use of org.apache.drill.exec.vector.accessor.writer.UnionWriterImpl in project drill by apache.
the class BaseWriterBuilder method buildUnion.
private AbstractObjectWriter buildUnion(UnionVector vector, VectorDescrip descrip) {
if (vector == null) {
throw new UnsupportedOperationException("Dummy variant writer not yet supported");
}
final AbstractObjectWriter[] variants = new AbstractObjectWriter[MinorType.values().length];
final MetadataProvider mdProvider = descrip.childProvider();
int i = 0;
for (final MinorType type : vector.getField().getType().getSubTypeList()) {
// This call will create the vector if it does not yet exist.
// Will throw an exception for unsupported types.
// so call this only if the MajorType reports that the type
// already exists.
final ValueVector memberVector = vector.getMember(type);
final VectorDescrip memberDescrip = new VectorDescrip(mdProvider, i++, memberVector.getField());
variants[type.ordinal()] = buildVectorWriter(memberVector, memberDescrip);
}
return new VariantObjectWriter(new UnionWriterImpl(descrip.metadata, vector, variants));
}
use of org.apache.drill.exec.vector.accessor.writer.UnionWriterImpl in project drill by apache.
the class ColumnBuilder method buildUnionList.
/**
* Create a list based on a (possible) union. The list starts empty here. The client
* can then add types as they are discovered. The list itself will transition from a
* list of nulls (no child type), to a list of a single type, to a list of unions.
* The writer interface will consistently present the list as a list of unions, even
* when the list itself has no subtype or a single subtype.
* <p>
* List vectors (lists of unions) are not supported in
* Drill. The code here works up through the scan operator. But, other operators do
* not support the {@code ListVector} type.
*
* @param parent the parent (tuple, union or list) that holds this list
* @param columnSchema description of the list (must be empty of
* subtypes)
* @return the column state for the list
*/
private ColumnState buildUnionList(ContainerState parent, ColumnMetadata columnSchema) {
// The variant must start out empty.
assert columnSchema.variantSchema().size() == 0;
// Create the union writer, bound to an empty list shim.
final UnionWriterImpl unionWriter = new UnionWriterImpl(columnSchema);
unionWriter.bindShim(new EmptyListShim());
final VariantObjectWriter unionObjWriter = new VariantObjectWriter(unionWriter);
// Create the list vector. Starts with the default (dummy) data
// vector which corresponds to the empty union shim above.
// Don't get the list vector from the vector cache. List vectors may
// have content that varies from batch to batch. Only the leaf
// vectors can be cached.
final ListVector listVector = new ListVector(columnSchema.schema(), parent.loader().allocator(), null);
// Create the list vector state that tracks the list vector lifecycle.
final ListVectorState vectorState = new ListVectorState(unionWriter, listVector);
// Create the list writer: an array of unions.
final AbstractObjectWriter listWriter = new ArrayObjectWriter(new ListWriterImpl(columnSchema, listVector, unionObjWriter));
// Create the manager for the columns within the list (which may or
// may not be grouped into a union.)
final ListState listState = new ListState(parent.loader(), parent.vectorCache().childCache(columnSchema.name()));
// Bind the union state to the union writer to handle column additions.
unionWriter.bindListener(listState);
// Assemble it all into a union column state.
return new UnionColumnState(parent.loader(), listWriter, vectorState, listState);
}
use of org.apache.drill.exec.vector.accessor.writer.UnionWriterImpl in project drill by apache.
the class ColumnBuilder method buildUnion.
/**
* Builds a union column.
* <p>
* The union vector type is not well supported in Drill. The idea is that
* arbitrary operators can absorb schema changes by converting vectors to
* unions so that an operator can handle, say, a nullable int and a varchar.
* In practice, most operators don't support this feature. (Sort does -- but
* does not manage memory for the union case.) In principal, union can't solve
* the problem because ODBC and JDBC don't support unions, and it is easy
* to envision changes that unions won't solve (int and varchar types combining
* in a join column, say.) Still, Drill supports unions, so the code here
* does so. Unions are fully tested in the row set writer mechanism.
*
* @param parent container of vectors
* @param columnSchema implied projection type for the column
* @return column
*/
private ColumnState buildUnion(ContainerState parent, ColumnMetadata columnSchema) {
assert columnSchema.isVariant() && !columnSchema.isArray();
// vectors can be cached.
assert columnSchema.variantSchema().size() == 0;
final UnionVector vector = new UnionVector(columnSchema.schema(), parent.loader().allocator(), null);
// Then the union writer.
final UnionWriterImpl unionWriter = new UnionWriterImpl(columnSchema, vector, null);
final VariantObjectWriter writer = new VariantObjectWriter(unionWriter);
// The union vector state which manages the types vector.
final UnionVectorState vectorState = new UnionVectorState(vector, unionWriter);
// Create the manager for the columns within the union.
final UnionState unionState = new UnionState(parent.loader(), parent.vectorCache().childCache(columnSchema.name()));
// Bind the union state to the union writer to handle column additions.
unionWriter.bindListener(unionState);
// Assemble it all into a union column state.
return new UnionColumnState(parent.loader(), writer, vectorState, unionState);
}
use of org.apache.drill.exec.vector.accessor.writer.UnionWriterImpl in project drill by apache.
the class ListState method addSecondType.
/**
* Perform the delicate dance of promoting a list vector from a single type to
* a union, while leaving the writer client blissfully ignorant that the underlying
* vector representation just did a radical change. Key tasks:
* <ul>
* <li>Create the new column (type member) requested by the client.</li>
* <li>The List vector currently has a single type. Promote the list to
* a union, adding the existing type (column) as the first union member.</li>
* <li>Initialize the union's type vector with either the type of the existing
* column, or null, depending on the setting of the is-set bits in the existing
* column vector.</li>
* <li>Since we've written values into the union's type vector, mark the
* last-write position in the union vector's type vector writer to reflect
* these writes. (Otherwise, the writer will helpfully zero-fill the previous
* positions as part of it's back-fill handling.</li>
* <li>Replace the single-type shim in the union vector with a full union
* shim.</li>
* <li>Move the existing column writer or the member column across from the
* single-writer shim to the new union shim.</li>
* <li>Augment the list vector's vector state to include a vector state for
* the newly created union vector.</li>
* </ul>
* <p>
* Here, yet again, an editorial comment might be useful. List vectors are
* very strange and not at all well designed for high-speed writing. They are
* too complex; too much can go wrong and there are too many states to handle.
* Not only that, variant types don't play well with a relational model like
* SQL. This code works, but the overall list concept really needs rethinking.
*
* @param colState the column state for the newly added type column; the
* one causing the list to change from single-type to a union
*/
private void addSecondType(ColumnState colState) {
final UnionWriterImpl unionWriter = unionWriter();
final ListVector listVector = listVector();
// Going from one type to a union
// Convert the list from single type to a union,
// moving across the previous type vector.
final int typeFillCount = unionWriter.elementPosition().writeIndex();
final UnionVector unionVector = listVector.convertToUnion(innerCardinality(), typeFillCount);
unionVector.addType(colState.vector());
// Replace the single-type shim with a union shim, copying
// across the existing writer.
final SimpleListShim oldShim = (SimpleListShim) unionWriter.shim();
final UnionVectorShim newShim = new UnionVectorShim(unionVector);
unionWriter.bindShim(newShim);
newShim.addMemberWriter(oldShim.memberWriter());
newShim.initTypeIndex(typeFillCount);
// The union vector will be managed within the list vector state.
// (Do this last because the union vector state expects the union
// writer to be operating in "union mode".
listVectorState().replaceMember(new UnionVectorState(unionVector, unionWriter));
}
use of org.apache.drill.exec.vector.accessor.writer.UnionWriterImpl in project drill by apache.
the class TestResultSetLoaderUnions method testVariantListDynamic.
/**
* Test a variant list created dynamically at load time.
* The list starts with no type, at which time it can hold
* only null values. Then we add a Varchar, and finally an
* Int.
* <p>
* This test is superficial. There are many odd cases to consider.
* <ul>
* <li>Write nulls to a list with no type. (This test ensures that
* adding a (nullable) scalar "does the right thing."</li>
* <li>Add a map to the list. Maps carry no "bits" vector, so null
* list entries to that point are lost. (For maps, we could go straight
* to a union, with just a map, to preserve the null states. This whole
* area is a huge mess...)</li>
* <li>Do the type transitions when writing to a row. (The tests here
* do the transition between rows.)</li>
* </ul>
*
* The reason for the sparse coverage is that Drill barely supports lists
* and unions; most code is just plain broken. Our goal here is not to fix
* all those problems, just to leave things no more broken than before.
*/
@Test
public void testVariantListDynamic() {
final ResultSetLoader rsLoader = new ResultSetLoaderImpl(fixture.allocator());
final RowSetLoader writer = rsLoader.writer();
// Can write a batch as if this was a repeated Varchar, except
// that any value can also be null.
rsLoader.startBatch();
writer.addColumn(MaterializedField.create("id", Types.required(MinorType.INT)));
writer.addColumn(MaterializedField.create("list", Types.optional(MinorType.LIST)));
// Sanity check: should be an array of variants because we said the
// types within the list are expandable (which is the default.)
final ArrayWriter arrWriter = writer.array("list");
assertEquals(ObjectType.VARIANT, arrWriter.entryType());
final VariantWriter variant = arrWriter.variant();
// We need to verify that the internal state is what we expect, so
// the next assertion peeks inside the private bits of the union
// writer. No client code should ever need to do this, of course.
assertTrue(((UnionWriterImpl) variant).shim() instanceof EmptyListShim);
// No types, so all we can do is add a null list, or a list of nulls.
writer.addRow(1, null).addRow(2, variantArray()).addRow(3, variantArray(null, null));
// Add a String. Now we can create a list of strings and/or nulls.
variant.addMember(MinorType.VARCHAR);
assertTrue(variant.hasType(MinorType.VARCHAR));
// Sanity check: sniff inside to ensure that the list contains a single
// type.
assertTrue(((UnionWriterImpl) variant).shim() instanceof SimpleListShim);
assertTrue(((ListWriterImpl) arrWriter).vector().getDataVector() instanceof NullableVarCharVector);
writer.addRow(4, variantArray("fred", null, "barney"));
// Add an integer. The list vector should be promoted to union.
// Now we can add both types.
variant.addMember(MinorType.INT);
// Sanity check: sniff inside to ensure promotion to union occurred
assertTrue(((UnionWriterImpl) variant).shim() instanceof UnionVectorShim);
assertTrue(((ListWriterImpl) arrWriter).vector().getDataVector() instanceof UnionVector);
writer.addRow(5, variantArray("wilma", null, 30));
// Verify
final RowSet result = fixture.wrap(rsLoader.harvest());
final TupleMetadata schema = new SchemaBuilder().add("id", MinorType.INT).addList("list").addType(MinorType.VARCHAR).addType(MinorType.INT).resumeSchema().buildSchema();
final SingleRowSet expected = fixture.rowSetBuilder(schema).addRow(1, null).addRow(2, variantArray()).addRow(3, variantArray(null, null)).addRow(4, variantArray("fred", null, "barney")).addRow(5, variantArray("wilma", null, 30)).build();
RowSetUtilities.verify(expected, result);
}
Aggregations