public SingleTableColumnResolver(
PhoenixConnection connection, NamedTableNode table, long timeStamp) throws SQLException {
super(connection);
List<PColumnFamily> families =
Lists.newArrayListWithExpectedSize(table.getDynamicColumns().size());
for (ColumnDef def : table.getDynamicColumns()) {
if (def.getColumnDefName().getFamilyName() != null) {
families.add(
new PColumnFamilyImpl(
PNameFactory.newName(def.getColumnDefName().getFamilyName()),
Collections.<PColumn>emptyList()));
}
}
Long scn = connection.getSCN();
PTable theTable =
new PTableImpl(
connection.getTenantId(),
table.getName().getSchemaName(),
table.getName().getTableName(),
scn == null ? HConstants.LATEST_TIMESTAMP : scn,
families);
theTable = this.addDynamicColumns(table.getDynamicColumns(), theTable);
alias = null;
tableRefs =
ImmutableList.of(
new TableRef(alias, theTable, timeStamp, !table.getDynamicColumns().isEmpty()));
}
/**
* Encode the primary key values from the table as a byte array. The values must be in the same
* order as the primary key constraint. If the connection and table are both tenant-specific, the
* tenant ID column must not be present in the values.
*
* @param conn an open connection
* @param fullTableName the full table name
* @param values the values of the primary key columns ordered in the same order as the primary
* key constraint
* @return the encoded byte array
* @throws SQLException if the table cannot be found or the incorrect number of of values are
* provided
* @see #decodePK(Connection, String, byte[]) to decode the byte[] back to the values
*/
public static byte[] encodePK(Connection conn, String fullTableName, Object[] values)
throws SQLException {
PTable table = getTable(conn, fullTableName);
PhoenixConnection pconn = conn.unwrap(PhoenixConnection.class);
int offset =
(table.getBucketNum() == null ? 0 : 1)
+ (table.isMultiTenant() && pconn.getTenantId() != null ? 1 : 0);
List<PColumn> pkColumns = table.getPKColumns();
if (pkColumns.size() - offset != values.length) {
throw new SQLException(
"Expected " + (pkColumns.size() - offset) + " but got " + values.length);
}
PDataType type = null;
TrustedByteArrayOutputStream output =
new TrustedByteArrayOutputStream(table.getRowKeySchema().getEstimatedValueLength());
try {
for (int i = offset; i < pkColumns.size(); i++) {
if (type != null && !type.isFixedWidth()) {
output.write(QueryConstants.SEPARATOR_BYTE);
}
type = pkColumns.get(i).getDataType();
byte[] value = type.toBytes(values[i - offset]);
output.write(value);
}
return output.toByteArray();
} finally {
try {
output.close();
} catch (IOException e) {
throw new RuntimeException(e); // Impossible
}
}
}
protected TableRef createTableRef(NamedTableNode tableNode, boolean updateCacheImmediately)
throws SQLException {
String tableName = tableNode.getName().getTableName();
String schemaName = tableNode.getName().getSchemaName();
long timeStamp = QueryConstants.UNSET_TIMESTAMP;
String fullTableName = SchemaUtil.getTableName(schemaName, tableName);
PName tenantId = connection.getTenantId();
PTable theTable = null;
if (updateCacheImmediately || connection.getAutoCommit()) {
MetaDataMutationResult result = client.updateCache(schemaName, tableName);
timeStamp = result.getMutationTime();
theTable = result.getTable();
if (theTable == null) {
throw new TableNotFoundException(schemaName, tableName, timeStamp);
}
} else {
try {
theTable = connection.getMetaDataCache().getTable(new PTableKey(tenantId, fullTableName));
} catch (TableNotFoundException e1) {
if (tenantId != null) { // Check with null tenantId next
try {
theTable = connection.getMetaDataCache().getTable(new PTableKey(null, fullTableName));
} catch (TableNotFoundException e2) {
}
}
}
// We always attempt to update the cache in the event of a TableNotFoundException
if (theTable == null) {
MetaDataMutationResult result = client.updateCache(schemaName, tableName);
if (result.wasUpdated()) {
timeStamp = result.getMutationTime();
theTable = result.getTable();
}
}
if (theTable == null) {
throw new TableNotFoundException(schemaName, tableName, timeStamp);
}
}
// Add any dynamic columns to the table declaration
List<ColumnDef> dynamicColumns = tableNode.getDynamicColumns();
theTable = addDynamicColumns(dynamicColumns, theTable);
TableRef tableRef =
new TableRef(tableNode.getAlias(), theTable, timeStamp, !dynamicColumns.isEmpty());
if (logger.isDebugEnabled() && timeStamp != QueryConstants.UNSET_TIMESTAMP) {
logger.debug(
"Re-resolved stale table "
+ fullTableName
+ " with seqNum "
+ tableRef.getTable().getSequenceNumber()
+ " at timestamp "
+ tableRef.getTable().getTimeStamp()
+ " with "
+ tableRef.getTable().getColumns().size()
+ " columns: "
+ tableRef.getTable().getColumns());
}
return tableRef;
}
/**
* 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;
}
public static PTable getTable(Connection conn, String name) throws SQLException {
PTable table = null;
PhoenixConnection pconn = conn.unwrap(PhoenixConnection.class);
try {
table = pconn.getMetaDataCache().getTable(new PTableKey(pconn.getTenantId(), name));
} catch (TableNotFoundException e) {
String schemaName = SchemaUtil.getSchemaNameFromFullName(name);
String tableName = SchemaUtil.getTableNameFromFullName(name);
MetaDataMutationResult result = new MetaDataClient(pconn).updateCache(schemaName, tableName);
if (result.getMutationCode() != MutationCode.TABLE_ALREADY_EXISTS) {
throw e;
}
table = result.getTable();
}
return table;
}
/**
* Decode a byte array value back into the Object values of the primary key constraint. If the
* connection and table are both tenant-specific, the tenant ID column is not expected to have
* been encoded and will not appear in the returned values.
*
* @param conn an open connection
* @param name the full table name
* @param value the value that was encoded with {@link #encodePK(Connection, String, Object[])}
* @return the Object values encoded in the byte array value
* @throws SQLException
*/
public static Object[] decodePK(Connection conn, String name, byte[] value) throws SQLException {
PTable table = getTable(conn, name);
PhoenixConnection pconn = conn.unwrap(PhoenixConnection.class);
int offset =
(table.getBucketNum() == null ? 0 : 1)
+ (table.isMultiTenant() && pconn.getTenantId() != null ? 1 : 0);
RowKeySchema schema = table.getRowKeySchema();
int nValues = schema.getMaxFields() - offset;
Object[] values = new Object[nValues];
ImmutableBytesWritable ptr = new ImmutableBytesWritable();
int i = 0;
schema.iterator(value, ptr, offset);
while (i < nValues && schema.next(ptr, i, value.length) != null) {
values[i] = schema.getField(i).getDataType().toObject(ptr);
i++;
}
return values;
}
@Test
public void testGetSplitsWithSkipScanFilter() throws Exception {
byte[][] splits = new byte[][] {Ka1A, Ka1B, Ka1E, Ka1G, Ka1I, Ka2A};
createTestTable(getUrl(), DDL, splits, null);
Properties props = PropertiesUtil.deepCopy(TEST_PROPERTIES);
Connection conn = DriverManager.getConnection(getUrl(), props);
PhoenixConnection pconn = conn.unwrap(PhoenixConnection.class);
PTable table =
pconn.getMetaDataCache().getTable(new PTableKey(pconn.getTenantId(), TABLE_NAME));
TableRef tableRef = new TableRef(table);
List<HRegionLocation> regions =
pconn
.getQueryServices()
.getAllTableRegions(tableRef.getTable().getPhysicalName().getBytes());
List<KeyRange> ranges = getSplits(tableRef, scan, regions, scanRanges);
assertEquals(
"Unexpected number of splits: " + ranges.size(), expectedSplits.size(), ranges.size());
for (int i = 0; i < expectedSplits.size(); i++) {
assertEquals(expectedSplits.get(i), ranges.get(i));
}
}
public static ColumnResolver getResolverForCreation(
final CreateTableStatement statement, final PhoenixConnection connection)
throws SQLException {
TableName baseTable = statement.getBaseTableName();
if (baseTable == null) {
return EMPTY_TABLE_RESOLVER;
}
NamedTableNode tableNode =
NamedTableNode.create(null, baseTable, Collections.<ColumnDef>emptyList());
// Always use non-tenant-specific connection here
try {
SingleTableColumnResolver visitor =
new SingleTableColumnResolver(connection, tableNode, true);
return visitor;
} catch (TableNotFoundException e) {
// Used for mapped VIEW, since we won't be able to resolve that.
// Instead, we create a table with just the dynamic columns.
// A tenant-specific connection may not create a mapped VIEW.
if (connection.getTenantId() == null && statement.getTableType() == PTableType.VIEW) {
ConnectionQueryServices services = connection.getQueryServices();
byte[] fullTableName =
SchemaUtil.getTableNameAsBytes(baseTable.getSchemaName(), baseTable.getTableName());
HTableInterface htable = null;
try {
htable = services.getTable(fullTableName);
} catch (UnsupportedOperationException ignore) {
throw e; // For Connectionless
} finally {
if (htable != null) Closeables.closeQuietly(htable);
}
tableNode = NamedTableNode.create(null, baseTable, statement.getColumnDefs());
return new SingleTableColumnResolver(connection, tableNode, e.getTimeStamp());
}
throw e;
}
}
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);
}
};
}
@SuppressWarnings("deprecation")
public void commit() throws SQLException {
int i = 0;
byte[] tenantId = connection.getTenantId() == null ? null : connection.getTenantId().getBytes();
long[] serverTimeStamps = validate();
Iterator<Map.Entry<TableRef, Map<ImmutableBytesPtr, Map<PColumn, byte[]>>>> iterator =
this.mutations.entrySet().iterator();
List<Map.Entry<TableRef, Map<ImmutableBytesPtr, Map<PColumn, byte[]>>>> committedList =
Lists.newArrayListWithCapacity(this.mutations.size());
// add tracing for this operation
TraceScope trace = Tracing.startNewSpan(connection, "Committing mutations to tables");
Span span = trace.getSpan();
while (iterator.hasNext()) {
Map.Entry<TableRef, Map<ImmutableBytesPtr, Map<PColumn, byte[]>>> entry = iterator.next();
Map<ImmutableBytesPtr, Map<PColumn, byte[]>> valuesMap = entry.getValue();
TableRef tableRef = entry.getKey();
PTable table = tableRef.getTable();
table.getIndexMaintainers(tempPtr);
boolean hasIndexMaintainers = tempPtr.getLength() > 0;
boolean isDataTable = true;
long serverTimestamp = serverTimeStamps[i++];
Iterator<Pair<byte[], List<Mutation>>> mutationsIterator =
addRowMutations(tableRef, valuesMap, serverTimestamp, false);
while (mutationsIterator.hasNext()) {
Pair<byte[], List<Mutation>> pair = mutationsIterator.next();
byte[] htableName = pair.getFirst();
List<Mutation> mutations = pair.getSecond();
// create a span per target table
// TODO maybe we can be smarter about the table name to string here?
Span child =
Tracing.child(span, "Writing mutation batch for table: " + Bytes.toString(htableName));
int retryCount = 0;
boolean shouldRetry = false;
do {
ServerCache cache = null;
if (hasIndexMaintainers && isDataTable) {
byte[] attribValue = null;
byte[] uuidValue;
if (IndexMetaDataCacheClient.useIndexMetadataCache(
connection, mutations, tempPtr.getLength())) {
IndexMetaDataCacheClient client = new IndexMetaDataCacheClient(connection, tableRef);
cache = client.addIndexMetadataCache(mutations, tempPtr);
child.addTimelineAnnotation("Updated index metadata cache");
uuidValue = cache.getId();
// If we haven't retried yet, retry for this case only, as it's possible that
// a split will occur after we send the index metadata cache to all known
// region servers.
shouldRetry = true;
} else {
attribValue = ByteUtil.copyKeyBytesIfNecessary(tempPtr);
uuidValue = ServerCacheClient.generateId();
}
// Either set the UUID to be able to access the index metadata from the cache
// or set the index metadata directly on the Mutation
for (Mutation mutation : mutations) {
if (tenantId != null) {
mutation.setAttribute(PhoenixRuntime.TENANT_ID_ATTRIB, tenantId);
}
mutation.setAttribute(PhoenixIndexCodec.INDEX_UUID, uuidValue);
if (attribValue != null) {
mutation.setAttribute(PhoenixIndexCodec.INDEX_MD, attribValue);
}
}
}
SQLException sqlE = null;
HTableInterface hTable = connection.getQueryServices().getTable(htableName);
try {
if (logger.isDebugEnabled()) logMutationSize(hTable, mutations);
long startTime = System.currentTimeMillis();
child.addTimelineAnnotation("Attempt " + retryCount);
hTable.batch(mutations);
child.stop();
shouldRetry = false;
if (logger.isDebugEnabled())
logger.debug(
"Total time for batch call of "
+ mutations.size()
+ " mutations into "
+ table.getName().getString()
+ ": "
+ (System.currentTimeMillis() - startTime)
+ " ms");
committedList.add(entry);
} catch (Exception e) {
SQLException inferredE = ServerUtil.parseServerExceptionOrNull(e);
if (inferredE != null) {
if (shouldRetry
&& retryCount == 0
&& inferredE.getErrorCode()
== SQLExceptionCode.INDEX_METADATA_NOT_FOUND.getErrorCode()) {
// Swallow this exception once, as it's possible that we split after sending the
// index metadata
// and one of the region servers doesn't have it. This will cause it to have it the
// next go around.
// If it fails again, we don't retry.
String msg =
"Swallowing exception and retrying after clearing meta cache on connection. "
+ inferredE;
logger.warn(msg);
connection.getQueryServices().clearTableRegionCache(htableName);
// add a new child span as this one failed
child.addTimelineAnnotation(msg);
child.stop();
child = Tracing.child(span, "Failed batch, attempting retry");
continue;
}
e = inferredE;
}
// Throw to client with both what was committed so far and what is left to be committed.
// That way, client can either undo what was done or try again with what was not done.
sqlE =
new CommitException(
e,
this,
new MutationState(
committedList, this.sizeOffset, this.maxSize, this.connection));
} finally {
try {
hTable.close();
} catch (IOException e) {
if (sqlE != null) {
sqlE.setNextException(ServerUtil.parseServerException(e));
} else {
sqlE = ServerUtil.parseServerException(e);
}
} finally {
try {
if (cache != null) {
cache.close();
}
} finally {
if (sqlE != null) {
throw sqlE;
}
}
}
}
} while (shouldRetry && retryCount++ < 1);
isDataTable = false;
}
numRows -= entry.getValue().size();
iterator.remove(); // Remove batches as we process them
}
trace.close();
assert (numRows == 0);
assert (this.mutations.isEmpty());
}
private void assertIteration(String dataColumns, String pk, Object[] values, String dataProps)
throws Exception {
String schemaName = "";
String tableName = "T";
Connection conn = DriverManager.getConnection(getUrl());
String fullTableName =
SchemaUtil.getTableName(
SchemaUtil.normalizeIdentifier(schemaName), SchemaUtil.normalizeIdentifier(tableName));
conn.createStatement()
.execute(
"CREATE TABLE "
+ fullTableName
+ "("
+ dataColumns
+ " CONSTRAINT pk PRIMARY KEY ("
+ pk
+ ")) "
+ (dataProps.isEmpty() ? "" : dataProps));
PhoenixConnection pconn = conn.unwrap(PhoenixConnection.class);
PTable table = pconn.getTable(new PTableKey(pconn.getTenantId(), fullTableName));
conn.close();
StringBuilder buf = new StringBuilder("UPSERT INTO " + fullTableName + " VALUES(");
for (int i = 0; i < values.length; i++) {
buf.append("?,");
}
buf.setCharAt(buf.length() - 1, ')');
PreparedStatement stmt = conn.prepareStatement(buf.toString());
for (int i = 0; i < values.length; i++) {
stmt.setObject(i + 1, values[i]);
}
stmt.execute();
Iterator<Pair<byte[], List<KeyValue>>> iterator =
PhoenixRuntime.getUncommittedDataIterator(conn);
List<KeyValue> dataKeyValues = iterator.next().getSecond();
KeyValue keyValue = dataKeyValues.get(0);
List<SortOrder> sortOrders = Lists.newArrayListWithExpectedSize(table.getPKColumns().size());
for (PColumn col : table.getPKColumns()) {
sortOrders.add(col.getSortOrder());
}
RowKeySchema schema = table.getRowKeySchema();
int minOffset = keyValue.getRowOffset();
ImmutableBytesWritable ptr = new ImmutableBytesWritable();
int nExpectedValues = values.length;
for (int i = values.length - 1; i >= 0; i--) {
if (values[i] == null) {
nExpectedValues--;
} else {
break;
}
}
int i = 0;
int maxOffset =
schema.iterator(keyValue.getRowArray(), minOffset, keyValue.getRowLength(), ptr);
for (i = 0; i < schema.getFieldCount(); i++) {
Boolean hasValue = schema.next(ptr, i, maxOffset);
if (hasValue == null) {
break;
}
assertTrue(hasValue);
PDataType type = PDataType.fromLiteral(values[i]);
SortOrder sortOrder = sortOrders.get(i);
Object value = type.toObject(ptr, schema.getField(i).getDataType(), sortOrder);
assertEquals(values[i], value);
}
assertEquals(nExpectedValues, i);
assertNull(schema.next(ptr, i, maxOffset));
for (i--; i >= 0; i--) {
Boolean hasValue = schema.previous(ptr, i, minOffset);
if (hasValue == null) {
break;
}
assertTrue(hasValue);
PDataType type = PDataType.fromLiteral(values[i]);
SortOrder sortOrder = sortOrders.get(i);
Object value = type.toObject(ptr, schema.getField(i).getDataType(), sortOrder);
assertEquals(values[i], value);
}
assertEquals(-1, i);
assertNull(schema.previous(ptr, i, minOffset));
}