/** 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);
}
}
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();
}
@Test
public void testSplit() {
// split the range (zero, zero] into:
// (zero,four], (four,six], (six,seven] and (seven, zero]
mt.split(tok(4));
mt.split(tok(6));
mt.split(tok(7));
assertEquals(4, mt.size());
assertEquals(new Range(tok(7), tok(-1)), mt.get(tok(-1)));
assertEquals(new Range(tok(-1), tok(4)), mt.get(tok(3)));
assertEquals(new Range(tok(-1), tok(4)), mt.get(tok(4)));
assertEquals(new Range(tok(4), tok(6)), mt.get(tok(6)));
assertEquals(new Range(tok(6), tok(7)), mt.get(tok(7)));
// check depths
assertEquals((byte) 1, mt.get(tok(4)).depth);
assertEquals((byte) 2, mt.get(tok(6)).depth);
assertEquals((byte) 3, mt.get(tok(7)).depth);
assertEquals((byte) 3, mt.get(tok(-1)).depth);
try {
mt.split(tok(-1));
fail("Shouldn't be able to split outside the initial range.");
} catch (AssertionError e) {
// pass
}
}
/**
* @param ltree First tree.
* @param rtree Second tree.
* @return A list of the largest contiguous ranges where the given trees disagree.
*/
public static List<TreeRange> difference(MerkleTree ltree, MerkleTree rtree) {
List<TreeRange> diff = new ArrayList<TreeRange>();
Token mintoken = ltree.partitioner.getMinimumToken();
TreeRange active = new TreeRange(null, mintoken, mintoken, (byte) 0, null);
byte[] lhash = ltree.hash(active);
byte[] rhash = rtree.hash(active);
if (lhash != null && rhash != null && !Arrays.equals(lhash, rhash)) {
if (FULLY_INCONSISTENT == differenceHelper(ltree, rtree, diff, active)) diff.add(active);
} else if (lhash == null || rhash == null) diff.add(active);
return diff;
}
@Test
public void testHashFull() {
byte[] val = DUMMY;
Range range = new Range(tok(-1), tok(-1));
// (zero, zero]
assertNull(mt.hash(range));
// validate the range
mt.get(tok(-1)).hash(val);
assertHashEquals(val, mt.hash(range));
}
@Test
public void testCompactHash() {
byte[] val = DUMMY;
byte[] valXval = hashed(val, 1, 1);
// (zero, four], (four,zero]
mt.split(tok(4));
// validate both ranges
mt.get(tok(4)).hash(val);
mt.get(tok(-1)).hash(val);
// compact (zero, eight]
mt.compact(tok(4));
assertHashEquals(valXval, mt.get(tok(-1)).hash());
}
@Test
public void testHashRandom() {
int max = 1000000;
TOKEN_SCALE = new BigInteger("" + max);
mt = new MerkleTree(partitioner, RECOMMENDED_DEPTH, 32);
Random random = new Random();
while (true) {
if (!mt.split(tok(random.nextInt(max)))) break;
}
// validate the tree
TreeRangeIterator ranges = mt.invalids(new Range(tok(-1), tok(-1)));
for (TreeRange range : ranges) range.addHash(new RowHash(range.right, new byte[0]));
assert null != mt.hash(new Range(tok(-1), tok(-1))) : "Could not hash tree " + mt;
}
TreeRangeIterator(MerkleTree tree, Range range) {
Token mintoken = tree.partitioner().getMinimumToken();
tovisit = new ArrayDeque<TreeRange>();
tovisit.add(new TreeRange(tree, mintoken, mintoken, (byte) 0, tree.root));
this.tree = tree;
this.range = range;
}
/**
* TODO: This function could be optimized into a depth first traversal of the two trees in
* parallel.
*
* <p>Takes two trees and a range for which they have hashes, but are inconsistent.
*
* @return FULLY_INCONSISTENT if active is inconsistent, PARTIALLY_INCONSISTENT if only a subrange
* is inconsistent.
*/
static int differenceHelper(
MerkleTree ltree, MerkleTree rtree, List<TreeRange> diff, TreeRange active) {
Token midpoint = ltree.partitioner().midpoint(active.left, active.right);
TreeRange left = new TreeRange(null, active.left, midpoint, inc(active.depth), null);
TreeRange right = new TreeRange(null, midpoint, active.right, inc(active.depth), null);
byte[] lhash;
byte[] rhash;
// see if we should recurse left
lhash = ltree.hash(left);
rhash = rtree.hash(left);
int ldiff = CONSISTENT;
boolean lreso = lhash != null && rhash != null;
if (lreso && !Arrays.equals(lhash, rhash)) ldiff = differenceHelper(ltree, rtree, diff, left);
else if (!lreso) ldiff = FULLY_INCONSISTENT;
// see if we should recurse right
lhash = ltree.hash(right);
rhash = rtree.hash(right);
int rdiff = CONSISTENT;
boolean rreso = lhash != null && rhash != null;
if (rreso && !Arrays.equals(lhash, rhash)) rdiff = differenceHelper(ltree, rtree, diff, right);
else if (!rreso) rdiff = FULLY_INCONSISTENT;
if (ldiff == FULLY_INCONSISTENT && rdiff == FULLY_INCONSISTENT) {
// both children are fully inconsistent
return FULLY_INCONSISTENT;
} else if (ldiff == FULLY_INCONSISTENT) {
diff.add(left);
return PARTIALLY_INCONSISTENT;
} else if (rdiff == FULLY_INCONSISTENT) {
diff.add(right);
return PARTIALLY_INCONSISTENT;
}
return PARTIALLY_INCONSISTENT;
}
@Test
public void testSplitLimitDepth() {
mt = new MerkleTree(partitioner, (byte) 2, Integer.MAX_VALUE);
assertTrue(mt.split(tok(4)));
assertTrue(mt.split(tok(2)));
assertEquals(3, mt.size());
// should fail to split below hashdepth
assertFalse(mt.split(tok(1)));
assertEquals(3, mt.size());
assertEquals(new Range(tok(4), tok(-1)), mt.get(tok(-1)));
assertEquals(new Range(tok(-1), tok(2)), mt.get(tok(2)));
assertEquals(new Range(tok(2), tok(4)), mt.get(tok(4)));
}
@Test
public void testSplitLimitSize() {
mt = new MerkleTree(partitioner, RECOMMENDED_DEPTH, 2);
assertTrue(mt.split(tok(4)));
assertEquals(2, mt.size());
// should fail to split above maxsize
assertFalse(mt.split(tok(2)));
assertEquals(2, mt.size());
assertEquals(new Range(tok(4), tok(-1)), mt.get(tok(-1)));
assertEquals(new Range(tok(-1), tok(4)), mt.get(tok(4)));
}
/** Compares our trees, and triggers repairs for any ranges that mismatch. */
public void run() {
// restore partitioners (in case we were serialized)
if (r1.tree.partitioner() == null) r1.tree.partitioner(StorageService.getPartitioner());
if (r2.tree.partitioner() == null) r2.tree.partitioner(StorageService.getPartitioner());
// compare trees, and collect differences
differences.addAll(MerkleTree.difference(r1.tree, r2.tree));
// choose a repair method based on the significance of the difference
String format =
String.format(
"[repair #%s] Endpoints %s and %s %%s for %s",
getName(), r1.endpoint, r2.endpoint, cfname);
if (differences.isEmpty()) {
logger.info(String.format(format, "are consistent"));
completed(this);
return;
}
// non-0 difference: perform streaming repair
logger.info(String.format(format, "have " + differences.size() + " range(s) out of sync"));
performStreamingRepair();
}
@Test
public void testSerialization() throws Exception {
Range full = new Range(tok(-1), tok(-1));
ByteArrayOutputStream bout = new ByteArrayOutputStream();
ObjectOutputStream oout = new ObjectOutputStream(bout);
// populate and validate the tree
mt.maxsize(256);
mt.init();
for (TreeRange range : mt.invalids(full)) range.addAll(new HIterator(range.right));
byte[] initialhash = mt.hash(full);
oout.writeObject(mt);
oout.close();
ByteArrayInputStream bin = new ByteArrayInputStream(bout.toByteArray());
ObjectInputStream oin = new ObjectInputStream(bin);
MerkleTree restored = (MerkleTree) oin.readObject();
// restore partitioner after serialization
restored.partitioner(partitioner);
assertHashEquals(initialhash, restored.hash(full));
}
@Test
public void testInvalids() {
Iterator<TreeRange> ranges;
// (zero, zero]
ranges = mt.invalids(new Range(tok(-1), tok(-1)));
assertEquals(new Range(tok(-1), tok(-1)), ranges.next());
assertFalse(ranges.hasNext());
// all invalid
mt.split(tok(4));
mt.split(tok(2));
mt.split(tok(6));
mt.split(tok(3));
mt.split(tok(5));
ranges = mt.invalids(new Range(tok(-1), tok(-1)));
assertEquals(new Range(tok(-1), tok(2)), ranges.next());
assertEquals(new Range(tok(2), tok(3)), ranges.next());
assertEquals(new Range(tok(3), tok(4)), ranges.next());
assertEquals(new Range(tok(4), tok(5)), ranges.next());
assertEquals(new Range(tok(5), tok(6)), ranges.next());
assertEquals(new Range(tok(6), tok(-1)), ranges.next());
assertFalse(ranges.hasNext());
// some invalid
mt.get(tok(2)).hash("non-null!".getBytes());
mt.get(tok(4)).hash("non-null!".getBytes());
mt.get(tok(5)).hash("non-null!".getBytes());
mt.get(tok(-1)).hash("non-null!".getBytes());
ranges = mt.invalids(new Range(tok(-1), tok(-1)));
assertEquals(new Range(tok(2), tok(3)), ranges.next());
assertEquals(new Range(tok(5), tok(6)), ranges.next());
assertFalse(ranges.hasNext());
// some invalid in left subrange
ranges = mt.invalids(new Range(tok(-1), tok(6)));
assertEquals(new Range(tok(2), tok(3)), ranges.next());
assertEquals(new Range(tok(5), tok(6)), ranges.next());
assertFalse(ranges.hasNext());
// some invalid in right subrange
ranges = mt.invalids(new Range(tok(2), tok(-1)));
assertEquals(new Range(tok(2), tok(3)), ranges.next());
assertEquals(new Range(tok(5), tok(6)), ranges.next());
assertFalse(ranges.hasNext());
}
@Test
public void testDifference() {
Range full = new Range(tok(-1), tok(-1));
int maxsize = 16;
mt.maxsize(maxsize);
MerkleTree mt2 = new MerkleTree(partitioner, RECOMMENDED_DEPTH, maxsize);
mt.init();
mt2.init();
TreeRange leftmost = null;
TreeRange middle = null;
TreeRange rightmost = null;
// compact the leftmost, and split the rightmost
Iterator<TreeRange> ranges = mt.invalids(full);
leftmost = ranges.next();
rightmost = null;
while (ranges.hasNext()) rightmost = ranges.next();
mt.compact(leftmost.right);
leftmost = mt.get(leftmost.right); // leftmost is now a larger range
mt.split(rightmost.right);
// set the hash for the left neighbor of rightmost
middle = mt.get(rightmost.left);
middle.hash("arbitrary!".getBytes());
byte depth = middle.depth;
// add dummy hashes to the rest of both trees
for (TreeRange range : mt.invalids(full)) range.addAll(new HIterator(range.right));
for (TreeRange range : mt2.invalids(full)) range.addAll(new HIterator(range.right));
// trees should disagree for leftmost, (middle.left, rightmost.right]
List<TreeRange> diffs = MerkleTree.difference(mt, mt2);
assertEquals(diffs + " contains wrong number of differences:", 2, diffs.size());
assertTrue(diffs.contains(leftmost));
assertTrue(diffs.contains(new Range(middle.left, rightmost.right)));
}
/**
* Generate two trees with different splits, but containing the same keys, and check that they
* compare equally.
*
* <p>The set of keys used in this test is: #{2,4,6,8,12,14,0}
*/
@Test
public void testValidateTree() {
TOKEN_SCALE = new BigInteger("16"); // this test needs slightly more resolution
Range full = new Range(tok(-1), tok(-1));
Iterator<TreeRange> ranges;
MerkleTree mt2 = new MerkleTree(partitioner, RECOMMENDED_DEPTH, Integer.MAX_VALUE);
mt.split(tok(8));
mt.split(tok(4));
mt.split(tok(12));
mt.split(tok(6));
mt.split(tok(10));
ranges = mt.invalids(full);
ranges.next().addAll(new HIterator(2, 4)); // (-1,4]: depth 2
ranges.next().addAll(new HIterator(6)); // (4,6]
ranges.next().addAll(new HIterator(8)); // (6,8]
ranges.next().addAll(new HIterator(/*empty*/ new int[0])); // (8,10]
ranges.next().addAll(new HIterator(12)); // (10,12]
ranges.next().addAll(new HIterator(14, -1)); // (12,-1]: depth 2
mt2.split(tok(8));
mt2.split(tok(4));
mt2.split(tok(12));
mt2.split(tok(2));
mt2.split(tok(10));
mt2.split(tok(9));
mt2.split(tok(11));
ranges = mt2.invalids(full);
ranges.next().addAll(new HIterator(2)); // (-1,2]
ranges.next().addAll(new HIterator(4)); // (2,4]
ranges.next().addAll(new HIterator(6, 8)); // (4,8]: depth 2
ranges.next().addAll(new HIterator(/*empty*/ new int[0])); // (8,9]
ranges.next().addAll(new HIterator(/*empty*/ new int[0])); // (9,10]
ranges.next().addAll(new HIterator(/*empty*/ new int[0])); // (10,11]: depth 4
ranges.next().addAll(new HIterator(12)); // (11,12]: depth 4
ranges.next().addAll(new HIterator(14, -1)); // (12,-1]: depth 2
byte[] mthash = mt.hash(full);
byte[] mt2hash = mt2.hash(full);
assertHashEquals("Tree hashes did not match: " + mt + " && " + mt2, mthash, mt2hash);
}
@Test
public void testHashDegenerate() {
TOKEN_SCALE = new BigInteger("32");
byte[] val = DUMMY;
byte[] childfullval = hashed(val, 5, 5, 4);
byte[] fullval = hashed(val, 5, 5, 4, 3, 2, 1);
Range childfull = new Range(tok(-1), tok(4));
Range full = new Range(tok(-1), tok(-1));
Range invalid = new Range(tok(4), tok(-1));
mt = new MerkleTree(partitioner, RECOMMENDED_DEPTH, Integer.MAX_VALUE);
mt.split(tok(16));
mt.split(tok(8));
mt.split(tok(4));
mt.split(tok(2));
mt.split(tok(1));
assertNull(mt.hash(full));
assertNull(mt.hash(childfull));
assertNull(mt.hash(invalid));
// validate the range
mt.get(tok(1)).hash(val);
mt.get(tok(2)).hash(val);
mt.get(tok(4)).hash(val);
mt.get(tok(8)).hash(val);
mt.get(tok(16)).hash(val);
mt.get(tok(-1)).hash(val);
assertHashEquals(fullval, mt.hash(full));
assertHashEquals(childfullval, mt.hash(childfull));
assertNull(mt.hash(invalid));
}
@Test
public void testHashInner() {
byte[] val = DUMMY;
byte[] lchildval = hashed(val, 3, 3, 2);
byte[] rchildval = hashed(val, 2, 2);
byte[] fullval = hashed(val, 3, 3, 2, 2, 2);
Range full = new Range(tok(-1), tok(-1));
Range lchild = new Range(tok(-1), tok(4));
Range rchild = new Range(tok(4), tok(-1));
Range invalid = new Range(tok(1), tok(-1));
// (zero,one] (one, two] (two,four] (four, six] (six, zero]
mt.split(tok(4));
mt.split(tok(2));
mt.split(tok(6));
mt.split(tok(1));
assertNull(mt.hash(full));
assertNull(mt.hash(lchild));
assertNull(mt.hash(rchild));
assertNull(mt.hash(invalid));
// validate the range
mt.get(tok(1)).hash(val);
mt.get(tok(2)).hash(val);
mt.get(tok(4)).hash(val);
mt.get(tok(6)).hash(val);
mt.get(tok(-1)).hash(val);
assertHashEquals(fullval, mt.hash(full));
assertHashEquals(lchildval, mt.hash(lchild));
assertHashEquals(rchildval, mt.hash(rchild));
assertNull(mt.hash(invalid));
}
@Test
public void testHashPartial() {
byte[] val = DUMMY;
byte[] leftval = hashed(val, 1, 1);
byte[] partialval = hashed(val, 1);
Range left = new Range(tok(-1), tok(4));
Range partial = new Range(tok(2), tok(4));
Range right = new Range(tok(4), tok(-1));
Range linvalid = new Range(tok(1), tok(4));
Range rinvalid = new Range(tok(4), tok(6));
// (zero,two] (two,four] (four, zero]
mt.split(tok(4));
mt.split(tok(2));
assertNull(mt.hash(left));
assertNull(mt.hash(partial));
assertNull(mt.hash(right));
assertNull(mt.hash(linvalid));
assertNull(mt.hash(rinvalid));
// validate the range
mt.get(tok(2)).hash(val);
mt.get(tok(4)).hash(val);
mt.get(tok(-1)).hash(val);
assertHashEquals(leftval, mt.hash(left));
assertHashEquals(partialval, mt.hash(partial));
assertHashEquals(val, mt.hash(right));
assertNull(mt.hash(linvalid));
assertNull(mt.hash(rinvalid));
}
@Test
public void testCompact() {
// (zero, one], (one,two], ... (seven, zero]
mt.split(tok(4));
mt.split(tok(2));
mt.split(tok(6));
mt.split(tok(1));
mt.split(tok(3));
mt.split(tok(5));
mt.split(tok(7));
// compact (zero,two] and then (four,six]
mt.compact(tok(1));
mt.compact(tok(5));
assertEquals(6, mt.size());
assertEquals(new Range(tok(-1), tok(2)), mt.get(tok(2)));
assertEquals(new Range(tok(2), tok(3)), mt.get(tok(3)));
assertEquals(new Range(tok(3), tok(4)), mt.get(tok(4)));
assertEquals(new Range(tok(4), tok(6)), mt.get(tok(5)));
assertEquals(new Range(tok(6), tok(7)), mt.get(tok(7)));
assertEquals(new Range(tok(7), tok(-1)), mt.get(tok(-1)));
// compacted ranges should be at depth 2, and the rest at 3
for (int i : new int[] {2, 6}) {
assertEquals((byte) 2, mt.get(tok(i)).depth);
}
for (int i : new int[] {3, 4, 7, -1}) {
assertEquals((byte) 3, mt.get(tok(i)).depth);
}
// compact (two,four] and then (six,zero]
mt.compact(tok(3));
mt.compact(tok(7));
assertEquals(4, mt.size());
assertEquals(new Range(tok(-1), tok(2)), mt.get(tok(2)));
assertEquals(new Range(tok(2), tok(4)), mt.get(tok(4)));
assertEquals(new Range(tok(4), tok(6)), mt.get(tok(5)));
assertEquals(new Range(tok(6), tok(-1)), mt.get(tok(-1)));
for (int i : new int[] {2, 4, 5, -1}) {
assertEquals((byte) 2, mt.get(tok(i)).depth);
}
// compact (zero,four]
mt.compact(tok(2));
assertEquals(3, mt.size());
assertEquals(new Range(tok(-1), tok(4)), mt.get(tok(2)));
assertEquals(new Range(tok(4), tok(6)), mt.get(tok(6)));
assertEquals(new Range(tok(6), tok(-1)), mt.get(tok(-1)));
// compact (four, zero]
mt.compact(tok(6));
assertEquals(2, mt.size());
assertEquals(new Range(tok(-1), tok(4)), mt.get(tok(2)));
assertEquals(new Range(tok(4), tok(-1)), mt.get(tok(6)));
assertEquals((byte) 1, mt.get(tok(2)).depth);
assertEquals((byte) 1, mt.get(tok(6)).depth);
// compact (zero, zero] (the root)
mt.compact(tok(4));
assertEquals(1, mt.size());
assertEquals(new Range(tok(-1), tok(-1)), mt.get(tok(-1)));
assertEquals((byte) 0, mt.get(tok(-1)).depth);
}