public UpsertCompiler(PhoenixStatement statement) {
this.statement = statement;
}
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);
List<ColumnName> columnNodes = upsert.getColumns();
TableRef tableRefToBe = null;
PTable table = null;
Set<PColumn> addViewColumnsToBe = Collections.emptySet();
Set<PColumn> overlapViewColumnsToBe = Collections.emptySet();
List<PColumn> allColumnsToBe = Collections.emptyList();
boolean isTenantSpecific = false;
boolean isSharedViewIndex = false;
String tenantId = null;
ColumnResolver resolver = null;
int[] columnIndexesToBe;
int nColumnsToSet = 0;
int[] pkSlotIndexesToBe;
List<ParseNode> valueNodes = upsert.getValues();
List<PColumn> targetColumns;
NamedTableNode tableNode = upsert.getTable();
String tableName = tableNode.getName().getTableName();
String schemaName = tableNode.getName().getSchemaName();
QueryPlan queryPlanToBe = null;
int nValuesToSet;
boolean sameTable = false;
boolean runOnServer = false;
UpsertingParallelIteratorFactory parallelIteratorFactoryToBe = null;
// Retry once if auto commit is off, as the meta data may
// be out of date. We do not retry if auto commit is on, as we
// update the cache up front when we create the resolver in that case.
boolean retryOnce = !connection.getAutoCommit();
while (true) {
try {
resolver = FromCompiler.getResolverForMutation(upsert, connection);
tableRefToBe = resolver.getTables().get(0);
table = tableRefToBe.getTable();
if (table.getType() == PTableType.VIEW) {
if (table.getViewType().isReadOnly()) {
throw new ReadOnlyTableException(schemaName,tableName);
}
}
boolean isSalted = table.getBucketNum() != null;
isTenantSpecific = table.isMultiTenant() && connection.getTenantId() != null;
isSharedViewIndex = table.getViewIndexId() != null;
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
allColumnsToBe = table.getColumns();
nColumnsToSet = 0;
if (table.getViewType() == ViewType.UPDATABLE) {
addViewColumnsToBe = Sets.newLinkedHashSetWithExpectedSize(allColumnsToBe.size());
for (PColumn column : allColumnsToBe) {
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 = allColumnsToBe.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 < allColumnsToBe.size(); i++) {
PColumn column = allColumnsToBe.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");
}
}
}
}
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
&& tableRefToBe.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 (which requires a client-side final merge)
* 3) autoCommit is on
* 4) the table is not immutable with indexes, as the client is the one that figures out the additional
* puts for index tables.
* 5) no limit clause, as the limit clause requires client-side post processing
* 6) no sequences, as sequences imply that the order of upsert must match the order of
* selection.
* Otherwise, run the query to pull the data from the server
* and populate the MutationState (upto a limit).
*/
if (! (select.isAggregate() || select.isDistinct() || select.getLimit() != null || select.hasSequence()) ) {
// We can pipeline the upsert select instead of spooling everything to disk first,
// if we don't have any post processing that's required.
parallelIteratorFactoryToBe = new UpsertingParallelIteratorFactory(connection, tableRefToBe);
// If we're in the else, then it's not an aggregate, distinct, limted, or sequence using query,
// so we might be able to run it entirely on the server side.
runOnServer = sameTable && isAutoCommit && !(table.isImmutableRows() && !table.getIndexes().isEmpty());
}
// 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
try {
QueryCompiler compiler = new QueryCompiler(statement, select, selectResolver, targetColumns, parallelIteratorFactoryToBe, new SequenceManager(statement));
queryPlanToBe = compiler.compile();
} catch (MetaDataEntityNotFoundException e) {
retryOnce = false; // don't retry if select clause has meta data entities that aren't found, as we already updated the cache
throw e;
}
nValuesToSet = queryPlanToBe.getProjector().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);
}
// 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) {
// We might have added columns, so refresh cache and try again if stale.
// Note that this check is not really sufficient, as a column could have
// been removed and the added back and we wouldn't detect that here.
if (retryOnce) {
retryOnce = false;
if (new MetaDataClient(connection).updateCache(schemaName, tableName).wasUpdated()) {
continue;
}
}
throw new SQLExceptionInfo.Builder(SQLExceptionCode.UPSERT_COLUMN_NUMBERS_MISMATCH)
.setMessage("Numbers of columns: " + nColumnsToSet + ". Number of values: " + nValuesToSet)
.build().buildException();
}
} catch (MetaDataEntityNotFoundException e) {
// Catch column/column family not found exception, as our meta data may
// be out of sync. Update the cache once and retry if we were out of sync.
// Otherwise throw, as we'll just get the same error next time.
if (retryOnce) {
retryOnce = false;
if (new MetaDataClient(connection).updateCache(schemaName, tableName).wasUpdated()) {
continue;
}
}
throw e;
}
break;
}
RowProjector projectorToBe = null;
// Optimize only after all checks have been performed
if (valueNodes == null) {
queryPlanToBe = new QueryOptimizer(services).optimize(queryPlanToBe, statement, targetColumns, parallelIteratorFactoryToBe);
projectorToBe = queryPlanToBe.getProjector();
runOnServer &= queryPlanToBe.getTableRef().equals(tableRefToBe);
}
final List<PColumn> allColumns = allColumnsToBe;
final RowProjector projector = projectorToBe;
final QueryPlan queryPlan = queryPlanToBe;
final TableRef tableRef = tableRefToBe;
final int[] columnIndexes = columnIndexesToBe;
final int[] pkSlotIndexes = pkSlotIndexesToBe;
final Set<PColumn> addViewColumns = addViewColumnsToBe;
final Set<PColumn> overlapViewColumns = overlapViewColumnsToBe;
final UpsertingParallelIteratorFactory parallelIteratorFactory = parallelIteratorFactoryToBe;
// 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 (parallelIteratorFactory == null) {
return upsertSelect(statement, tableRef, projector, iterator, columnIndexes, pkSlotIndexes);
}
parallelIteratorFactory.setRowProjector(projector);
parallelIteratorFactory.setColumnIndexes(columnIndexes);
parallelIteratorFactory.setPkSlotIndexes(pkSlotIndexes);
Tuple tuple;
long totalRowCount = 0;
while ((tuple=iterator.next()) != null) {// Runs query
KeyValue kv = tuple.getValue(0);
totalRowCount += PDataType.LONG.getCodec().decodeLong(kv.getBuffer(), 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;