ValueVector#

ValueVector interface (which called Array in C++ implementation and the the specification) is an abstraction that is used to store a sequence of values having the same type in an individual column. Internally, those values are represented by one or several buffers, the number and meaning of which depend on the vector’s data type.

There are concrete subclasses of ValueVector for each primitive data type and nested type described in the specification. There are a few differences in naming with the type names described in the specification: Table with non-intuitive names (BigInt = 64 bit integer, etc).

It is important that vector is allocated before attempting to read or write, ValueVector “should” strive to guarantee this order of operation: create > allocate > mutate > set value count > access > clear (or allocate to start the process over). We will go through a concrete example to demonstrate each operation in the next section.

Vector Life Cycle#

As discussed above, each vector goes through several steps in its life cycle, and each step is triggered by a vector operation. In particular, we have the following vector operations:

1. Vector creation: we create a new vector object by, for example, the vector constructor. The following code creates a new IntVector by the constructor:

RootAllocator allocator = new RootAllocator(Long.MAX_VALUE);
...
IntVector vector = new IntVector("int vector", allocator);

By now, a vector object is created. However, no underlying memory has been allocated, so we need the following step.

2. Vector allocation: in this step, we allocate memory for the vector. For most vectors, we have two options: 1) if we know the maximum vector capacity, we can specify it by calling the allocateNew(int) method; 2) otherwise, we should call the allocateNew() method, and a default capacity will be allocated for it. For our running example, we assume that the vector capacity never exceeds 10:

vector.allocateNew(10);

3. Vector mutation: now we can populate the vector with values we desire. For all vectors, we can populate vector values through vector writers (An example will be given in the next section). For primitive types, we can also mutate the vector by the set methods. There are two classes of set methods: 1) if we can be sure the vector has enough capacity, we can call the set(index, value) method. 2) if we are not sure about the vector capacity, we should call the setSafe(index, value) method, which will automatically take care of vector reallocation, if the capacity is not sufficient. For our running example, we know the vector has enough capacity, so we can call

vector.set(/*index*/5, /*value*/25);

4. Set value count: for this step, we set the value count of the vector by calling the setValueCount(int) method:

vector.setValueCount(10);

After this step, the vector enters an immutable state. In other words, we should no longer mutate it. (Unless we reuse the vector by allocating it again. This will be discussed shortly.)

5. Vector access: it is time to access vector values. Similarly, we have two options to access values: 1) get methods and 2) vector reader. Vector reader works for all types of vectors, while get methods are only available for primitive vectors. A concrete example for vector reader will be given in the next section. Below is an example of vector access by get method:

int value = vector.get(5);  // value == 25

6. Vector clear: when we are done with the vector, we should clear it to release its memory. This is done by calling the close() method:

vector.close();

Some points to note about the steps above:

  • The steps are not necessarily performed in a linear sequence. Instead, they can be in a loop. For example, when a vector enters the access step, we can also go back to the vector mutation step, and then set value count, access vector, and so on.

  • We should try to make sure the above steps are carried out in order. Otherwise, the vector may be in an undefined state, and some unexpected behavior may occur. However, this restriction is not strict. That means it is possible that we violates the order above, but still get correct results.

  • When mutating vector values through set methods, we should prefer set(index, value) methods to setSafe(index, value) methods whenever possible, to avoid unnecessary performance overhead of handling vector capacity.

  • All vectors implement the AutoCloseable interface. So they must be closed explicitly when they are no longer used, to avoid resource leak. To make sure of this, it is recommended to place vector related operations into a try-with-resources block.

  • For fixed width vectors (e.g. IntVector), we can set values at different indices in arbitrary orders. For variable width vectors (e.g. VarCharVector), however, we must set values in non-decreasing order of the indices. Otherwise, the values after the set position will become invalid. For example, suppose we use the following statements to populate a variable width vector:

VarCharVector vector = new VarCharVector("vector", allocator);
vector.allocateNew();
vector.setSafe(0, "zero");
vector.setSafe(1, "one");
...
vector.setSafe(9, "nine");

Then we set the value at position 5 again:

vector.setSafe(5, "5");

After that, the values at positions 6, 7, 8, and 9 of the vector will become invalid.

Building ValueVector#

Note that the current implementation doesn’t enforce the rule that Arrow objects are immutable. ValueVector instances could be created directly by using new keyword, there are set/setSafe APIs and concrete subclasses of FieldWriter for populating values.

For example, the code below shows how to build a BigIntVector, in this case, we build a vector of the range 0 to 7 where the element that should hold the fourth value is nulled

try (BufferAllocator allocator = new RootAllocator(Long.MAX_VALUE);
  BigIntVector vector = new BigIntVector("vector", allocator)) {
  vector.allocateNew(8);
  vector.set(0, 1);
  vector.set(1, 2);
  vector.set(2, 3);
  vector.setNull(3);
  vector.set(4, 5);
  vector.set(5, 6);
  vector.set(6, 7);
  vector.set(7, 8);
  vector.setValueCount(8); // this will finalizes the vector by convention.
  ...
}

The BigIntVector holds two ArrowBufs. The first buffer holds the null bitmap, which consists here of a single byte with the bits 1|1|1|1|0|1|1|1 (the bit is 1 if the value is non-null). The second buffer contains all the above values. As the fourth entry is null, the value at that position in the buffer is undefined. Note compared with set API, setSafe API would check value capacity before setting values and reallocate buffers if necessary.

Here is how to build a vector using writer

try (BigIntVector vector = new BigIntVector("vector", allocator);
  BigIntWriter writer = new BigIntWriterImpl(vector)) {
  writer.setPosition(0);
  writer.writeBigInt(1);
  writer.setPosition(1);
  writer.writeBigInt(2);
  writer.setPosition(2);
  writer.writeBigInt(3);
  // writer.setPosition(3) is not called which means the fourth value is null.
  writer.setPosition(4);
  writer.writeBigInt(5);
  writer.setPosition(5);
  writer.writeBigInt(6);
  writer.setPosition(6);
  writer.writeBigInt(7);
  writer.setPosition(7);
  writer.writeBigInt(8);
}

There are get API and concrete subclasses of FieldReader for accessing vector values, what needs to be declared is that writer/reader is not as efficient as direct access

// access via get API
for (int i = 0; i < vector.getValueCount(); i++) {
  if (!vector.isNull(i)) {
    System.out.println(vector.get(i));
  }
}

// access via reader
BigIntReader reader = vector.getReader();
for (int i = 0; i < vector.getValueCount(); i++) {
  reader.setPosition(i);
  if (reader.isSet()) {
    System.out.println(reader.readLong());
  }
}

Building ListVector#

A ListVector is a vector that holds a list of values for each index. Working with one you need to handle the same steps as mentioned above (create > allocate > mutate > set value count > access > clear), but the details of how you accomplish this are slightly different since you need to both create the vector and set the list of values for each index.

For example, the code below shows how to build a ListVector of int’s using the writer UnionListWriter. We build a vector from 0 to 9 and each index contains a list with values [[0, 0, 0, 0, 0], [0, 1, 2, 3, 4], [0, 2, 4, 6, 8], …, [0, 9, 18, 27, 36]]. List values can be added in any order so writing a list such as [3, 1, 2] would be just as valid.

try (BufferAllocator allocator = new RootAllocator(Long.MAX_VALUE);
  ListVector listVector = ListVector.empty("vector", allocator)) {
  UnionListWriter writer = listVector.getWriter();
  for (int i = 0; i < 10; i++) {
     writer.startList();
     writer.setPosition(i);
     for (int j = 0; j < 5; j++) {
         writer.writeInt(j * i);
     }
     writer.setValueCount(5);
     writer.endList();
  }
  listVector.setValueCount(10);
}

ListVector values can be accessed either through the get API or through the reader class UnionListReader. To read all the values, first enumerate through the indexes, and then enumerate through the inner list values.

// access via get API
for (int i = 0; i < listVector.getValueCount(); i++) {
   if (!listVector.isNull(i)) {
       ArrayList<Integer> elements = (ArrayList<Integer>) listVector.getObject(i);
       for (Integer element : elements) {
           System.out.println(element);
       }
   }
}

// access via reader
UnionListReader reader = listVector.getReader();
for (int i = 0; i < listVector.getValueCount(); i++) {
   reader.setPosition(i);
   while (reader.next()) {
       IntReader intReader = reader.reader();
       if (intReader.isSet()) {
           System.out.println(intReader.readInteger());
       }
   }
}

Dictionary Encoding#

Dictionary encoding is a form of compression where values of one type are replaced by values of a smaller type: an array of ints replacing an array of strings is a common example. The mapping between the original values and the replacements is held in a ‘dictionary’. Since the dictionary needs only one copy of each of the longer values, the combination of the dictionary and the array of smaller values may use less memory. The more repetitive the original data, the greater the savings.

A FieldVector can be dictionary encoded for performance or improved memory efficiency. Nearly any type of vector might be encoded if there are many values, but few unique values.

There are a few steps involved in the encoding process:

  1. Create a regular, un-encoded vector and populate it

  2. Create a dictionary vector of the same type as the un-encoded vector. This vector must have the same values, but each unique value in the un-encoded vector need appear here only once.

  3. Create a Dictionary. It will contain the dictionary vector, plus a DictionaryEncoding object that holds the encoding’s metadata and settings values.

  4. Create a DictionaryEncoder.

  5. Call the encode() method on the DictionaryEncoder to produce an encoded version of the original vector.

  6. (Optional) Call the decode() method on the encoded vector to re-create the original values.

The encoded values will be integers. Depending on how many unique values you have, you can use TinyIntVector, SmallIntVector, IntVector, or BigIntVector to hold them. You specify the type when you create your DictionaryEncoding instance. You might wonder where those integers come from: the dictionary vector is a regular vector, so the value’s index position in that vector is used as its encoded value.

Another critical attribute in DictionaryEncoding is the id. It’s important to understand how the id is used, so we cover that later in this section.

This result will be a new vector (for example, an IntVector) that can act in place of the original vector (for example, a VarCharVector). When you write the data in arrow format, it is both the new IntVector plus the dictionary that is written: you will need the dictionary later to retrieve the original values.

// 1. create a vector for the un-encoded data and populate it
VarCharVector unencoded = new VarCharVector("unencoded", allocator);
// now put some data in it before continuing

// 2. create a vector to hold the dictionary and populate it
VarCharVector dictionaryVector = new VarCharVector("dictionary", allocator);

// 3. create a dictionary object
Dictionary dictionary = new Dictionary(dictionaryVector, new DictionaryEncoding(1L, false, null));

// 4. create a dictionary encoder
DictionaryEncoder encoder = new DictionaryEncoder.encode(dictionary, allocator);

// 5. encode the data
IntVector encoded = (IntVector) encoder.encode(unencoded);

// 6. re-create an un-encoded version from the encoded vector
VarCharVector decoded = (VarCharVector) encoder.decode(encoded);

One thing we haven’t discussed is how to create the dictionary vector from the original un-encoded values. That is left to the library user since a custom method will likely be more efficient than a general utility. Since the dictionary vector is just a normal vector, you can populate its values with the standard APIs.

Finally, you can package a number of dictionaries together, which is useful if you’re working with a VectorSchemaRoot with several dictionary-encoded vectors. This is done using an object called a DictionaryProvider. as shown in the example below. Note that we don’t put the dictionary vectors in the same VectorSchemaRoot as the data vectors, as they will generally have fewer values.

DictionaryProvider.MapDictionaryProvider provider =
    new DictionaryProvider.MapDictionaryProvider();

provider.put(dictionary);

The DictionaryProvider is simply a map of identifiers to Dictionary objects, where each identifier is a long value. In the above code you will see it as the first argument to the DictionaryEncoding constructor.

This is where the DictionaryEncoding’s ‘id’ attribute comes in. This value is used to connect dictionaries to instances of VectorSchemaRoot, using a DictionaryProvider. Here’s how that works:

  • The VectorSchemaRoot has a Schema object containing a list of Field objects.

  • The field has an attribute called ‘dictionary’, but it holds a DictionaryEncoding rather than a Dictionary

  • As mentioned, the DictionaryProvider holds dictionaries indexed by a long value. This value is the id from your DictionaryEncoding.

  • To retrieve the dictionary for a vector in a VectorSchemaRoot, you get the field associated with the vector, get its dictionary attribute, and use that object’s id to look up the correct dictionary in the provider.

// create the encoded vector, the Dictionary and DictionaryProvider as discussed above

// Create a VectorSchemaRoot with one encoded vector
VectorSchemaRoot vsr = new VectorSchemaRoot(List.of(encoded));

// now we want to decode our vector, so we retrieve its dictionary from the provider
Field f = vsr.getField(encoded.getName());
DictionaryEncoding encoding = f.getDictionary();
Dictionary dictionary = provider.get(encoding.getId());

As you can see, a DictionaryProvider is handy for managing the dictionaries associated with a VectorSchemaRoot. More importantly, it helps package the dictionaries for a VectorSchemaRoot when it’s written. The classes ArrowFileWriter and ArrowStreamWriter both accept an optional DictionaryProvider argument for that purpose. You can find example code for writing dictionaries in the documentation for (Reading/Writing IPC formats). ArrowReader and its subclasses also implement the DictionaryProvider interface, so you can retrieve the actual dictionaries when reading a file.

Slicing#

Similar with C++ implementation, it is possible to make zero-copy slices of vectors to obtain a vector referring to some logical sub-sequence of the data through TransferPair

IntVector vector = new IntVector("intVector", allocator);
for (int i = 0; i < 10; i++) {
  vector.setSafe(i, i);
}
vector.setValueCount(10);

TransferPair tp = vector.getTransferPair(allocator);
tp.splitAndTransfer(0, 5);
IntVector sliced = (IntVector) tp.getTo();
// In this case, the vector values are [0, 1, 2, 3, 4, 5, 6, 7, 8, 9] and the sliceVector values are [0, 1, 2, 3, 4].