public static String findExistingColumn(PTable table, List<PColumn> columns) {
for (PColumn column : columns) {
PName familyName = column.getFamilyName();
if (familyName == null) {
try {
return table.getPKColumn(column.getName().getString()).getName().getString();
} catch (ColumnNotFoundException e) {
continue;
}
} else {
try {
return table
.getColumnFamily(familyName.getString())
.getColumn(column.getName().getString())
.getName()
.getString();
} catch (ColumnFamilyNotFoundException e) {
continue; // Shouldn't happen
} catch (ColumnNotFoundException e) {
continue;
}
}
}
return null;
}
/**
* Check that none of no columns in our updatable VIEW are changing values.
*
* @param tableRef
* @param overlapViewColumns
* @param targetColumns
* @param projector
* @throws SQLException
*/
private static void throwIfNotUpdatable(
TableRef tableRef,
Set<PColumn> overlapViewColumns,
List<PColumn> targetColumns,
RowProjector projector,
boolean sameTable)
throws SQLException {
PTable table = tableRef.getTable();
if (table.getViewType() == ViewType.UPDATABLE && !overlapViewColumns.isEmpty()) {
ImmutableBytesWritable ptr = new ImmutableBytesWritable();
for (int i = 0; i < targetColumns.size(); i++) {
PColumn targetColumn = targetColumns.get(i);
if (overlapViewColumns.contains(targetColumn)) {
Expression source = projector.getColumnProjector(i).getExpression();
if (source.isStateless()) {
source.evaluate(null, ptr);
if (Bytes.compareTo(
ptr.get(),
ptr.getOffset(),
ptr.getLength(),
targetColumn.getViewConstant(),
0,
targetColumn.getViewConstant().length - 1)
== 0) {
continue;
}
}
throw new SQLExceptionInfo.Builder(SQLExceptionCode.CANNOT_UPDATE_VIEW_COLUMN)
.setColumnName(targetColumn.getName().getString())
.build()
.buildException();
}
}
}
}
/**
* @return a fully qualified column name in the format: "CFNAME"."COLNAME" or "COLNAME" depending
* on whether or not there is a column family name present.
*/
public static String getQuotedFullColumnName(PColumn pCol) {
checkNotNull(pCol);
String columnName = pCol.getName().getString();
String columnFamilyName =
pCol.getFamilyName() != null ? pCol.getFamilyName().getString() : null;
return getQuotedFullColumnName(columnFamilyName, columnName);
}
@Override
public ColumnRef resolveColumn(String schemaName, String tableName, String colName)
throws SQLException {
TableRef tableRef = tableRefs.get(0);
boolean resolveCF = false;
if (schemaName != null || tableName != null) {
String resolvedTableName = tableRef.getTable().getTableName().getString();
String resolvedSchemaName = tableRef.getTable().getSchemaName().getString();
if (schemaName != null && tableName != null) {
if (!(schemaName.equals(resolvedSchemaName) && tableName.equals(resolvedTableName))) {
if (!(resolveCF = schemaName.equals(alias))) {
throw new ColumnNotFoundException(schemaName, tableName, null, colName);
}
}
} else { // schemaName == null && tableName != null
if (tableName != null
&& !tableName.equals(alias)
&& (!tableName.equals(resolvedTableName) || !resolvedSchemaName.equals(""))) {
resolveCF = true;
}
}
}
PColumn column =
resolveCF
? tableRef.getTable().getColumnFamily(tableName).getColumn(colName)
: tableRef.getTable().getColumn(colName);
return new ColumnRef(tableRef, column.getPosition());
}
/**
* Imperfect estimate of row size given a PTable TODO: keep row count in stats table and use total
* size / row count instead
*
* @param table
* @return estimate of size in bytes of a row
*/
public static long estimateRowSize(PTable table) {
int keyLength = estimateKeyLength(table);
long rowSize = 0;
for (PColumn column : table.getColumns()) {
if (!SchemaUtil.isPKColumn(column)) {
PDataType type = column.getDataType();
Integer maxLength = column.getMaxLength();
int valueLength =
!type.isFixedWidth()
? VAR_KV_LENGTH_ESTIMATE
: maxLength == null ? type.getByteSize() : maxLength;
rowSize +=
KeyValue.getKeyValueDataStructureSize(
keyLength,
column.getFamilyName().getBytes().length,
column.getName().getBytes().length,
valueLength);
}
}
// Empty key value
rowSize +=
KeyValue.getKeyValueDataStructureSize(
keyLength,
getEmptyColumnFamily(table).length,
QueryConstants.EMPTY_COLUMN_BYTES.length,
0);
return rowSize;
}
/*
Map all unique pairs <family, name> to index. Table name is part of TableRowkey, so we do
not care about it
*/
private void initColumnIndexes() throws SQLException {
columnIndexes = new TreeMap(Bytes.BYTES_COMPARATOR);
int columnIndex = 0;
for (int index = 0; index < logicalNames.size(); index++) {
PTable table = PhoenixRuntime.getTable(conn, logicalNames.get(index));
List<PColumn> cls = table.getColumns();
for (int i = 0; i < cls.size(); i++) {
PColumn c = cls.get(i);
byte[] family = new byte[0];
if (c.getFamilyName() != null) // Skip PK column
family = c.getFamilyName().getBytes();
byte[] name = c.getName().getBytes();
byte[] cfn = Bytes.add(family, QueryConstants.NAMESPACE_SEPARATOR_BYTES, name);
if (!columnIndexes.containsKey(cfn)) {
columnIndexes.put(cfn, new Integer(columnIndex));
columnIndex++;
}
}
byte[] emptyColumnFamily = SchemaUtil.getEmptyColumnFamily(table);
byte[] cfn =
Bytes.add(
emptyColumnFamily,
QueryConstants.NAMESPACE_SEPARATOR_BYTES,
QueryConstants.EMPTY_COLUMN_BYTES);
columnIndexes.put(cfn, new Integer(columnIndex));
columnIndex++;
}
}
/**
* Constructs a column info for the supplied pColumn
*
* @param pColumn
* @return columnInfo
* @throws SQLException if the parameter is null.
*/
public static ColumnInfo getColumnInfo(PColumn pColumn) throws SQLException {
if (pColumn == null) {
throw new SQLException("pColumn must not be null.");
}
int sqlType = pColumn.getDataType().getResultSetSqlType();
ColumnInfo columnInfo = new ColumnInfo(pColumn.toString(), sqlType);
return columnInfo;
}
@Override
public Expression visit(LiteralParseNode node) throws SQLException {
if (isTopLevel()) {
return LiteralExpression.newConstant(
node.getValue(), column.getDataType(), column.getSortOrder(), true);
}
return super.visit(node);
}
/**
* Validates that the meta data is valid against the server meta data if we haven't yet done so.
* Otherwise, for every UPSERT VALUES call, we'd need to hit the server to see if the meta data
* has changed.
*
* @param connection
* @return the server time to use for the upsert
* @throws SQLException if the table or any columns no longer exist
*/
private long[] validate() throws SQLException {
int i = 0;
Long scn = connection.getSCN();
PName tenantId = connection.getTenantId();
MetaDataClient client = new MetaDataClient(connection);
long[] timeStamps = new long[this.mutations.size()];
for (Map.Entry<TableRef, Map<ImmutableBytesPtr, Map<PColumn, byte[]>>> entry :
mutations.entrySet()) {
TableRef tableRef = entry.getKey();
long serverTimeStamp = tableRef.getTimeStamp();
PTable table = tableRef.getTable();
// If we're auto committing, we've already validated the schema when we got the
// ColumnResolver,
// so no need to do it again here.
if (!connection.getAutoCommit()) {
MetaDataMutationResult result =
client.updateCache(table.getSchemaName().getString(), table.getTableName().getString());
long timestamp = result.getMutationTime();
if (timestamp != QueryConstants.UNSET_TIMESTAMP) {
serverTimeStamp = timestamp;
if (result.wasUpdated()) {
// TODO: use bitset?
table =
connection
.getMetaDataCache()
.getTable(new PTableKey(tenantId, table.getName().getString()));
PColumn[] columns = new PColumn[table.getColumns().size()];
for (Map.Entry<ImmutableBytesPtr, Map<PColumn, byte[]>> rowEntry :
entry.getValue().entrySet()) {
Map<PColumn, byte[]> valueEntry = rowEntry.getValue();
if (valueEntry != PRow.DELETE_MARKER) {
for (PColumn column : valueEntry.keySet()) {
columns[column.getPosition()] = column;
}
}
}
for (PColumn column : columns) {
if (column != null) {
table
.getColumnFamily(column.getFamilyName().getString())
.getColumn(column.getName().getString());
}
}
tableRef.setTable(table);
}
}
}
timeStamps[i++] =
scn == null
? serverTimeStamp == QueryConstants.UNSET_TIMESTAMP
? HConstants.LATEST_TIMESTAMP
: serverTimeStamp
: scn;
}
return timeStamps;
}
@Override
public Expression visit(BindParseNode node) throws SQLException {
if (isTopLevel()) {
context.getBindManager().addParamMetaData(node, column);
Object value = context.getBindManager().getBindValue(node);
return LiteralExpression.newConstant(
value, column.getDataType(), column.getSortOrder(), true);
}
return super.visit(node);
}
// Estimate the key length after pos slot for schema.
private static int estimatePartLength(int pos, Iterator<PColumn> iterator) {
int length = 0;
while (iterator.hasNext()) {
PColumn column = iterator.next();
if (column.getDataType().isFixedWidth()) {
length += SchemaUtil.getFixedByteSize(column);
} else {
length += 1; // SEPARATOR byte.
}
}
return length;
}
/**
* Estimate the max key length in bytes of the PK for a given table
*
* @param table the table
* @return the max PK length
*/
public static int estimateKeyLength(PTable table) {
int maxKeyLength = 0;
// Calculate the max length of a key (each part must currently be of a fixed width)
int i = 0;
List<PColumn> columns = table.getPKColumns();
while (i < columns.size()) {
PColumn keyColumn = columns.get(i++);
PDataType type = keyColumn.getDataType();
Integer maxLength = keyColumn.getMaxLength();
maxKeyLength +=
!type.isFixedWidth()
? VAR_LENGTH_ESTIMATE
: maxLength == null ? type.getByteSize() : maxLength;
}
return maxKeyLength;
}
/**
* Get list of ColumnInfos that contain Column Name and its associated PDataType for an import.
* The supplied list of columns can be null -- if it is non-null, it represents a user-supplied
* list of columns to be imported.
*
* @param conn Phoenix connection from which metadata will be read
* @param tableName Phoenix table name whose columns are to be checked. Can include a schema name
* @param columns user-supplied list of import columns, can be null
*/
public static List<ColumnInfo> generateColumnInfo(
Connection conn, String tableName, List<String> columns) throws SQLException {
PTable table = PhoenixRuntime.getTable(conn, tableName);
List<ColumnInfo> columnInfoList = Lists.newArrayList();
Set<String> unresolvedColumnNames = new TreeSet<String>();
if (columns == null) {
// use all columns in the table
for (PColumn pColumn : table.getColumns()) {
int sqlType = pColumn.getDataType().getResultSetSqlType();
columnInfoList.add(new ColumnInfo(pColumn.toString(), sqlType));
}
} else {
// Leave "null" as indication to skip b/c it doesn't exist
for (int i = 0; i < columns.size(); i++) {
String columnName = columns.get(i);
try {
ColumnInfo columnInfo = PhoenixRuntime.getColumnInfo(table, columnName);
columnInfoList.add(columnInfo);
} catch (ColumnNotFoundException cnfe) {
unresolvedColumnNames.add(columnName.trim());
} catch (AmbiguousColumnException ace) {
unresolvedColumnNames.add(columnName.trim());
}
}
}
// if there exists columns that cannot be resolved, error out.
if (unresolvedColumnNames.size() > 0) {
StringBuilder exceptionMessage = new StringBuilder();
boolean first = true;
exceptionMessage.append("Unable to resolve these column names:\n");
for (String col : unresolvedColumnNames) {
if (first) first = false;
else exceptionMessage.append(",");
exceptionMessage.append(col);
}
exceptionMessage.append("\nAvailable columns with column families:\n");
first = true;
for (PColumn pColumn : table.getColumns()) {
if (first) first = false;
else exceptionMessage.append(",");
exceptionMessage.append(pColumn.toString());
}
throw new SQLException(exceptionMessage.toString());
}
return columnInfoList;
}
// Go through each slot in the schema and try match it with the split byte array. If the split
// does not confer to the schema, extends its length to match the schema.
private static byte[] processSplit(byte[] split, LinkedHashSet<PColumn> pkColumns) {
int pos = 0, offset = 0, maxOffset = split.length;
Iterator<PColumn> iterator = pkColumns.iterator();
while (pos < pkColumns.size()) {
PColumn column = iterator.next();
if (column.getDataType().isFixedWidth()) { // Fixed width
int length = SchemaUtil.getFixedByteSize(column);
if (maxOffset - offset < length) {
// The split truncates the field. Fill in the rest of the part and any fields that
// are missing after this field.
int fillInLength = length - (maxOffset - offset);
fillInLength += estimatePartLength(pos + 1, iterator);
return ByteUtil.fillKey(split, split.length + fillInLength);
}
// Account for this field, move to next position;
offset += length;
pos++;
} else { // Variable length
// If we are the last slot, then we are done. Nothing needs to be filled in.
if (pos == pkColumns.size() - 1) {
break;
}
while (offset < maxOffset && split[offset] != QueryConstants.SEPARATOR_BYTE) {
offset++;
}
if (offset == maxOffset) {
// The var-length field does not end with a separator and it's not the last field.
int fillInLength = 1; // SEPARATOR byte for the current var-length slot.
fillInLength += estimatePartLength(pos + 1, iterator);
return ByteUtil.fillKey(split, split.length + fillInLength);
}
// Move to the next position;
offset += 1; // skip separator;
pos++;
}
}
return split;
}
private static SelectStatement prependTenantAndViewConstants(
PTable table, SelectStatement select, String tenantId, Set<PColumn> addViewColumns) {
if ((!table.isMultiTenant() || tenantId == null)
&& table.getViewIndexId() == null
&& addViewColumns.isEmpty()) {
return select;
}
List<AliasedNode> selectNodes =
newArrayListWithCapacity(select.getSelect().size() + 1 + addViewColumns.size());
if (table.isMultiTenant() && tenantId != null) {
selectNodes.add(new AliasedNode(null, new LiteralParseNode(tenantId)));
}
if (table.getViewIndexId() != null) {
selectNodes.add(new AliasedNode(null, new LiteralParseNode(table.getViewIndexId())));
}
selectNodes.addAll(select.getSelect());
for (PColumn column : addViewColumns) {
byte[] byteValue = column.getViewConstant();
Object value = column.getDataType().toObject(byteValue, 0, byteValue.length - 1);
selectNodes.add(new AliasedNode(null, new LiteralParseNode(value)));
}
return SelectStatement.create(select, selectNodes);
}
@Override
public ColumnRef resolveColumn(String schemaName, String tableName, String colName)
throws SQLException {
if (tableName == null) {
int theColumnPosition = -1;
TableRef theTableRef = null;
Iterator<TableRef> iterator = tables.iterator();
while (iterator.hasNext()) {
TableRef tableRef = iterator.next();
try {
PColumn column = tableRef.getTable().getColumn(colName);
if (theTableRef != null) {
throw new AmbiguousColumnException(colName);
}
theTableRef = tableRef;
theColumnPosition = column.getPosition();
} catch (ColumnNotFoundException e) {
}
}
if (theTableRef != null) {
return new ColumnRef(theTableRef, theColumnPosition);
}
throw new ColumnNotFoundException(colName);
} else {
try {
TableRef tableRef = resolveTable(schemaName, tableName);
PColumn column = tableRef.getTable().getColumn(colName);
return new ColumnRef(tableRef, column.getPosition());
} catch (TableNotFoundException e) {
// Try using the tableName as a columnFamily reference instead
ColumnFamilyRef cfRef = resolveColumnFamily(schemaName, tableName);
PColumn column = cfRef.getFamily().getColumn(colName);
return new ColumnRef(cfRef.getTableRef(), column.getPosition());
}
}
}
private static QueryPlan addPlan(
PhoenixStatement statement,
SelectStatement select,
PTable index,
List<? extends PDatum> targetColumns,
ParallelIteratorFactory parallelIteratorFactory,
QueryPlan dataPlan,
boolean isHinted)
throws SQLException {
int nColumns = dataPlan.getProjector().getColumnCount();
String tableAlias = dataPlan.getTableRef().getTableAlias();
String alias =
tableAlias == null
? null
: '"' + tableAlias + '"'; // double quote in case it's case sensitive
String schemaName = index.getParentSchemaName().getString();
schemaName = schemaName.length() == 0 ? null : '"' + schemaName + '"';
String tableName = '"' + index.getTableName().getString() + '"';
TableNode table = FACTORY.namedTable(alias, FACTORY.table(schemaName, tableName));
SelectStatement indexSelect = FACTORY.select(select, table);
ColumnResolver resolver =
FromCompiler.getResolverForQuery(indexSelect, statement.getConnection());
// We will or will not do tuple projection according to the data plan.
boolean isProjected =
dataPlan.getContext().getResolver().getTables().get(0).getTable().getType()
== PTableType.PROJECTED;
// Check index state of now potentially updated index table to make sure it's active
if (PIndexState.ACTIVE.equals(resolver.getTables().get(0).getTable().getIndexState())) {
try {
// translate nodes that match expressions that are indexed to the associated column parse
// node
indexSelect =
ParseNodeRewriter.rewrite(
indexSelect,
new IndexExpressionParseNodeRewriter(
index, statement.getConnection(), indexSelect.getUdfParseNodes()));
QueryCompiler compiler =
new QueryCompiler(
statement,
indexSelect,
resolver,
targetColumns,
parallelIteratorFactory,
dataPlan.getContext().getSequenceManager(),
isProjected);
QueryPlan plan = compiler.compile();
// If query doesn't have where clause and some of columns to project are missing
// in the index then we need to get missing columns from main table for each row in
// local index. It's like full scan of both local index and data table which is inefficient.
// Then we don't use the index. If all the columns to project are present in the index
// then we can use the index even the query doesn't have where clause.
if (index.getIndexType() == IndexType.LOCAL
&& indexSelect.getWhere() == null
&& !plan.getContext().getDataColumns().isEmpty()) {
return null;
}
// Checking number of columns handles the wildcard cases correctly, as in that case the
// index
// must contain all columns from the data table to be able to be used.
if (plan.getTableRef().getTable().getIndexState() == PIndexState.ACTIVE) {
if (plan.getProjector().getColumnCount() == nColumns) {
return plan;
} else if (index.getIndexType() == IndexType.GLOBAL) {
throw new ColumnNotFoundException("*");
}
}
} catch (ColumnNotFoundException e) {
/* Means that a column is being used that's not in our index.
* Since we currently don't keep stats, we don't know the selectivity of the index.
* For now, if this is a hinted plan, we will try rewriting the query as a subquery;
* otherwise we just don't use this index (as opposed to trying to join back from
* the index table to the data table.
*/
SelectStatement dataSelect = (SelectStatement) dataPlan.getStatement();
ParseNode where = dataSelect.getWhere();
if (isHinted && where != null) {
StatementContext context = new StatementContext(statement, resolver);
WhereConditionRewriter whereRewriter =
new WhereConditionRewriter(FromCompiler.getResolver(dataPlan.getTableRef()), context);
where = where.accept(whereRewriter);
if (where != null) {
PTable dataTable = dataPlan.getTableRef().getTable();
List<PColumn> pkColumns = dataTable.getPKColumns();
List<AliasedNode> aliasedNodes =
Lists.<AliasedNode>newArrayListWithExpectedSize(pkColumns.size());
List<ParseNode> nodes = Lists.<ParseNode>newArrayListWithExpectedSize(pkColumns.size());
boolean isSalted = dataTable.getBucketNum() != null;
boolean isTenantSpecific =
dataTable.isMultiTenant() && statement.getConnection().getTenantId() != null;
int posOffset = (isSalted ? 1 : 0) + (isTenantSpecific ? 1 : 0);
for (int i = posOffset; i < pkColumns.size(); i++) {
PColumn column = pkColumns.get(i);
String indexColName = IndexUtil.getIndexColumnName(column);
ParseNode indexColNode =
new ColumnParseNode(null, '"' + indexColName + '"', indexColName);
PDataType indexColType = IndexUtil.getIndexColumnDataType(column);
PDataType dataColType = column.getDataType();
if (indexColType != dataColType) {
indexColNode = FACTORY.cast(indexColNode, dataColType, null, null);
}
aliasedNodes.add(FACTORY.aliasedNode(null, indexColNode));
nodes.add(new ColumnParseNode(null, '"' + column.getName().getString() + '"'));
}
SelectStatement innerSelect =
FACTORY.select(
indexSelect.getFrom(),
indexSelect.getHint(),
false,
aliasedNodes,
where,
null,
null,
null,
null,
null,
indexSelect.getBindCount(),
false,
indexSelect.hasSequence(),
Collections.<SelectStatement>emptyList(),
indexSelect.getUdfParseNodes());
ParseNode outerWhere =
FACTORY.in(
nodes.size() == 1 ? nodes.get(0) : FACTORY.rowValueConstructor(nodes),
FACTORY.subquery(innerSelect, false),
false,
true);
ParseNode extractedCondition = whereRewriter.getExtractedCondition();
if (extractedCondition != null) {
outerWhere = FACTORY.and(Lists.newArrayList(outerWhere, extractedCondition));
}
HintNode hint =
HintNode.combine(
HintNode.subtract(
indexSelect.getHint(),
new Hint[] {Hint.INDEX, Hint.NO_CHILD_PARENT_JOIN_OPTIMIZATION}),
FACTORY.hint("NO_INDEX"));
SelectStatement query = FACTORY.select(dataSelect, hint, outerWhere);
ColumnResolver queryResolver =
FromCompiler.getResolverForQuery(query, statement.getConnection());
query = SubqueryRewriter.transform(query, queryResolver, statement.getConnection());
queryResolver = FromCompiler.getResolverForQuery(query, statement.getConnection());
query = StatementNormalizer.normalize(query, queryResolver);
QueryPlan plan =
new QueryCompiler(
statement,
query,
queryResolver,
targetColumns,
parallelIteratorFactory,
dataPlan.getContext().getSequenceManager(),
isProjected)
.compile();
return plan;
}
}
}
}
return null;
}
public MutationPlan compile(UpsertStatement upsert) throws SQLException {
final PhoenixConnection connection = statement.getConnection();
ConnectionQueryServices services = connection.getQueryServices();
final int maxSize =
services
.getProps()
.getInt(
QueryServices.MAX_MUTATION_SIZE_ATTRIB,
QueryServicesOptions.DEFAULT_MAX_MUTATION_SIZE);
final ColumnResolver resolver = FromCompiler.getResolverForMutation(upsert, connection);
final TableRef tableRef = resolver.getTables().get(0);
final PTable table = tableRef.getTable();
if (table.getType() == PTableType.VIEW) {
if (table.getViewType().isReadOnly()) {
throw new ReadOnlyTableException(
table.getSchemaName().getString(), table.getTableName().getString());
}
}
boolean isSalted = table.getBucketNum() != null;
final boolean isTenantSpecific = table.isMultiTenant() && connection.getTenantId() != null;
final boolean isSharedViewIndex = table.getViewIndexId() != null;
String tenantId = isTenantSpecific ? connection.getTenantId().getString() : null;
int posOffset = isSalted ? 1 : 0;
// Setup array of column indexes parallel to values that are going to be set
List<ColumnName> columnNodes = upsert.getColumns();
final List<PColumn> allColumns = table.getColumns();
Set<PColumn> addViewColumnsToBe = Collections.emptySet();
Set<PColumn> overlapViewColumnsToBe = Collections.emptySet();
int[] columnIndexesToBe;
int nColumnsToSet = 0;
int[] pkSlotIndexesToBe;
List<PColumn> targetColumns;
if (table.getViewType() == ViewType.UPDATABLE) {
addViewColumnsToBe = Sets.newLinkedHashSetWithExpectedSize(allColumns.size());
for (PColumn column : allColumns) {
if (column.getViewConstant() != null) {
addViewColumnsToBe.add(column);
}
}
}
ImmutableBytesWritable ptr = new ImmutableBytesWritable();
// Allow full row upsert if no columns or only dynamic ones are specified and values count match
if (columnNodes.isEmpty()
|| columnNodes.size() == upsert.getTable().getDynamicColumns().size()) {
nColumnsToSet = allColumns.size() - posOffset;
columnIndexesToBe = new int[nColumnsToSet];
pkSlotIndexesToBe = new int[columnIndexesToBe.length];
targetColumns = Lists.newArrayListWithExpectedSize(columnIndexesToBe.length);
targetColumns.addAll(Collections.<PColumn>nCopies(columnIndexesToBe.length, null));
int minPKPos = 0;
if (isTenantSpecific) {
PColumn tenantColumn = table.getPKColumns().get(minPKPos);
columnIndexesToBe[minPKPos] = tenantColumn.getPosition();
targetColumns.set(minPKPos, tenantColumn);
minPKPos++;
}
if (isSharedViewIndex) {
PColumn indexIdColumn = table.getPKColumns().get(minPKPos);
columnIndexesToBe[minPKPos] = indexIdColumn.getPosition();
targetColumns.set(minPKPos, indexIdColumn);
minPKPos++;
}
for (int i = posOffset, j = 0; i < allColumns.size(); i++) {
PColumn column = allColumns.get(i);
if (SchemaUtil.isPKColumn(column)) {
pkSlotIndexesToBe[i - posOffset] = j + posOffset;
if (j++ < minPKPos) { // Skip, as it's already been set above
continue;
}
minPKPos = 0;
}
columnIndexesToBe[i - posOffset + minPKPos] = i;
targetColumns.set(i - posOffset + minPKPos, column);
}
if (!addViewColumnsToBe.isEmpty()) {
// All view columns overlap in this case
overlapViewColumnsToBe = addViewColumnsToBe;
addViewColumnsToBe = Collections.emptySet();
}
} else {
// Size for worse case
int numColsInUpsert = columnNodes.size();
nColumnsToSet =
numColsInUpsert
+ addViewColumnsToBe.size()
+ (isTenantSpecific ? 1 : 0)
+ +(isSharedViewIndex ? 1 : 0);
columnIndexesToBe = new int[nColumnsToSet];
pkSlotIndexesToBe = new int[columnIndexesToBe.length];
targetColumns = Lists.newArrayListWithExpectedSize(columnIndexesToBe.length);
targetColumns.addAll(Collections.<PColumn>nCopies(columnIndexesToBe.length, null));
Arrays.fill(
columnIndexesToBe,
-1); // TODO: necessary? So we'll get an AIOB exception if it's not replaced
Arrays.fill(
pkSlotIndexesToBe,
-1); // TODO: necessary? So we'll get an AIOB exception if it's not replaced
BitSet pkColumnsSet = new BitSet(table.getPKColumns().size());
int i = 0;
// Add tenant column directly, as we don't want to resolve it as this will fail
if (isTenantSpecific) {
PColumn tenantColumn = table.getPKColumns().get(i + posOffset);
columnIndexesToBe[i] = tenantColumn.getPosition();
pkColumnsSet.set(pkSlotIndexesToBe[i] = i + posOffset);
targetColumns.set(i, tenantColumn);
i++;
}
if (isSharedViewIndex) {
PColumn indexIdColumn = table.getPKColumns().get(i + posOffset);
columnIndexesToBe[i] = indexIdColumn.getPosition();
pkColumnsSet.set(pkSlotIndexesToBe[i] = i + posOffset);
targetColumns.set(i, indexIdColumn);
i++;
}
for (ColumnName colName : columnNodes) {
ColumnRef ref =
resolver.resolveColumn(null, colName.getFamilyName(), colName.getColumnName());
PColumn column = ref.getColumn();
if (IndexUtil.getViewConstantValue(column, ptr)) {
if (overlapViewColumnsToBe.isEmpty()) {
overlapViewColumnsToBe = Sets.newHashSetWithExpectedSize(addViewColumnsToBe.size());
}
nColumnsToSet--;
overlapViewColumnsToBe.add(column);
addViewColumnsToBe.remove(column);
}
columnIndexesToBe[i] = ref.getColumnPosition();
targetColumns.set(i, column);
if (SchemaUtil.isPKColumn(column)) {
pkColumnsSet.set(pkSlotIndexesToBe[i] = ref.getPKSlotPosition());
}
i++;
}
for (PColumn column : addViewColumnsToBe) {
columnIndexesToBe[i] = column.getPosition();
targetColumns.set(i, column);
if (SchemaUtil.isPKColumn(column)) {
pkColumnsSet.set(pkSlotIndexesToBe[i] = SchemaUtil.getPKPosition(table, column));
}
i++;
}
for (i = posOffset; i < table.getPKColumns().size(); i++) {
PColumn pkCol = table.getPKColumns().get(i);
if (!pkColumnsSet.get(i)) {
if (!pkCol.isNullable()) {
throw new ConstraintViolationException(
table.getName().getString()
+ "."
+ pkCol.getName().getString()
+ " may not be null");
}
}
}
}
List<ParseNode> valueNodes = upsert.getValues();
QueryPlan plan = null;
RowProjector rowProjectorToBe = null;
final int nValuesToSet;
boolean sameTable = false;
boolean runOnServer = false;
UpsertingParallelIteratorFactory upsertParallelIteratorFactoryToBe = null;
final boolean isAutoCommit = connection.getAutoCommit();
if (valueNodes == null) {
SelectStatement select = upsert.getSelect();
assert (select != null);
select = SubselectRewriter.flatten(select, connection);
ColumnResolver selectResolver = FromCompiler.getResolverForQuery(select, connection);
select = StatementNormalizer.normalize(select, selectResolver);
select = prependTenantAndViewConstants(table, select, tenantId, addViewColumnsToBe);
sameTable =
select.getFrom().size() == 1 && tableRef.equals(selectResolver.getTables().get(0));
/* We can run the upsert in a coprocessor if:
* 1) from has only 1 table and the into table matches from table
* 2) the select query isn't doing aggregation
* 3) autoCommit is on
* 4) the table is not immutable, as the client is the one that figures out the additional
* puts for index tables.
* 5) no limit clause
* Otherwise, run the query to pull the data from the server
* and populate the MutationState (upto a limit).
*/
runOnServer =
sameTable
&& isAutoCommit
&& !table.isImmutableRows()
&& !select.isAggregate()
&& !select.isDistinct()
&& select.getLimit() == null
&& table.getBucketNum() == null;
ParallelIteratorFactory parallelIteratorFactory;
if (select.isAggregate() || select.isDistinct() || select.getLimit() != null) {
parallelIteratorFactory = null;
} else {
// We can pipeline the upsert select instead of spooling everything to disk first,
// if we don't have any post processing that's required.
parallelIteratorFactory =
upsertParallelIteratorFactoryToBe =
new UpsertingParallelIteratorFactory(connection, tableRef);
}
// If we may be able to run on the server, add a hint that favors using the data table
// if all else is equal.
// TODO: it'd be nice if we could figure out in advance if the PK is potentially changing,
// as this would disallow running on the server. We currently use the row projector we
// get back to figure this out.
HintNode hint = upsert.getHint();
if (!upsert.getHint().hasHint(Hint.USE_INDEX_OVER_DATA_TABLE)) {
hint = HintNode.create(hint, Hint.USE_DATA_OVER_INDEX_TABLE);
}
select = SelectStatement.create(select, hint);
// Pass scan through if same table in upsert and select so that projection is computed
// correctly
// Use optimizer to choose the best plan
plan =
new QueryOptimizer(services)
.optimize(statement, select, selectResolver, targetColumns, parallelIteratorFactory);
runOnServer &= plan.getTableRef().equals(tableRef);
rowProjectorToBe = plan.getProjector();
nValuesToSet = rowProjectorToBe.getColumnCount();
// Cannot auto commit if doing aggregation or topN or salted
// Salted causes problems because the row may end up living on a different region
} else {
nValuesToSet =
valueNodes.size()
+ addViewColumnsToBe.size()
+ (isTenantSpecific ? 1 : 0)
+ (isSharedViewIndex ? 1 : 0);
}
final RowProjector projector = rowProjectorToBe;
final UpsertingParallelIteratorFactory upsertParallelIteratorFactory =
upsertParallelIteratorFactoryToBe;
final QueryPlan queryPlan = plan;
// Resize down to allow a subset of columns to be specifiable
if (columnNodes.isEmpty() && columnIndexesToBe.length >= nValuesToSet) {
nColumnsToSet = nValuesToSet;
columnIndexesToBe = Arrays.copyOf(columnIndexesToBe, nValuesToSet);
pkSlotIndexesToBe = Arrays.copyOf(pkSlotIndexesToBe, nValuesToSet);
}
if (nValuesToSet != nColumnsToSet) {
throw new SQLExceptionInfo.Builder(SQLExceptionCode.UPSERT_COLUMN_NUMBERS_MISMATCH)
.setMessage(
"Numbers of columns: " + nColumnsToSet + ". Number of values: " + nValuesToSet)
.build()
.buildException();
}
final int[] columnIndexes = columnIndexesToBe;
final int[] pkSlotIndexes = pkSlotIndexesToBe;
final Set<PColumn> addViewColumns = addViewColumnsToBe;
final Set<PColumn> overlapViewColumns = overlapViewColumnsToBe;
// TODO: break this up into multiple functions
////////////////////////////////////////////////////////////////////
// UPSERT SELECT
/////////////////////////////////////////////////////////////////////
if (valueNodes == null) {
// Before we re-order, check that for updatable view columns
// the projected expression either matches the column name or
// is a constant with the same required value.
throwIfNotUpdatable(tableRef, overlapViewColumnsToBe, targetColumns, projector, sameTable);
////////////////////////////////////////////////////////////////////
// UPSERT SELECT run server-side (maybe)
/////////////////////////////////////////////////////////////////////
if (runOnServer) {
// At most this array will grow bigger by the number of PK columns
int[] allColumnsIndexes = Arrays.copyOf(columnIndexes, columnIndexes.length + nValuesToSet);
int[] reverseColumnIndexes = new int[table.getColumns().size()];
List<Expression> projectedExpressions =
Lists.newArrayListWithExpectedSize(reverseColumnIndexes.length);
Arrays.fill(reverseColumnIndexes, -1);
for (int i = 0; i < nValuesToSet; i++) {
projectedExpressions.add(projector.getColumnProjector(i).getExpression());
reverseColumnIndexes[columnIndexes[i]] = i;
}
/*
* Order projected columns and projected expressions with PK columns
* leading order by slot position
*/
int offset = table.getBucketNum() == null ? 0 : 1;
for (int i = 0; i < table.getPKColumns().size() - offset; i++) {
PColumn column = table.getPKColumns().get(i + offset);
int pos = reverseColumnIndexes[column.getPosition()];
if (pos == -1) {
// Last PK column may be fixed width and nullable
// We don't want to insert a null expression b/c
// it's not valid to set a fixed width type to null.
if (column.getDataType().isFixedWidth()) {
continue;
}
// Add literal null for missing PK columns
pos = projectedExpressions.size();
Expression literalNull =
LiteralExpression.newConstant(null, column.getDataType(), true);
projectedExpressions.add(literalNull);
allColumnsIndexes[pos] = column.getPosition();
}
// Swap select expression at pos with i
Collections.swap(projectedExpressions, i, pos);
// Swap column indexes and reverse column indexes too
int tempPos = allColumnsIndexes[i];
allColumnsIndexes[i] = allColumnsIndexes[pos];
allColumnsIndexes[pos] = tempPos;
reverseColumnIndexes[tempPos] = reverseColumnIndexes[i];
reverseColumnIndexes[i] = i;
}
// If any pk slots are changing, be conservative and don't run this server side.
// If the row ends up living in a different region, we'll get an error otherwise.
for (int i = 0; i < table.getPKColumns().size(); i++) {
PColumn column = table.getPKColumns().get(i);
Expression source = projectedExpressions.get(i);
if (source == null
|| !source.equals(
new ColumnRef(tableRef, column.getPosition()).newColumnExpression())) {
// TODO: we could check the region boundaries to see if the pk will still be in it.
runOnServer = false; // bail on running server side, since PK may be changing
break;
}
}
////////////////////////////////////////////////////////////////////
// UPSERT SELECT run server-side
/////////////////////////////////////////////////////////////////////
if (runOnServer) {
// Iterate through columns being projected
List<PColumn> projectedColumns =
Lists.newArrayListWithExpectedSize(projectedExpressions.size());
for (int i = 0; i < projectedExpressions.size(); i++) {
// Must make new column if position has changed
PColumn column = allColumns.get(allColumnsIndexes[i]);
projectedColumns.add(column.getPosition() == i ? column : new PColumnImpl(column, i));
}
// Build table from projectedColumns
PTable projectedTable = PTableImpl.makePTable(table, projectedColumns);
SelectStatement select =
SelectStatement.create(SelectStatement.COUNT_ONE, upsert.getHint());
final RowProjector aggProjector =
ProjectionCompiler.compile(queryPlan.getContext(), select, GroupBy.EMPTY_GROUP_BY);
/*
* Transfer over PTable representing subset of columns selected, but all PK columns.
* Move columns setting PK first in pkSlot order, adding LiteralExpression of null for any missing ones.
* Transfer over List<Expression> for projection.
* In region scan, evaluate expressions in order, collecting first n columns for PK and collection non PK in mutation Map
* Create the PRow and get the mutations, adding them to the batch
*/
final StatementContext context = queryPlan.getContext();
final Scan scan = context.getScan();
scan.setAttribute(
BaseScannerRegionObserver.UPSERT_SELECT_TABLE,
UngroupedAggregateRegionObserver.serialize(projectedTable));
scan.setAttribute(
BaseScannerRegionObserver.UPSERT_SELECT_EXPRS,
UngroupedAggregateRegionObserver.serialize(projectedExpressions));
// Ignore order by - it has no impact
final QueryPlan aggPlan =
new AggregatePlan(
context,
select,
tableRef,
aggProjector,
null,
OrderBy.EMPTY_ORDER_BY,
null,
GroupBy.EMPTY_GROUP_BY,
null);
return new MutationPlan() {
@Override
public PhoenixConnection getConnection() {
return connection;
}
@Override
public ParameterMetaData getParameterMetaData() {
return queryPlan.getContext().getBindManager().getParameterMetaData();
}
@Override
public StatementContext getContext() {
return queryPlan.getContext();
}
@Override
public MutationState execute() throws SQLException {
ImmutableBytesWritable ptr = context.getTempPtr();
tableRef.getTable().getIndexMaintainers(ptr);
ServerCache cache = null;
try {
if (ptr.getLength() > 0) {
IndexMetaDataCacheClient client =
new IndexMetaDataCacheClient(connection, tableRef);
cache = client.addIndexMetadataCache(context.getScanRanges(), ptr);
byte[] uuidValue = cache.getId();
scan.setAttribute(PhoenixIndexCodec.INDEX_UUID, uuidValue);
}
ResultIterator iterator = aggPlan.iterator();
try {
Tuple row = iterator.next();
final long mutationCount =
(Long) aggProjector.getColumnProjector(0).getValue(row, PDataType.LONG, ptr);
return new MutationState(maxSize, connection) {
@Override
public long getUpdateCount() {
return mutationCount;
}
};
} finally {
iterator.close();
}
} finally {
if (cache != null) {
cache.close();
}
}
}
@Override
public ExplainPlan getExplainPlan() throws SQLException {
List<String> queryPlanSteps = aggPlan.getExplainPlan().getPlanSteps();
List<String> planSteps =
Lists.newArrayListWithExpectedSize(queryPlanSteps.size() + 1);
planSteps.add("UPSERT ROWS");
planSteps.addAll(queryPlanSteps);
return new ExplainPlan(planSteps);
}
};
}
}
////////////////////////////////////////////////////////////////////
// UPSERT SELECT run client-side
/////////////////////////////////////////////////////////////////////
return new MutationPlan() {
@Override
public PhoenixConnection getConnection() {
return connection;
}
@Override
public ParameterMetaData getParameterMetaData() {
return queryPlan.getContext().getBindManager().getParameterMetaData();
}
@Override
public StatementContext getContext() {
return queryPlan.getContext();
}
@Override
public MutationState execute() throws SQLException {
ResultIterator iterator = queryPlan.iterator();
if (upsertParallelIteratorFactory == null) {
return upsertSelect(
statement, tableRef, projector, iterator, columnIndexes, pkSlotIndexes);
}
upsertParallelIteratorFactory.setRowProjector(projector);
upsertParallelIteratorFactory.setColumnIndexes(columnIndexes);
upsertParallelIteratorFactory.setPkSlotIndexes(pkSlotIndexes);
Tuple tuple;
long totalRowCount = 0;
while ((tuple = iterator.next()) != null) { // Runs query
Cell kv = tuple.getValue(0);
totalRowCount +=
PDataType.LONG
.getCodec()
.decodeLong(kv.getValueArray(), kv.getValueOffset(), SortOrder.getDefault());
}
// Return total number of rows that have been updated. In the case of auto commit being
// off
// the mutations will all be in the mutation state of the current connection.
return new MutationState(maxSize, statement.getConnection(), totalRowCount);
}
@Override
public ExplainPlan getExplainPlan() throws SQLException {
List<String> queryPlanSteps = queryPlan.getExplainPlan().getPlanSteps();
List<String> planSteps = Lists.newArrayListWithExpectedSize(queryPlanSteps.size() + 1);
planSteps.add("UPSERT SELECT");
planSteps.addAll(queryPlanSteps);
return new ExplainPlan(planSteps);
}
};
}
////////////////////////////////////////////////////////////////////
// UPSERT VALUES
/////////////////////////////////////////////////////////////////////
int nodeIndex = 0;
// initialze values with constant byte values first
final byte[][] values = new byte[nValuesToSet][];
if (isTenantSpecific) {
values[nodeIndex++] = connection.getTenantId().getBytes();
}
if (isSharedViewIndex) {
values[nodeIndex++] = MetaDataUtil.getViewIndexIdDataType().toBytes(table.getViewIndexId());
}
final int nodeIndexOffset = nodeIndex;
// Allocate array based on size of all columns in table,
// since some values may not be set (if they're nullable).
final StatementContext context = new StatementContext(statement, resolver, new Scan());
UpsertValuesCompiler expressionBuilder = new UpsertValuesCompiler(context);
final List<Expression> constantExpressions =
Lists.newArrayListWithExpectedSize(valueNodes.size());
// First build all the expressions, as with sequences we want to collect them all first
// and initialize them in one batch
for (ParseNode valueNode : valueNodes) {
if (!valueNode.isStateless()) {
throw new SQLExceptionInfo.Builder(SQLExceptionCode.VALUE_IN_UPSERT_NOT_CONSTANT)
.build()
.buildException();
}
PColumn column = allColumns.get(columnIndexes[nodeIndex]);
expressionBuilder.setColumn(column);
Expression expression = valueNode.accept(expressionBuilder);
if (expression.getDataType() != null
&& !expression.getDataType().isCastableTo(column.getDataType())) {
throw TypeMismatchException.newException(
expression.getDataType(),
column.getDataType(),
"expression: " + expression.toString() + " in column " + column);
}
constantExpressions.add(expression);
nodeIndex++;
}
return new MutationPlan() {
@Override
public PhoenixConnection getConnection() {
return connection;
}
@Override
public ParameterMetaData getParameterMetaData() {
return context.getBindManager().getParameterMetaData();
}
@Override
public StatementContext getContext() {
return context;
}
@Override
public MutationState execute() throws SQLException {
ImmutableBytesWritable ptr = context.getTempPtr();
final SequenceManager sequenceManager = context.getSequenceManager();
// Next evaluate all the expressions
int nodeIndex = nodeIndexOffset;
Tuple tuple =
sequenceManager.getSequenceCount() == 0 ? null : sequenceManager.newSequenceTuple(null);
for (Expression constantExpression : constantExpressions) {
PColumn column = allColumns.get(columnIndexes[nodeIndex]);
constantExpression.evaluate(tuple, ptr);
Object value = null;
if (constantExpression.getDataType() != null) {
value =
constantExpression
.getDataType()
.toObject(
ptr,
constantExpression.getSortOrder(),
constantExpression.getMaxLength(),
constantExpression.getScale());
if (!constantExpression.getDataType().isCoercibleTo(column.getDataType(), value)) {
throw TypeMismatchException.newException(
constantExpression.getDataType(),
column.getDataType(),
"expression: " + constantExpression.toString() + " in column " + column);
}
if (!column
.getDataType()
.isSizeCompatible(
ptr,
value,
constantExpression.getDataType(),
constantExpression.getMaxLength(),
constantExpression.getScale(),
column.getMaxLength(),
column.getScale())) {
throw new SQLExceptionInfo.Builder(SQLExceptionCode.DATA_EXCEEDS_MAX_CAPACITY)
.setColumnName(column.getName().getString())
.setMessage("value=" + constantExpression.toString())
.build()
.buildException();
}
}
column
.getDataType()
.coerceBytes(
ptr,
value,
constantExpression.getDataType(),
constantExpression.getMaxLength(),
constantExpression.getScale(),
constantExpression.getSortOrder(),
column.getMaxLength(),
column.getScale(),
column.getSortOrder());
if (overlapViewColumns.contains(column)
&& Bytes.compareTo(
ptr.get(),
ptr.getOffset(),
ptr.getLength(),
column.getViewConstant(),
0,
column.getViewConstant().length - 1)
!= 0) {
throw new SQLExceptionInfo.Builder(SQLExceptionCode.CANNOT_UPDATE_VIEW_COLUMN)
.setColumnName(column.getName().getString())
.setMessage("value=" + constantExpression.toString())
.build()
.buildException();
}
values[nodeIndex] = ByteUtil.copyKeyBytesIfNecessary(ptr);
nodeIndex++;
}
// Add columns based on view
for (PColumn column : addViewColumns) {
if (IndexUtil.getViewConstantValue(column, ptr)) {
values[nodeIndex++] = ByteUtil.copyKeyBytesIfNecessary(ptr);
} else {
throw new IllegalStateException();
}
}
Map<ImmutableBytesPtr, Map<PColumn, byte[]>> mutation = Maps.newHashMapWithExpectedSize(1);
setValues(values, pkSlotIndexes, columnIndexes, tableRef.getTable(), mutation);
return new MutationState(tableRef, mutation, 0, maxSize, connection);
}
@Override
public ExplainPlan getExplainPlan() throws SQLException {
List<String> planSteps = Lists.newArrayListWithExpectedSize(2);
if (context.getSequenceManager().getSequenceCount() > 0) {
planSteps.add(
"CLIENT RESERVE " + context.getSequenceManager().getSequenceCount() + " SEQUENCES");
}
planSteps.add("PUT SINGLE ROW");
return new ExplainPlan(planSteps);
}
};
}
private static MutationState upsertSelect(
PhoenixStatement statement,
TableRef tableRef,
RowProjector projector,
ResultIterator iterator,
int[] columnIndexes,
int[] pkSlotIndexes)
throws SQLException {
try {
PhoenixConnection connection = statement.getConnection();
ConnectionQueryServices services = connection.getQueryServices();
int maxSize =
services
.getProps()
.getInt(
QueryServices.MAX_MUTATION_SIZE_ATTRIB,
QueryServicesOptions.DEFAULT_MAX_MUTATION_SIZE);
int batchSize = Math.min(connection.getMutateBatchSize(), maxSize);
boolean isAutoCommit = connection.getAutoCommit();
byte[][] values = new byte[columnIndexes.length][];
int rowCount = 0;
Map<ImmutableBytesPtr, Map<PColumn, byte[]>> mutation =
Maps.newHashMapWithExpectedSize(batchSize);
PTable table = tableRef.getTable();
ResultSet rs = new PhoenixResultSet(iterator, projector, statement);
ImmutableBytesWritable ptr = new ImmutableBytesWritable();
while (rs.next()) {
for (int i = 0; i < values.length; i++) {
PColumn column = table.getColumns().get(columnIndexes[i]);
byte[] bytes = rs.getBytes(i + 1);
ptr.set(bytes == null ? ByteUtil.EMPTY_BYTE_ARRAY : bytes);
Object value = rs.getObject(i + 1);
int rsPrecision = rs.getMetaData().getPrecision(i + 1);
Integer precision = rsPrecision == 0 ? null : rsPrecision;
int rsScale = rs.getMetaData().getScale(i + 1);
Integer scale = rsScale == 0 ? null : rsScale;
// We are guaranteed that the two column will have compatible types,
// as we checked that before.
if (!column
.getDataType()
.isSizeCompatible(
ptr,
value,
column.getDataType(),
precision,
scale,
column.getMaxLength(),
column.getScale())) {
throw new SQLExceptionInfo.Builder(SQLExceptionCode.DATA_EXCEEDS_MAX_CAPACITY)
.setColumnName(column.getName().getString())
.setMessage("value=" + column.getDataType().toStringLiteral(ptr, null))
.build()
.buildException();
}
column
.getDataType()
.coerceBytes(
ptr,
value,
column.getDataType(),
precision,
scale,
SortOrder.getDefault(),
column.getMaxLength(),
column.getScale(),
column.getSortOrder());
values[i] = ByteUtil.copyKeyBytesIfNecessary(ptr);
}
setValues(values, pkSlotIndexes, columnIndexes, table, mutation);
rowCount++;
// Commit a batch if auto commit is true and we're at our batch size
if (isAutoCommit && rowCount % batchSize == 0) {
MutationState state = new MutationState(tableRef, mutation, 0, maxSize, connection);
connection.getMutationState().join(state);
connection.commit();
mutation.clear();
}
}
// If auto commit is true, this last batch will be committed upon return
return new MutationState(
tableRef, mutation, rowCount / batchSize * batchSize, maxSize, connection);
} finally {
iterator.close();
}
}
public static boolean isPKColumn(PColumn column) {
return column.getFamilyName() == null;
}
@Override
protected RegionScanner doPostScannerOpen(
final ObserverContext<RegionCoprocessorEnvironment> c, final Scan scan, final RegionScanner s)
throws IOException {
byte[] isUngroupedAgg = scan.getAttribute(BaseScannerRegionObserver.UNGROUPED_AGG);
if (isUngroupedAgg == null) {
return s;
}
final ScanProjector p = ScanProjector.deserializeProjectorFromScan(scan);
final HashJoinInfo j = HashJoinInfo.deserializeHashJoinFromScan(scan);
RegionScanner theScanner = s;
if (p != null || j != null) {
theScanner =
new HashJoinRegionScanner(s, p, j, ScanUtil.getTenantId(scan), c.getEnvironment());
}
final RegionScanner innerScanner = theScanner;
byte[] indexUUID = scan.getAttribute(PhoenixIndexCodec.INDEX_UUID);
PTable projectedTable = null;
List<Expression> selectExpressions = null;
byte[] upsertSelectTable = scan.getAttribute(BaseScannerRegionObserver.UPSERT_SELECT_TABLE);
boolean isUpsert = false;
boolean isDelete = false;
byte[] deleteCQ = null;
byte[] deleteCF = null;
byte[][] values = null;
byte[] emptyCF = null;
ImmutableBytesWritable ptr = null;
if (upsertSelectTable != null) {
isUpsert = true;
projectedTable = deserializeTable(upsertSelectTable);
selectExpressions =
deserializeExpressions(scan.getAttribute(BaseScannerRegionObserver.UPSERT_SELECT_EXPRS));
values = new byte[projectedTable.getPKColumns().size()][];
ptr = new ImmutableBytesWritable();
} else {
byte[] isDeleteAgg = scan.getAttribute(BaseScannerRegionObserver.DELETE_AGG);
isDelete = isDeleteAgg != null && Bytes.compareTo(PDataType.TRUE_BYTES, isDeleteAgg) == 0;
if (!isDelete) {
deleteCF = scan.getAttribute(BaseScannerRegionObserver.DELETE_CF);
deleteCQ = scan.getAttribute(BaseScannerRegionObserver.DELETE_CQ);
}
emptyCF = scan.getAttribute(BaseScannerRegionObserver.EMPTY_CF);
}
int batchSize = 0;
long ts = scan.getTimeRange().getMax();
HRegion region = c.getEnvironment().getRegion();
List<Mutation> mutations = Collections.emptyList();
if (isDelete || isUpsert || (deleteCQ != null && deleteCF != null) || emptyCF != null) {
// TODO: size better
mutations = Lists.newArrayListWithExpectedSize(1024);
batchSize =
c.getEnvironment()
.getConfiguration()
.getInt(MUTATE_BATCH_SIZE_ATTRIB, QueryServicesOptions.DEFAULT_MUTATE_BATCH_SIZE);
}
Aggregators aggregators =
ServerAggregators.deserialize(
scan.getAttribute(BaseScannerRegionObserver.AGGREGATORS),
c.getEnvironment().getConfiguration());
Aggregator[] rowAggregators = aggregators.getAggregators();
boolean hasMore;
boolean hasAny = false;
MultiKeyValueTuple result = new MultiKeyValueTuple();
if (logger.isInfoEnabled()) {
logger.info("Starting ungrouped coprocessor scan " + scan);
}
long rowCount = 0;
region.startRegionOperation();
try {
do {
List<Cell> results = new ArrayList<Cell>();
// Results are potentially returned even when the return value of s.next is false
// since this is an indication of whether or not there are more values after the
// ones returned
hasMore = innerScanner.nextRaw(results);
if (!results.isEmpty()) {
rowCount++;
result.setKeyValues(results);
try {
if (isDelete) {
// FIXME: the version of the Delete constructor without the lock args was introduced
// in 0.94.4, thus if we try to use it here we can no longer use the 0.94.2 version
// of the client.
Cell firstKV = results.get(0);
Delete delete =
new Delete(
firstKV.getRowArray(), firstKV.getRowOffset(), firstKV.getRowLength(), ts);
mutations.add(delete);
} else if (isUpsert) {
Arrays.fill(values, null);
int i = 0;
List<PColumn> projectedColumns = projectedTable.getColumns();
for (; i < projectedTable.getPKColumns().size(); i++) {
Expression expression = selectExpressions.get(i);
if (expression.evaluate(result, ptr)) {
values[i] = ptr.copyBytes();
// If SortOrder from expression in SELECT doesn't match the
// column being projected into then invert the bits.
if (expression.getSortOrder() != projectedColumns.get(i).getSortOrder()) {
SortOrder.invert(values[i], 0, values[i], 0, values[i].length);
}
}
}
projectedTable.newKey(ptr, values);
PRow row = projectedTable.newRow(kvBuilder, ts, ptr);
for (; i < projectedColumns.size(); i++) {
Expression expression = selectExpressions.get(i);
if (expression.evaluate(result, ptr)) {
PColumn column = projectedColumns.get(i);
Object value = expression.getDataType().toObject(ptr, column.getSortOrder());
// We are guaranteed that the two column will have the same type.
if (!column
.getDataType()
.isSizeCompatible(
ptr,
value,
column.getDataType(),
expression.getMaxLength(),
expression.getScale(),
column.getMaxLength(),
column.getScale())) {
throw new ValueTypeIncompatibleException(
column.getDataType(), column.getMaxLength(), column.getScale());
}
column
.getDataType()
.coerceBytes(
ptr,
value,
expression.getDataType(),
expression.getMaxLength(),
expression.getScale(),
expression.getSortOrder(),
column.getMaxLength(),
column.getScale(),
column.getSortOrder());
byte[] bytes = ByteUtil.copyKeyBytesIfNecessary(ptr);
row.setValue(column, bytes);
}
}
for (Mutation mutation : row.toRowMutations()) {
mutations.add(mutation);
}
} else if (deleteCF != null && deleteCQ != null) {
// No need to search for delete column, since we project only it
// if no empty key value is being set
if (emptyCF == null || result.getValue(deleteCF, deleteCQ) != null) {
Delete delete =
new Delete(
results.get(0).getRowArray(),
results.get(0).getRowOffset(),
results.get(0).getRowLength());
delete.deleteColumns(deleteCF, deleteCQ, ts);
mutations.add(delete);
}
}
if (emptyCF != null) {
/*
* If we've specified an emptyCF, then we need to insert an empty
* key value "retroactively" for any key value that is visible at
* the timestamp that the DDL was issued. Key values that are not
* visible at this timestamp will not ever be projected up to
* scans past this timestamp, so don't need to be considered.
* We insert one empty key value per row per timestamp.
*/
Set<Long> timeStamps = Sets.newHashSetWithExpectedSize(results.size());
for (Cell kv : results) {
long kvts = kv.getTimestamp();
if (!timeStamps.contains(kvts)) {
Put put = new Put(kv.getRowArray(), kv.getRowOffset(), kv.getRowLength());
put.add(
emptyCF, QueryConstants.EMPTY_COLUMN_BYTES, kvts, ByteUtil.EMPTY_BYTE_ARRAY);
mutations.add(put);
}
}
}
// Commit in batches based on UPSERT_BATCH_SIZE_ATTRIB in config
if (!mutations.isEmpty() && batchSize > 0 && mutations.size() % batchSize == 0) {
commitBatch(region, mutations, indexUUID);
mutations.clear();
}
} catch (ConstraintViolationException e) {
// Log and ignore in count
logger.error(
"Failed to create row in "
+ region.getRegionNameAsString()
+ " with values "
+ SchemaUtil.toString(values),
e);
continue;
}
aggregators.aggregate(rowAggregators, result);
hasAny = true;
}
} while (hasMore);
} finally {
innerScanner.close();
region.closeRegionOperation();
}
if (logger.isInfoEnabled()) {
logger.info("Finished scanning " + rowCount + " rows for ungrouped coprocessor scan " + scan);
}
if (!mutations.isEmpty()) {
commitBatch(region, mutations, indexUUID);
}
final boolean hadAny = hasAny;
KeyValue keyValue = null;
if (hadAny) {
byte[] value = aggregators.toBytes(rowAggregators);
keyValue =
KeyValueUtil.newKeyValue(
UNGROUPED_AGG_ROW_KEY,
SINGLE_COLUMN_FAMILY,
SINGLE_COLUMN,
AGG_TIMESTAMP,
value,
0,
value.length);
}
final KeyValue aggKeyValue = keyValue;
RegionScanner scanner =
new BaseRegionScanner() {
private boolean done = !hadAny;
@Override
public HRegionInfo getRegionInfo() {
return innerScanner.getRegionInfo();
}
@Override
public boolean isFilterDone() {
return done;
}
@Override
public void close() throws IOException {
innerScanner.close();
}
@Override
public boolean next(List<Cell> results) throws IOException {
if (done) return false;
done = true;
results.add(aggKeyValue);
return false;
}
@Override
public long getMaxResultSize() {
return scan.getMaxResultSize();
}
};
return scanner;
}
/**
* Order the plans among all the possible ones from best to worst. Since we don't keep stats yet,
* we use the following simple algorithm: 1) If the query is a point lookup (i.e. we have a set of
* exact row keys), choose among those. 2) If the query has an ORDER BY and a LIMIT, choose the
* plan that has all the ORDER BY expression in the same order as the row key columns. 3) If there
* are more than one plan that meets (1&2), choose the plan with: a) the most row key columns that
* may be used to form the start/stop scan key. b) the plan that preserves ordering for a group
* by. c) the data table plan
*
* @param plans the list of candidate plans
* @return list of plans ordered from best to worst.
*/
private List<QueryPlan> orderPlansBestToWorst(
SelectStatement select, List<QueryPlan> plans, boolean stopAtBestPlan) {
final QueryPlan dataPlan = plans.get(0);
if (plans.size() == 1) {
return plans;
}
/**
* If we have a plan(s) that are just point lookups (i.e. fully qualified row keys), then favor
* those first.
*/
List<QueryPlan> candidates = Lists.newArrayListWithExpectedSize(plans.size());
if (stopAtBestPlan) { // If we're stopping at the best plan, only consider point lookups if
// there are any
for (QueryPlan plan : plans) {
if (plan.getContext().getScanRanges().isPointLookup()) {
candidates.add(plan);
}
}
} else {
candidates.addAll(plans);
}
/**
* If we have a plan(s) that removes the order by, choose from among these, as this is typically
* the most expensive operation. Once we have stats, if there's a limit on the query, we might
* choose a different plan. For example if the limit was a very large number and the combination
* of applying other filters on the row key are estimated to choose fewer rows, we'd choose that
* one.
*/
List<QueryPlan> stillCandidates = plans;
List<QueryPlan> bestCandidates = candidates;
if (!candidates.isEmpty()) {
stillCandidates = candidates;
bestCandidates = Lists.<QueryPlan>newArrayListWithExpectedSize(candidates.size());
}
for (QueryPlan plan : stillCandidates) {
// If ORDER BY optimized out (or not present at all)
if (plan.getOrderBy().getOrderByExpressions().isEmpty()) {
bestCandidates.add(plan);
}
}
if (bestCandidates.isEmpty()) {
bestCandidates.addAll(stillCandidates);
}
int nViewConstants = 0;
PTable dataTable = dataPlan.getTableRef().getTable();
if (dataTable.getType() == PTableType.VIEW) {
for (PColumn column : dataTable.getColumns()) {
if (column.getViewConstant() != null) {
nViewConstants++;
}
}
}
final int boundRanges = nViewConstants;
final int comparisonOfDataVersusIndexTable =
select.getHint().hasHint(Hint.USE_DATA_OVER_INDEX_TABLE) ? -1 : 1;
Collections.sort(
bestCandidates,
new Comparator<QueryPlan>() {
@Override
public int compare(QueryPlan plan1, QueryPlan plan2) {
PTable table1 = plan1.getTableRef().getTable();
PTable table2 = plan2.getTableRef().getTable();
// For shared indexes (i.e. indexes on views and local indexes),
// a) add back any view constants as these won't be in the index, and
// b) ignore the viewIndexId which will be part of the row key columns.
int c =
(plan2.getContext().getScanRanges().getBoundPkColumnCount()
+ (table2.getViewIndexId() == null ? 0 : (boundRanges - 1)))
- (plan1.getContext().getScanRanges().getBoundPkColumnCount()
+ (table1.getViewIndexId() == null ? 0 : (boundRanges - 1)));
if (c != 0) return c;
if (plan1.getGroupBy() != null && plan2.getGroupBy() != null) {
if (plan1.getGroupBy().isOrderPreserving()
!= plan2.getGroupBy().isOrderPreserving()) {
return plan1.getGroupBy().isOrderPreserving() ? -1 : 1;
}
}
// Use smaller table (table with fewest kv columns)
c =
(table1.getColumns().size() - table1.getPKColumns().size())
- (table2.getColumns().size() - table2.getPKColumns().size());
if (c != 0) return c;
// If all things are equal, don't choose local index as it forces scan
// on every region (unless there's no start/stop key)
if (table1.getIndexType() == IndexType.LOCAL) {
return plan1.getContext().getScanRanges().getRanges().isEmpty() ? -1 : 1;
}
if (table2.getIndexType() == IndexType.LOCAL) {
return plan2.getContext().getScanRanges().getRanges().isEmpty() ? 1 : -1;
}
// All things being equal, just use the table based on the
// Hint.USE_DATA_OVER_INDEX_TABLE
if (table1.getType() == PTableType.INDEX) {
return comparisonOfDataVersusIndexTable;
}
if (table2.getType() == PTableType.INDEX) {
return -comparisonOfDataVersusIndexTable;
}
return 0;
}
});
return bestCandidates;
}