/**
* because the field is NULL and it's source is the incoming mutation, we still need to accumulate
* it. We must be careful, however, to accumulate the proper null value.
*
* <p>In theory, we could use a sparse encoding here--just accumulate a length 0 entry, which will
* allow us to use a very short row key to determine nullity. However, that doesn't work
* correctly, because doubles and floats at the end of the index might decode the row key as a
* double, resulting in goofball answers.
*
* <p>Instead, we must use the dense encoding approach here. That means that we must select the
* proper dense type based on columnTypes[i]. For most data types, this is still a length-0 array,
* but for floats and doubles it will put the proper type into place.
*/
private void accumulateNull(EntryAccumulator keyAccumulator, int pos, int type) {
if (typeProvider.isScalar(type))
keyAccumulator.addScalar(pos, SIConstants.EMPTY_BYTE_ARRAY, 0, 0);
else if (typeProvider.isDouble(type))
keyAccumulator.addDouble(
pos, Encoding.encodedNullDouble(), 0, Encoding.encodedNullDoubleLength());
else if (typeProvider.isFloat(type))
keyAccumulator.addDouble(
pos, Encoding.encodedNullFloat(), 0, Encoding.encodedNullFloatLength());
else keyAccumulator.add(pos, SIConstants.EMPTY_BYTE_ARRAY, 0, 0);
}
public KVPair writeDirectIndex(LocatedRow locatedRow) throws IOException, StandardException {
assert locatedRow != null : "locatedRow passed in is null";
ExecRow execRow = locatedRow.getRow();
getSerializers(execRow);
EntryAccumulator keyAccumulator = getKeyAccumulator();
keyAccumulator.reset();
boolean hasNullKeyFields = false;
for (int i = 0; i < execRow.nColumns(); i++) {
if (execRow.getColumn(i + 1) == null || execRow.getColumn(i + 1).isNull()) {
hasNullKeyFields = true;
accumulateNull(keyAccumulator, i, indexFormatIds[i]);
} else {
byte[] data = serializers[i].encodeDirect(execRow.getColumn(i + 1), false);
accumulate(
keyAccumulator, i, indexFormatIds[i], index.getDescColumns(i), data, 0, data.length);
}
}
// add the row key to the end of the index key
byte[] srcRowKey = Encoding.encodeBytesUnsorted(locatedRow.getRowLocation().getBytes());
EntryEncoder rowEncoder = getRowEncoder();
MultiFieldEncoder entryEncoder = rowEncoder.getEntryEncoder();
entryEncoder.reset();
entryEncoder.setRawBytes(srcRowKey);
byte[] indexValue = rowEncoder.encode();
byte[] indexRowKey;
if (index.getUnique()) {
boolean nonUnique =
index.getUniqueWithDuplicateNulls() && (hasNullKeyFields || !keyAccumulator.isFinished());
indexRowKey = getIndexRowKey(srcRowKey, nonUnique);
} else indexRowKey = getIndexRowKey(srcRowKey, true);
return new KVPair(indexRowKey, indexValue, KVPair.Type.INSERT);
}
@Test
public void testAlwaysAcceptEntryWorks() throws Exception {
BitSet fields = new BitSet();
fields.set(0);
fields.set(2);
EntryPredicateFilter predicateFilter = new EntryPredicateFilter(fields);
EntryAccumulator accumulator = new ByteEntryAccumulator(predicateFilter, false, null);
byte[] encodedOne = Encoding.encode(1);
accumulator.add(2, encodedOne, 0, encodedOne.length);
byte[] encodedTwo = Encoding.encode(2);
accumulator.add(0, encodedTwo, 0, encodedTwo.length);
byte[] bytes = accumulator.finish();
MultiFieldDecoder decoder = MultiFieldDecoder.wrap(bytes);
Assert.assertEquals(2, decoder.decodeNextInt());
Assert.assertEquals(1, decoder.decodeNextInt());
}
/**
* Translate the given base table record mutation into its associated, referencing index record.
* <br>
* Encapsulates the logic required to create an index record for a given base table record with
* all the required discriminating and encoding rules (column is part of a PK, value is null,
* etc).
*
* @param mutation KVPair containing the rowKey of the base table record for which we want to
* translate to the associated index. This mutation should already have its requred {@link
* KVPair.Type Type} set.
* @return A KVPair representing the index record of the given base table mutation. This KVPair is
* suitable for performing the required modification of the index record associated with this
* mutation.
* @throws IOException for encoding/decoding problems.
*/
public KVPair translate(KVPair mutation) throws IOException {
if (mutation == null) {
return null;
}
EntryAccumulator keyAccumulator = getKeyAccumulator();
keyAccumulator.reset();
boolean hasNullKeyFields = false;
/*
* Handle index columns from the source table's primary key.
*/
if (table.getColumnOrderingCount() > 0) {
// we have key columns to check
MultiFieldDecoder keyDecoder = getSrcKeyDecoder();
keyDecoder.set(mutation.getRowKey());
for (int i = 0; i < table.getColumnOrderingCount(); i++) {
int sourceKeyColumnPos = table.getColumnOrdering(i);
int indexKeyPos =
sourceKeyColumnPos < mainColToIndexPosMap.length
? mainColToIndexPosMap[sourceKeyColumnPos]
: -1;
int offset = keyDecoder.offset();
boolean isNull = skip(keyDecoder, table.getFormatIds(sourceKeyColumnPos));
if (!indexedCols.get(sourceKeyColumnPos)) continue;
if (indexKeyPos >= 0) {
/*
* since primary keys have an implicit NOT NULL constraint here, we don't need to check for it,
* and isNull==true would represent a programmer error, rather than an actual state the
* system can be in.
*/
assert !isNull : "Programmer error: Cannot update a primary key to a null value!";
int length = keyDecoder.offset() - offset - 1;
/*
* A note about sort order:
*
* We are in the primary key section, which means that the element is ordered in
* ASCENDING order. In an ideal world, that wouldn't matter because
*/
accumulate(
keyAccumulator,
indexKeyPos,
table.getFormatIds(sourceKeyColumnPos),
index.getDescColumns(indexKeyPos),
keyDecoder.array(),
offset,
length);
}
}
}
/*
* Handle non-null index columns from the source tables non-primary key columns.
*
* this will set indexed columns with values taken from the incoming mutation (rather than
* backfilling them with existing values, which would occur elsewhere).
*/
EntryDecoder rowDecoder = getSrcValueDecoder();
rowDecoder.set(mutation.getValue());
BitIndex bitIndex = rowDecoder.getCurrentIndex();
MultiFieldDecoder rowFieldDecoder = rowDecoder.getEntryDecoder();
for (int i = bitIndex.nextSetBit(0); i >= 0; i = bitIndex.nextSetBit(i + 1)) {
if (!indexedCols.get(i)) {
// skip non-indexed columns
rowDecoder.seekForward(rowFieldDecoder, i);
continue;
}
int keyColumnPos = i < mainColToIndexPosMap.length ? mainColToIndexPosMap[i] : -1;
if (keyColumnPos < 0) {
rowDecoder.seekForward(rowFieldDecoder, i);
} else {
int offset = rowFieldDecoder.offset();
boolean isNull = rowDecoder.seekForward(rowFieldDecoder, i);
hasNullKeyFields = isNull || hasNullKeyFields;
int length;
if (!isNull) {
length = rowFieldDecoder.offset() - offset - 1;
accumulate(
keyAccumulator,
keyColumnPos,
table.getFormatIds(i),
index.getDescColumns(keyColumnPos),
rowFieldDecoder.array(),
offset,
length);
} else {
/*
* because the field is NULL and it's source is the incoming mutation, we
* still need to accumulate it. We must be careful, however, to accumulate the
* proper null value.
*
* In theory, we could use a sparse encoding here--just accumulate a length 0 entry,
* which will allow us to use a very short row key to determine nullity. However, that
* doesn't work correctly, because doubles and floats at the end of the index might decode
* the row key as a double, resulting in goofball answers.
*
* Instead, we must use the dense encoding approach here. That means that we must
* select the proper dense type based on columnTypes[i]. For most data types, this is still
* a length-0 array, but for floats and doubles it will put the proper type into place.
*/
accumulateNull(keyAccumulator, keyColumnPos, table.getFormatIds(i));
}
}
}
/*
* Handle NULL index columns from the source tables non-primary key columns.
*/
for (int srcColIndex = 0; srcColIndex < mainColToIndexPosMap.length; srcColIndex++) {
/* position of the source column within the index encoding */
int indexColumnPosition = mainColToIndexPosMap[srcColIndex];
if (!isSourceColumnPrimaryKey(srcColIndex)
&& indexColumnPosition >= 0
&& !bitIndex.isSet(srcColIndex)) {
hasNullKeyFields = true;
keyAccumulator.add(indexColumnPosition, new byte[] {}, 0, 0);
}
}
// add the row key to the end of the index key
byte[] srcRowKey = Encoding.encodeBytesUnsorted(mutation.getRowKey());
EntryEncoder rowEncoder = getRowEncoder();
MultiFieldEncoder entryEncoder = rowEncoder.getEntryEncoder();
entryEncoder.reset();
entryEncoder.setRawBytes(srcRowKey);
byte[] indexValue = rowEncoder.encode();
byte[] indexRowKey;
if (index.getUnique()) {
boolean nonUnique =
index.getUniqueWithDuplicateNulls() && (hasNullKeyFields || !keyAccumulator.isFinished());
indexRowKey = getIndexRowKey(srcRowKey, nonUnique);
} else indexRowKey = getIndexRowKey(srcRowKey, true);
return new KVPair(indexRowKey, indexValue, mutation.getType());
}