{
// Start with the most-restrictive indexed clause, then apply remaining clauses
// to each row matching that clause.
// TODO: allow merge join instead of just one index + loop
final IndexExpression primary = highestSelectivityPredicate(filter.getClause());
final SecondaryIndex index = indexManager.getIndexForColumn(primary.column_name);
assert index != null;
final DecoratedKey indexKey = index.getIndexKeyFor(primary.value);
if (logger.isDebugEnabled())
logger.debug("Most-selective indexed predicate is {}",
((AbstractSimplePerColumnSecondaryIndex) index).expressionString(primary));
/*
* XXX: If the range requested is a token range, we'll have to start at the beginning (and stop at the end) of
* the indexed row unfortunately (which will be inefficient), because we have not way to intuit the small
* possible key having a given token. A fix would be to actually store the token along the key in the
* indexed row.
*/
ByteBuffer startKey = range.left instanceof DecoratedKey ? ((DecoratedKey)range.left).key : ByteBufferUtil.EMPTY_BYTE_BUFFER;
ByteBuffer endKey = range.right instanceof DecoratedKey ? ((DecoratedKey)range.right).key : ByteBufferUtil.EMPTY_BYTE_BUFFER;
final CompositeType baseComparator = (CompositeType)baseCfs.getComparator();
final CompositeType indexComparator = (CompositeType)index.getIndexCfs().getComparator();
CompositeType.Builder builder = null;
if (startKey.remaining() > 0)
{
builder = indexComparator.builder().add(startKey);
// For names filter, we have no choice but to query from the beginning of the key. This can be highly inefficient however.
if (filter.originalFilter() instanceof SliceQueryFilter)
{
ByteBuffer[] components = baseComparator.split(((SliceQueryFilter)filter.originalFilter()).start());
for (int i = 0; i < Math.min(prefixSize, components.length); ++i)
builder.add(components[i]);
}
}
final ByteBuffer startPrefix = startKey.remaining() == 0 ? ByteBufferUtil.EMPTY_BYTE_BUFFER : builder.build();
if (endKey.remaining() > 0)
{
builder = indexComparator.builder().add(endKey);
// For names filter, we have no choice but to query until the end of the key. This can be highly inefficient however.
if (filter.originalFilter() instanceof SliceQueryFilter)
{
ByteBuffer[] components = baseComparator.split(((SliceQueryFilter)filter.originalFilter()).finish());
for (int i = 0; i < Math.min(prefixSize, components.length); ++i)
builder.add(components[i]);
}
}
final ByteBuffer endPrefix = endKey.remaining() == 0 ? ByteBufferUtil.EMPTY_BYTE_BUFFER : builder.buildAsEndOfRange();
// We will need to filter clustering keys based on the user filter. If
// it is a names filter, we are really interested on the clustering
// part, not the actual column name (NOTE: this is a hack that assumes CQL3).
final SliceQueryFilter originalFilter;
if (filter.originalFilter() instanceof SliceQueryFilter)
{
originalFilter = (SliceQueryFilter)filter.originalFilter();
}
else
{
ByteBuffer first = ((NamesQueryFilter)filter.originalFilter()).columns.iterator().next();
ByteBuffer[] components = baseComparator.split(first);
builder = baseComparator.builder();
// All all except the last component, since it's the column name
for (int i = 0; i < components.length - 1; i++)
builder.add(components[i]);
originalFilter = new SliceQueryFilter(builder.copy().build(), builder.copy().buildAsEndOfRange(), false, Integer.MAX_VALUE);
}
return new ColumnFamilyStore.AbstractScanIterator()
{
private ByteBuffer lastSeenPrefix = startPrefix;
private Deque<IColumn> indexColumns;
private final QueryPath path = new QueryPath(baseCfs.columnFamily);
private int columnsRead = Integer.MAX_VALUE;
private final int meanColumns = Math.max(index.getIndexCfs().getMeanColumns(), 1);
// We shouldn't fetch only 1 row as this provides buggy paging in case the first row doesn't satisfy all clauses
private final int rowsPerQuery = Math.max(Math.min(filter.maxRows(), filter.maxColumns() / meanColumns), 2);
public boolean needsFiltering()
{
return false;
}
private Row makeReturn(DecoratedKey key, ColumnFamily data)
{
if (data == null)
{
return endOfData();
}
else
{
assert key != null;
return new Row(key, data);
}
}
protected Row computeNext()
{
/*
* Our internal index code is wired toward internal rows. So we need to acumulate all results for a given
* row before returning from this method. Which unfortunately means that this method has to do what
* CFS.filter does for KeysIndex.
*/
DecoratedKey currentKey = null;
ColumnFamily data = null;
int columnsCount = 0;
int limit = ((SliceQueryFilter)filter.initialFilter()).count;
while (true)
{
// Did we got more columns that needed to respect the user limit?
// (but we still need to return was fetch already)
if (columnsCount > limit)
return makeReturn(currentKey, data);
if (indexColumns == null || indexColumns.isEmpty())
{
if (columnsRead < rowsPerQuery)
{
logger.trace("Read only {} (< {}) last page through, must be done", columnsRead, rowsPerQuery);
return makeReturn(currentKey, data);
}
if (logger.isTraceEnabled())
logger.trace("Scanning index {} starting with {}",
((AbstractSimplePerColumnSecondaryIndex)index).expressionString(primary), indexComparator.getString(startPrefix));
QueryFilter indexFilter = QueryFilter.getSliceFilter(indexKey,
new QueryPath(index.getIndexCfs().getColumnFamilyName()),
lastSeenPrefix,
endPrefix,
false,
rowsPerQuery);
ColumnFamily indexRow = index.getIndexCfs().getColumnFamily(indexFilter);
if (indexRow == null)
return makeReturn(currentKey, data);
Collection<IColumn> sortedColumns = indexRow.getSortedColumns();
columnsRead = sortedColumns.size();