltree.partitioner() == null) ltree.partitioner(StorageService.getPartitioner());
if (rtree.partitioner() == null) rtree.partitioner(StorageService.getPartitioner());
// compare trees, and collect differences
differences.addAll(MerkleTree.difference(ltree, rtree));
// choose a repair method based on the significance of the difference
String format =
"Endpoints " + local + " and " + remote + " %s for " + cfname + " on " + range;
if (differences.isEmpty()) {
logger.info(String.format(format, "are consistent"));
completed(remote, cfname);
return;
}
// non-0 difference: perform streaming repair
logger.info(String.format(format, "have " + differences.size() + " range(s) out of sync"));
try {
performStreamingRepair();
} catch (IOException e) {
throw new RuntimeException(e);
}
}
/**
* Submit differencers for running. All tree *must* have been received before this is called.
*/
public void submitDifferencers() {
// We need to difference all trees one against another
for (int i = 0; i < trees.size() - 1; ++i) {
TreeResponse r1 = trees.get(i);
for (int j = i + 1; j < trees.size(); ++j) {
TreeResponse r2 = trees.get(j);
Differencer differencer = new Differencer(cfname, r1, r2);
logger.debug("Queueing comparison {}", differencer);
differencers.add(differencer);
}
}
differencers.start();
trees.clear(); // allows gc to do its thing
}
/**
* Starts sending/receiving our list of differences to/from the remote endpoint: creates a
* callback that will be called out of band once the streams complete.
*/
void performStreamingRepair() throws IOException {
logger.info(
"Performing streaming repair of "
+ differences.size()
+ " ranges with "
+ remote
+ " for "
+ range);
ColumnFamilyStore cfstore = Table.open(tablename).getColumnFamilyStore(cfname);
try {
Collection<SSTableReader> sstables = cfstore.getSSTables();
Callback callback = new Callback();
// send ranges to the remote node
StreamOutSession outsession = StreamOutSession.create(tablename, remote, callback);
StreamOut.transferSSTables(outsession, sstables, differences, OperationType.AES);
// request ranges from the remote node
StreamIn.requestRanges(remote, tablename, differences, callback, OperationType.AES);
} catch (Exception e) {
throw new IOException("Streaming repair failed.", e);
}
}
