* 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();
final ByteBuffer startPrefix;
if (startKey.remaining() > 0)
{
CompositeType.Builder 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]);
}
startPrefix = builder.build();
}
else
{
startPrefix = ByteBufferUtil.EMPTY_BYTE_BUFFER;
}
final ByteBuffer endPrefix;
if (endKey.remaining() > 0)
{
CompositeType.Builder 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]);
}
endPrefix = builder.buildAsEndOfRange();
}
else
{
endPrefix = ByteBufferUtil.EMPTY_BYTE_BUFFER;
}
// 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);
CompositeType.Builder 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 ArrayDeque<IColumn> indexColumns;
private final QueryPath path = new QueryPath(baseCfs.columnFamily);
private int columnsRead = Integer.MAX_VALUE;
private int limit = ((SliceQueryFilter)filter.initialFilter()).count;
private int columnsCount = 0;
private 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();
assert key != null;
return new Row(key, data);
}
protected Row computeNext()
{
/*
* Our internal index code is wired toward internal rows. So we need to accumulate 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;
while (true)
{
// Did we get more columns that needed to respect the user limit?
// (but we still need to return what has been fetched 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();
indexColumns = new ArrayDeque<IColumn>(sortedColumns);
IColumn firstColumn = sortedColumns.iterator().next();
// Paging is racy, so it is possible the first column of a page is not the last seen one.
if (lastSeenPrefix != startPrefix && lastSeenPrefix.equals(firstColumn.name()))
{
// skip the row we already saw w/ the last page of results
indexColumns.poll();
logger.trace("Skipping {}", indexComparator.getString(firstColumn.name()));
}
else if (range instanceof Range && !indexColumns.isEmpty() && firstColumn.name().equals(startPrefix))
{
// skip key excluded by range
indexColumns.poll();
logger.trace("Skipping first key as range excludes it");
}
}
while (!indexColumns.isEmpty() && columnsCount <= limit)
{
IColumn column = indexColumns.poll();
lastSeenPrefix = column.name();
if (column.isMarkedForDelete())
{
logger.trace("skipping {}", column.name());
continue;
}
ByteBuffer[] components = indexComparator.split(lastSeenPrefix);
DecoratedKey dk = baseCfs.partitioner.decorateKey(components[0]);
// Are we done for this row?
if (currentKey == null)
{
currentKey = dk;
}
else if (!currentKey.equals(dk))
{
DecoratedKey previousKey = currentKey;
currentKey = dk;
// We're done with the previous row, return it if it had data, continue otherwise
indexColumns.addFirst(column);
if (data == null)
continue;
else
return makeReturn(previousKey, data);
}
if (!range.right.isMinimum(baseCfs.partitioner) && range.right.compareTo(dk) < 0)
{
logger.trace("Reached end of assigned scan range");
return endOfData();
}
if (!range.contains(dk))
{
logger.debug("Skipping entry {} outside of assigned scan range", dk.token);
continue;
}
logger.trace("Adding index hit to current row for {}", indexComparator.getString(lastSeenPrefix));
// For sparse composites, we're good querying the whole logical row
// Obviously if this index is used for other usage, that might be inefficient
CompositeType.Builder builder = baseComparator.builder();
for (int i = 0; i < prefixSize; i++)
builder.add(components[i + 1]);