@Override
public int compare(BytesRef left, BytesRef right) {
// Make shallow copy in case decode changes the BytesRef:
leftScratch.bytes = left.bytes;
leftScratch.offset = left.offset;
leftScratch.length = left.length;
rightScratch.bytes = right.bytes;
rightScratch.offset = right.offset;
rightScratch.length = right.length;
long leftCost = decode(leftScratch, input);
long rightCost = decode(rightScratch, input);
if (hasPayloads) {
decodePayload(leftScratch, input);
decodePayload(rightScratch, input);
}
if (hasContexts) {
decodeContexts(leftScratch, input);
decodeContexts(rightScratch, input);
}
int cmp = comparator.compare(leftScratch, rightScratch);
if (cmp != 0) {
return cmp;
}
return Long.compare(leftCost, rightCost);
}
@Override
public BytesRef writeToBytes() {
long start = System.nanoTime();
int size = set.size();
BytesRef bytes = new BytesRef(new byte[HEADER_SIZE + (int) bytesUsed.get()]);
// Encode encoding type
Bytes.writeInt(bytes, this.getEncoding().ordinal());
// Encode flag
bytes.bytes[bytes.offset++] = (byte) (this.isPruned() ? 1 : 0);
// Encode size of the set
Bytes.writeInt(bytes, size);
// Encode longs
BytesRef reusable = new BytesRef();
for (int i = 0; i < this.set.size(); i++) {
this.set.get(i, reusable);
Bytes.writeBytesRef(reusable, bytes);
}
logger.debug(
"Serialized {} terms - took {} ms", this.size(), (System.nanoTime() - start) / 1000000);
bytes.length = bytes.offset;
bytes.offset = 0;
return bytes;
}
/**
* Ensure we own term.bytes so that it's safe to modify. We detect via a kluge in which
* cellsByLevel[0].termBuf is non-null, which is a pre-allocated for use to replace term.bytes.
*/
void ensureOwnTermBytes() {
NRCell cell0 = cellsByLevel[0];
if (cell0.termBuf == null) return; // we already own the bytes
System.arraycopy(term.bytes, term.offset, cell0.termBuf, 0, term.length);
term.bytes = cell0.termBuf;
term.offset = 0;
cell0.termBuf = null;
}
protected static void copy(BytesRef from, BytesRef to) {
if (to.bytes.length < from.length) {
to.bytes = new byte[ArrayUtil.oversize(from.length, RamUsageEstimator.NUM_BYTES_BYTE)];
}
to.offset = 0;
to.length = from.length;
System.arraycopy(from.bytes, from.offset, to.bytes, 0, from.length);
}
@Override
public BytesRef getTokenBytesNoLeaf(BytesRef result) {
if (result == null) result = new BytesRef();
result.bytes = term.bytes;
result.offset = term.offset;
result.length = termLenByLevel[cellLevel];
assert result.length <= term.length;
return result;
}
private void decodeTermFreqs() throws IOException {
// logger.debug("Decode Term Freq in Node: {}", this.hashCode());
// logger.debug("Decode Term Freq in Node at {}", in.getFilePointer());
in.readBytes(termFreqCompressedBuffer.bytes, 0, termFreqCompressedBufferLength);
termFreqCompressedBuffer.offset = 0;
termFreqCompressedBuffer.length = termFreqCompressedBufferLength;
nodDecompressor.decompress(termFreqCompressedBuffer, termFreqBuffer);
// set length limit based on block size, as certain decompressor with
// large window size can set it larger than the blockSize, e.g., AFor
termFreqBuffer.length = termFreqBlockSize;
termFreqReadPending = false;
}
private void decodeNodeLengths() throws IOException {
// logger.debug("Decode Nodes Length: {}", this.hashCode());
// logger.debug("Decode Nodes Length at {}", in.getFilePointer());
in.readBytes(nodLenCompressedBuffer.bytes, 0, nodLenCompressedBufferLength);
nodLenCompressedBuffer.offset = 0;
nodLenCompressedBuffer.length = nodLenCompressedBufferLength;
nodDecompressor.decompress(nodLenCompressedBuffer, nodLenBuffer);
// set length limit based on block size, as certain decompressor with
// large window size can set it larger than the blockSize, e.g., AFor
nodLenBuffer.length = nodLenBlockSize;
nodLenReadPending = false;
}
@Test
public void testCreateAndSplitId() {
BytesRef createUid = Uid.createUidAsBytes("foo", "bar");
BytesRef[] splitUidIntoTypeAndId = Uid.splitUidIntoTypeAndId(createUid);
assertThat("foo", equalTo(splitUidIntoTypeAndId[0].utf8ToString()));
assertThat("bar", equalTo(splitUidIntoTypeAndId[1].utf8ToString()));
// split also with an offset
BytesRef ref = new BytesRef(createUid.length + 10);
ref.offset = 9;
ref.length = createUid.length;
System.arraycopy(createUid.bytes, createUid.offset, ref.bytes, ref.offset, ref.length);
splitUidIntoTypeAndId = Uid.splitUidIntoTypeAndId(ref);
assertThat("foo", equalTo(splitUidIntoTypeAndId[0].utf8ToString()));
assertThat("bar", equalTo(splitUidIntoTypeAndId[1].utf8ToString()));
}
@Override
public void compress(final IntsRef input, final BytesRef output) {
assert input.ints.length % 32 == 0;
final int[] uncompressedData = input.ints;
final byte[] compressedData = output.bytes;
// prepare the input buffer before starting the compression
this.prepareInputBuffer(input);
while (input.offset < input.length) {
for (final long compressorCode :
this.frameCompressorCodes(uncompressedData, input.offset, input.length)) {
compressedData[output.offset] = (byte) compressorCode;
this.compressors[(int) compressorCode].compress(input, output);
}
}
// flip buffer
input.offset = 0;
output.length = output.offset;
output.offset = 0;
}
@Override
public BytesRef next() throws IOException {
if (scratch == null) {
return null;
}
boolean success = false;
try {
byte[] next = reader.read();
if (next != null) {
scratch.bytes = next;
scratch.length = next.length;
scratch.offset = 0;
} else {
IOUtils.close(reader);
scratch = null;
}
success = true;
return scratch;
} finally {
if (!success) {
IOUtils.closeWhileHandlingException(reader);
}
}
}
@Override
public Cell readCell(BytesRef term, Cell scratch) {
if (scratch == null) scratch = getWorldCell();
// We decode level #, leaf boolean, and populate bytes by reference. We don't decode the stack.
// reverse lookup term length to the level and hence the cell
NRCell[] cellsByLevel = ((NRCell) scratch).cellsByLevel;
boolean isLeaf = term.bytes[term.offset + term.length - 1] == 0;
int lenNoLeaf = isLeaf ? term.length - 1 : term.length;
NRCell result = cellsByLevel[levelByTermLen[lenNoLeaf]];
if (cellsByLevel[0].termBuf == null)
cellsByLevel[0].termBuf = result.term.bytes; // a kluge; see cell.ensureOwnTermBytes()
result.term.bytes = term.bytes;
result.term.offset = term.offset;
result.term.length = lenNoLeaf; // technically this isn't used but may help debugging
result.reset();
if (isLeaf) result.setLeaf();
result.cellNumber = -1; // lazy decode flag
return result;
}
@Override
public void build(TermFreqIterator tfit) throws IOException {
if (tfit instanceof TermFreqPayloadIterator) {
throw new IllegalArgumentException("this suggester doesn't support payloads");
}
File tempInput =
File.createTempFile(
FSTCompletionLookup.class.getSimpleName(), ".input", Sort.defaultTempDir());
File tempSorted =
File.createTempFile(
FSTCompletionLookup.class.getSimpleName(), ".sorted", Sort.defaultTempDir());
Sort.ByteSequencesWriter writer = new Sort.ByteSequencesWriter(tempInput);
Sort.ByteSequencesReader reader = null;
ExternalRefSorter sorter = null;
// Push floats up front before sequences to sort them. For now, assume they are non-negative.
// If negative floats are allowed some trickery needs to be done to find their byte order.
boolean success = false;
try {
byte[] buffer = new byte[0];
ByteArrayDataOutput output = new ByteArrayDataOutput(buffer);
BytesRef spare;
while ((spare = tfit.next()) != null) {
if (spare.length + 4 >= buffer.length) {
buffer = ArrayUtil.grow(buffer, spare.length + 4);
}
output.reset(buffer);
output.writeInt(encodeWeight(tfit.weight()));
output.writeBytes(spare.bytes, spare.offset, spare.length);
writer.write(buffer, 0, output.getPosition());
}
writer.close();
// We don't know the distribution of scores and we need to bucket them, so we'll sort
// and divide into equal buckets.
SortInfo info = new Sort().sort(tempInput, tempSorted);
tempInput.delete();
FSTCompletionBuilder builder =
new FSTCompletionBuilder(
buckets, sorter = new ExternalRefSorter(new Sort()), sharedTailLength);
final int inputLines = info.lines;
reader = new Sort.ByteSequencesReader(tempSorted);
long line = 0;
int previousBucket = 0;
int previousScore = 0;
ByteArrayDataInput input = new ByteArrayDataInput();
BytesRef tmp1 = new BytesRef();
BytesRef tmp2 = new BytesRef();
while (reader.read(tmp1)) {
input.reset(tmp1.bytes);
int currentScore = input.readInt();
int bucket;
if (line > 0 && currentScore == previousScore) {
bucket = previousBucket;
} else {
bucket = (int) (line * buckets / inputLines);
}
previousScore = currentScore;
previousBucket = bucket;
// Only append the input, discard the weight.
tmp2.bytes = tmp1.bytes;
tmp2.offset = input.getPosition();
tmp2.length = tmp1.length - input.getPosition();
builder.add(tmp2, bucket);
line++;
}
// The two FSTCompletions share the same automaton.
this.higherWeightsCompletion = builder.build();
this.normalCompletion =
new FSTCompletion(higherWeightsCompletion.getFST(), false, exactMatchFirst);
success = true;
} finally {
if (success) IOUtils.close(reader, writer, sorter);
else IOUtils.closeWhileHandlingException(reader, writer, sorter);
tempInput.delete();
tempSorted.delete();
}
}
