Creating a Table from a VectorSchemaRoot¶

Tables are created from a VectorSchemaRoot as shown below. The memory buffers holding the data are transferred from the vector schema root to new vectors in the new table, clearing the source vectors in the process. This ensures that the data in your new table is never changed. Since the buffers are transferred rather than copied, this is a very low overhead operation.

Table t = new Table(someVectorSchemaRoot);

If you now update the vectors held by the VectorSchemaRoot (using some version of ValueVector#setSafe()), it would reflect those changes, but the values in table t are unchanged.

Creating a Table from FieldVectors¶

Tables can be created from FieldVectors as shown below, using ‘var-arg’ array arguments:

IntVector myVector = createMyIntVector();
VectorSchemaRoot vsr1 = new VectorSchemaRoot(myVector);

or by passing a collection:

IntVector myVector = createMyIntVector();
List<FieldVector> fvList = List.of(myVector);
VectorSchemaRoot vsr1 = new VectorSchemaRoot(fvList);

It is rarely a good idea to share vectors between multiple vector schema roots, and it would not be a good idea to share them between vector schema roots and tables. Creating a VectorSchemaRoot from a list of vectors does not cause the reference counts for the vectors to be incremented. Unless you manage the counts manually, the code below would lead to more references than reference counts, and that could lead to trouble. There is an implicit assumption that the vectors were created for use by one VectorSchemaRoot that this code violates.

Don’t do this:

IntVector myVector = createMyIntVector();  // Reference count for myVector = 1
VectorSchemaRoot vsr1 = new VectorSchemaRoot(myVector); // Still one reference
VectorSchemaRoot vsr2 = new VectorSchemaRoot(myVector);
// Ref count is still one, but there are two VSRs with a reference to myVector
vsr2.clear(); // Reference count for myVector is 0.

What is happening is that the reference counter works at a lower level than the VectorSchemaRoot interface. A reference counter counts references to ArrowBuf instances that control memory buffers. It doesn’t count references to the vectors that hold those ArrowBufs. In the example above, each ArrowBuf is held by one vector, so there is only one reference. This distinction is blurred when you call the VectorSchemaRoot’s clear() method, which frees the memory held by each of the vectors it references even though another instance references the same vectors.

When you create tables from vectors, it’s assumed that there are no external references to those vectors. To be certain, the buffers underlying these vectors are transferred to new vectors in the new table, and the original vectors are cleared.

Don’t do this either, but note the difference from above:

IntVector myVector = createMyIntVector(); // Reference count for myVector = 1
Table t1 = new Table(myVector);
// myVector is cleared; Table t1 has a new hidden vector with the data from myVector
Table t2 = new Table(myVector);
// t2 has no rows because myVector was just cleared
// t1 continues to have the data from the original vector
t2.clear();
// no change because t2 is already empty and t1 is independent

With tables, memory is explicitly transferred on instantiation so the buffers held by a table are held by only that table.

Creating Tables with dictionary-encoded vectors¶

Another point of difference is that VectorSchemaRoot is uninformed about any dictionary-encoding of its vectors, while tables hold an optional DictionaryProvider instance. If any vectors in the source data are encoded, a DictionaryProvider must be set to un-encode the values.

VectorSchemaRoot vsr = myVsr();
DictionaryProvider provider = myProvider();
Table t = new Table(vsr, provider);

In Table, dictionaries are used like they are with vectors. To decode a vector, the user provides the name of the vector to decode and the dictionary id:

Table t = new Table(vsr, provider);
ValueVector decodedName = t.decode("name", 1L);

To encode a vector from a table, a similar approach is used:

Table t = new Table(vsr, provider);
ValueVector encodedName = t.encode("name", 1L);

Freeing memory explicitly¶

Tables use off-heap memory that must be freed when it is no longer needed. Table implements AutoCloseable so the best way to create one is in a try-with-resources block:

try (VectorSchemaRoot vsr = myMethodForGettingVsrs();
    Table t = new Table(vsr)) {
    // do useful things.
}

If you don’t use a try-with-resources block, you must close the table manually:

try {
    VectorSchemaRoot vsr = myMethodForGettingVsrs();
    Table t = new Table(vsr);
    // do useful things.
} finally {
    vsr.close();
    t.close();
}

Manual closing should be performed in a finally block.

Getting the schema¶

You get the table’s schema just as you would with a vector schema root:

Schema s = table.getSchema();

Adding and removing vectors¶

Table provides facilities for adding and removing vectors modeled on the same functionality in VectorSchemaRoot. These operations return new instances rather than modifying the original instance in-place.

try (Table t = new Table(vectorList)) {
    IntVector v3 = new IntVector("3", intFieldType, allocator);
    Table t2 = t.addVector(2, v3);
    Table t3 = t2.removeVector(1);
    // don't forget to close t2 and t3
}

Slicing tables¶

Table supports slice() operations, where a slice of a source table is a second Table that refers to a single, contiguous range of rows in the source.

try (Table t = new Table(vectorList)) {
    Table t2 = t.slice(100, 200); // creates a slice referencing the values in range (100, 200]
    ...
}

This raises the question: If you create a slice with all the values in the source table (as shown below), how would that differ from a new Table constructed with the same vectors as the source?

try (Table t = new Table(vectorList)) {
    Table t2 = t.slice(0, t.getRowCount()); // creates a slice referencing all the values in t
    // ...
}

The difference is that when you construct a new table, the buffers are transferred from the source vectors to new vectors in the destination. With a slice, both tables share the same underlying vectors. That’s OK, though, since both tables are immutable.

Using FieldReaders¶

You can get a FieldReader for any vector in the Table passing either the Field, vector index, or vector name as an argument. The signatures are the same as in VectorSchemaRoot.

FieldReader nameReader = table.getReader("user_name");

Row operations¶

Row-based access is supported by the Row object. Row provides get() methods by both vector name and vector position, but no set() operations.

It is important to recognize that rows are NOT reified as objects, but rather operate like a cursor where the data from numerous logical rows in the table can be viewed (one at a time) using the same Row instance. See “Moving from row-to-row” below for information about navigating through the table.

Getting a row¶

Calling immutableRow() on any table instance returns a new Row instance.

Row r = table.immutableRow();

Moving from row-to-row¶

Since rows are iterable, you can traverse a table using a standard while loop:

Row r = table.immutableRow();
while (r.hasNext()) {
  r.next();
  // do something useful here
}

Table implements Iterable<Row> so you can access rows directly from a table in an enhanced for loop:

for (Row row: table) {
  int age = row.getInt("age");
  boolean nameIsNull = row.isNull("name");
  ...
}

Finally, while rows are usually iterated in the order of the underlying data vectors, but they are also positionable using the Row#setPosition() method, so you can skip to a specific row. Row numbers are 0-based.

Row r = table.immutableRow();
int age101 = r.setPosition(101); // change position directly to 101

Any changes to position are applied to all the columns in the table.

Note that you must call next(), or setPosition() before accessing values via a row. Failure to do so results in a runtime exception.

Read operations using rows¶

Methods are available for getting values by vector name and vector index, where index is the 0-based position of the vector in the table. For example, assuming ‘age’ is the 13th vector in ‘table’, the following two gets are equivalent:

Row r = table.immutableRow();
r.next(); // position the row at the first value
int age1 = r.get("age"); // gets the value of vector named 'age' in the table at row 0
int age2 = r.get(12);    // gets the value of the 13th vector in the table at row 0

You can also get value using a nullable ValueHolder. For example:

NullableIntHolder holder = new NullableIntHolder();
int b = row.getInt("age", holder);

This can be used to retrieve values without creating a new Object for each.

In addition to getting values, you can check if a value is null using isNull(). This is important if the vector contains any nulls, as asking for a value from a vector can cause NullPointerExceptions in some cases.

boolean name0isNull = row.isNull("name");

You can also get the current row number:

int row = row.getRowNumber();

Reading values as Objects¶

For any given vector type, the basic get() method returns a primitive value wherever possible. For example, getTimeStampMicro() returns a long value that encodes the timestamp. To get the LocalDateTime object representing that timestamp in Java, another method with ‘Obj’ appended to the name is provided. For example:

long ts = row.getTimeStampMicro();
LocalDateTime tsObject = row.getTimeStampMicroObj();

The exception to this naming scheme is for complex vector types (List, Map, Schema, Union, DenseUnion, and ExtensionType). These always return objects rather than primitives so no “Obj” extension is required. It is expected that some users may subclass Row to add getters that are more specific to their needs.

Reading VarChars and LargeVarChars¶

Strings in arrow are represented as byte arrays encoded with the UTF-8 charset. You can get either a String result or the actual byte array.

byte[] b = row.getVarChar("first_name");
String s = row.getVarCharObj("first_name");       // uses the default encoding (UTF-8)

Converting a Table to a VectorSchemaRoot¶

Tables can be converted to vector schema roots using the toVectorSchemaRoot() method. Buffers are transferred to the vector schema root and the source table is cleared.

VectorSchemaRoot root = myTable.toVectorSchemaRoot();

Working with the C-Data interface¶

The ability to work with native code is required for many Arrow features. This section describes how tables can be be exported for use with native code

Exporting works by converting the data to a VectorSchemaRoot instance and using the existing facilities to transfer the data. You could do it yourself, but that isn’t ideal because conversion to a vector schema root breaks the immutability guarantees. Using the exportTable() methods in the Data class avoids this concern.

Data.exportTable(bufferAllocator, table, dictionaryProvider, outArrowArray);

If the table contains dictionary-encoded vectors and was constructed with a DictionaryProvider, the provider argument to exportTable() can be omitted and the table’s provider attribute will be used:

Data.exportTable(bufferAllocator, table, outArrowArray);