Skip to content

Latest commit

 

History

History
733 lines (572 loc) · 24.8 KB

writing-dense.rst

File metadata and controls

733 lines (572 loc) · 24.8 KB
Raw

Writing Dense Arrays

In this tutorial you will learn how to write to dense arrays. It is highly recommended that you read the tutorials on dense arrays and tiling first, as well as writing sparse arrays.

Full programs
Program Links
writing_dense_padding writingdensepaddingcpp writingdensepaddingpy
writing_dense_multiple writingdensemultiplecpp writingdensemultiplepy
writing_dense_sparse writingdensesparsecpp
writing_dense_global writingdenseglobalcpp
writing_dense_global_expansion writingdenseglobalexpansioncpp

Basic concepts and definitions

TileDB allows you to partially populate a dense array, i.e., it permits empty cells. In this tutorial we explain that TileDB materializes only cells that belong to partially written tiles. Moreover, you can write sparse cells to a dense array. In this case empty cells are not materialized at all, similar to the case of sparse arrays.

When reading a subarray from a dense array, TileDB returns values for all cells in the subarray, even for the empty ones. For each empty cell, TileDB returns a special "fill value" which depends on the attribute data type. This is different from the case of sparse arrays where a subarray read query returns values only for the non-empty cells.

Writing to a dense array

Let us revisit the quickstart_dense example of tutorial dense-arrays. Here is how we wrote to the array:

cpp

Context ctx;
std::vector<int> data = {
    1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};
Array array(ctx, array_name, TILEDB_WRITE);
Query query(ctx, array);
query.set_layout(TILEDB_ROW_MAJOR).set_buffer("a", data);
query.submit();
array.close();

The above code omits setting one more parameter to the query object, namely the subarray inside which the cell values will be written. TileDB understands that you will be writing to the entire domain, i.e., [1,4], [1,4]. The result would be equivalent if we added the following lines before submitting the query:

std::vector<int> subarray = {1, 4, 1, 4};
query.set_subarray(subarray);

python

ctx = tiledb.Ctx()
with tiledb.DenseArray(ctx, array_name, mode='w') as A:
    data = np.array(([1, 2, 3, 4],
                     [5, 6, 7, 8],
                     [9, 10, 11, 12],
                     [13, 14, 15, 16]))
    A[:] = data

This writes values 1, 2, ..., 16 to the 4x4 dense TileDB array. You can write to any subarray in a dense array, i.e., you do not have to populate the entire array at once (see also the section on multiple writes below). For instance, the code below would just populate only two rows of the above array, i.e., subarray [1,2], [1,4]:

python

ctx = tiledb.Ctx()
# Open the array and write to it.
with tiledb.DenseArray(ctx, array_name, mode='w') as A:
    data = np.array(([1, 2, 3, 4],
                     [5, 6, 7, 8]))
    A[1:3, 1,5] = data

cpp

Observe that the above code sets the query layout to row-major. This means that the values 1, 2, 3, ..., 16 are laid out inside buffer data (provided to the query) in row-major order with respect to the subarray query. You can also set the layout to column-major or global order instead as well (we explain this in more detail below). TileDB knows how to efficiently re-organize the cells internally (if needed) and map them to the global order upon writing the values to disk.

To better demonstrate the effect of the query layout in writes, let us create the same array as above, but now with 2x2 space tiling, and experiment with row-major, column-major and global query layout upon writing. Substitute the create_array and write_array functions of quickstart_dense with the ones provided below.

void create_array() {
  Context ctx;
  if (Object::object(ctx, array_name).type() == Object::Type::Array)
    return;

  // 4x4 domain with 2x2 space tiling
  Domain domain(ctx);
  domain.add_dimension(Dimension::create<int>(ctx, "rows", {{1, 4}}, 2))
        .add_dimension(Dimension::create<int>(ctx, "cols", {{1, 4}}, 2));
  ArraySchema schema(ctx, TILEDB_DENSE);
  schema.set_domain(domain).set_order({{TILEDB_ROW_MAJOR, TILEDB_ROW_MAJOR}});
  schema.add_attribute(Attribute::create<int>(ctx, "a"));
  Array::create(array_name, schema);
}

void write_array() {
  std::vector<int> data = {
      1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};
  std::vector<int> subarray = {1, 4, 1, 4};

  Context ctx;
  Array array(ctx, array_name, TILEDB_WRITE);
  Query query(ctx, array);
  query.set_layout(TILEDB_ROW_MAJOR) // Try also with TILEDB_COL_MAJOR/TILEDB_GLOBAL_ORDER
       .set_buffer("a", data)
       .set_subarray(subarray);
  query.submit();
  array.close();
}

python

As explained above, we write to TileDB using numpy arrays. By default, a numpy array lays out the cell values internally in row-major order (or C-order). You can convert the layout of a numpy array x to column-major (or Fortran-order) prior to writing it to a TileDB array simply as follows:

np.asfortranarray(x)

TileDB will check the layout of the numpy values and will write in the corresponding layout.

Warning

Currently global writes are not supported in the Python API.

The figure below depicts the array contents when varying the query layout. Observe that the layout is always specified with respect to the query subarray.

Empty space / Padding

Since TileDB allows you to write to any subarray of your array, what happens if your array has empty areas? We demonstrate with example writing_dense_padding. We first create a 4x4 array with 2x2 space tiling. We then write only subarray [2,3], [12], leaving the rest of the array unpopulated:

python

ctx = tiledb.Ctx()
# Open the array and write to it.
with tiledb.DenseArray(ctx, array_name, mode='w') as A:
    # Write to [2,3], [1,2]
    data = np.array(([1, 2], [3, 4]))
    A[2:4, 1:3] = data

The array looks like in the figure below.

The example then reads the entire array (i.e., [1,4], [1,4]) in row-major order and prints the cell values on the screen. Here is the output after running the program:

cpp

$ g++ -std=c++11 writing_dense_padding.cc -o writing_dense_padding_cpp -ltiledb
$ ./writing_dense_padding_cpp
-2147483648
-2147483648
-2147483648
-2147483648
1
2
-2147483648
-2147483648
3
4
-2147483648
-2147483648
-2147483648
-2147483648
-2147483648
-2147483648

python

$ python writing_dense_padding.py
[[-2147483648 -2147483648 -2147483648 -2147483648]
 [          1           2 -2147483648 -2147483648]
 [          3           4 -2147483648 -2147483648]
 [-2147483648 -2147483648 -2147483648 -2147483648]]

Observe that for every empty cell in [1,4], [1,4], TileDB returned value -2147483648. This is the default fill value, which is is equal to the minumum value stored in an integer variable. The table below shows the default fill values for all supported attribute data types.

Data type Fill value
TILEDB_CHAR Minimum char value
TILEDB_INT8 Minimum int8 value
TILEDB_UINT8 Maximum uint8 value
TILEDB_INT16 Minimum int16 value
TILEDB_UINT16 Maximum uint16 value
TILEDB_INT32 Minimum int32 value
TILEDB_UINT32 Maximum uint32 value
TILEDB_INT64 Minimum int64 value
TILEDB_UINT64 Maximum uint64 value
TILEDB_FLOAT32 NaN
TILEDB_FLOAT64 NaN
TILEDB_ASCII 0
TILEDB_UTF8 0
TILEDB_UTF16 0
TILEDB_USC2 0
TILEDB_USC4 0
TILEDB_ANY 0

Note

In a future release, you will be able to define your own fill values for each of your arrays.

But you may wonder, how does TileDB handles empty spaces at the physical level?

Note

TileDB does not materialize empty cells for dense arrays, except for the ones belonging to partially written tiles.

Let us explain the above with our running example. The array has 2x2 space tiling, which means that we have two partially written tiles (the upper left and lower left), and two completely empty tiles (upper right and lower right). TileDB distinguishes between partially written tiles and completely empty tiles. We mentioned in an earlier tutorial that TileDB always writes integral tiles on the file, i.e., it cannot write just 2 out of 4 cells in our example. Hence, TileDB will explicitly store the fill value for each empty cell in a partially written tile. In contrast, it does not materialize any cells for the completely empty tiles (i.e., it entirely ignores empty tiles). This is depicted in the figure below.

Let us inspect the contents of the array folder, after running the above example:

$ ls -l writing_dense_padding/
total 8
drwx------  4 stavros  staff  136 Jun 25 10:28 __76679576e9454d1eb08e8a2f5bf8fd29_1529936885146
-rwx------  1 stavros  staff  109 Jun 25 10:28 __array_schema.tdb
-rwx------  1 stavros  staff    0 Jun 25 10:28 __lock.tdb
$ ls -l writing_dense_padding/__76679576e9454d1eb08e8a2f5bf8fd29_1529936885146/
total 16
-rwx------  1 stavros  staff  106 Jun 25 10:28 __fragment_metadata.tdb
-rwx------  1 stavros  staff   32 Jun 25 10:28 a.tdb

As expected, there is a single fragment subdirectory. Observe the size of attribute file a.tdb, which is 32 bytes. Given that the attribute type is int32 (and since there is no compression), this implies that 8 cell values were written to the file, i.e., two tiles. This confirms that padding applies only to partially written tiles, whereas completely empty tiles are not materialized at all.

Multiple writes / Updates

You can write to a dense array multiple times, similar to the case we described for sparse arrays in an earlier tutorial. Each write in row-major or column-major layout creates a new subfolder/fragment in the array directory (we explain unordered and global layout in the subsections below). Consider the following two writes to the dense array we have been using in the examples above (see code example writing_dense_multiple):

python

ctx = tiledb.Ctx()
# Open the array and write to it.
with tiledb.DenseArray(ctx, array_name, mode='w') as A:
    # First write
    data = np.array(([1, 2], [3, 4]))
    A[1:3, 1:3] = data

    # Second write
    data = np.array(([5, 6, 7, 8],
                     [9, 10, 11, 12]))
    A[2:4, 1:5] = data

The first writes to subarray [1,2], [1,2], whereas the second to [2,3], [1,4]. The figure below depicts the two writes, as well as the collective logical view of the array after the second write.

After running the program, we get the following output. Observe that the full read (in row-major) order results in retrieving the cell values as depicted in the collective logical view of the array (again, retrieving the default fill values for the empty cells).

cpp

$ g++ -std=c++11 writing_dense_multiple.cc -o writing_dense_multiple_cpp -ltiledb
$ ./writing_dense_multiple_cpp
1
2
-2147483648
-2147483648
5
6
7
8
9
10
11
12
-2147483648
-2147483648
-2147483648
-2147483648

python

$ python writing_dense_multiple.py
[[          1           2 -2147483648 -2147483648]
 [          5           6           7           8]
 [          9          10          11          12]
 [-2147483648 -2147483648 -2147483648 -2147483648]]

Listing the array directory, you can see that there are two subfolders/fragments created:

$ ls -l writing_dense_multiple/
total 8
drwx------  4 stavros  staff  136 Jun 25 14:51 __71c5c364bc4b4f49888668c912c4a01c_1529952665416
-rwx------  1 stavros  staff  109 Jun 25 14:51 __array_schema.tdb
drwx------  4 stavros  staff  136 Jun 25 14:51 __d13ece6b48ca470b8c42810cdf9d9206_1529952665410
-rwx------  1 stavros  staff    0 Jun 25 14:51 __lock.tdb

Writing sparse cells

Warning

Currently sparse writes to dense arrays are not supported in the Python API.

One exciting feature about dense arrays is that you can write sparse cells to them, i.e., you can write multiple cells that do not necessary fall in the same hyper-rectangular subarray in a single write query. The sparse writes in dense arrays are identical to those of sparse arrays, i.e., you need to add an extra buffer that holds the explicit coordinates of the cells you are writing into. Also here is where the unordered layout is relevant again. Everything discussed about writes in writing-sparse (e.g., even writing in global order) holds here as well.

Let us demonstrate with code example writing_dense_sparse. Here is how we write some sparse cells to the dense array of the previous examples:

The array resulting from the above write looks as follows:

Compiling and running the program gives the output shown below. Observe that, contrary to sparse arrays, when slicing dense arrays, TileDB returns fill values for empty areas as explain earlier in this tutorial. Recall that, in the case of sparse arrays, you get back only the values of the non-empty cells; no cell is ever materialized for sparse arrays (neither in writes nor reads). Also notice that you can explicitly request the cell coordinates even in dense arrays.

$ g++ -std=c++11 writing_dense_sparse.cc -o writing_dense_sparse_cpp -ltiledb
$ ./writing_dense_sparse_cpp
Cell (1, 1) has data -2147483648
Cell (1, 2) has data 1
Cell (1, 3) has data -2147483648
Cell (1, 4) has data 4
Cell (2, 1) has data 2
Cell (2, 2) has data -2147483648
Cell (2, 3) has data -2147483648
Cell (2, 4) has data -2147483648
Cell (3, 1) has data -2147483648
Cell (3, 2) has data -2147483648
Cell (3, 3) has data -2147483648
Cell (3, 4) has data -2147483648
Cell (4, 1) has data -2147483648
Cell (4, 2) has data -2147483648
Cell (4, 3) has data 3
Cell (4, 4) has data -2147483648

Let us inspect the contents of the dense array after the write:

$ ls -l writing_dense_sparse/
total 8
drwx------  5 stavros  staff  170 Jun 25 17:59 __88e0ebca9aa44442918ad93fa82209f2_1529963970336
-rwx------  1 stavros  staff  109 Jun 25 17:59 __array_schema.tdb
-rwx------  1 stavros  staff    0 Jun 25 17:59 __lock.tdb
$ ls -l writing_dense_sparse/__88e0ebca9aa44442918ad93fa82209f2_1529963970336/
total 24
-rwx------  1 stavros  staff   32 Jun 25 17:59 __coords.tdb
-rwx------  1 stavros  staff  110 Jun 25 17:59 __fragment_metadata.tdb
-rwx------  1 stavros  staff   16 Jun 25 17:59 a.tdb

Observe that the coordinates were written explicitly in file __coords.tdb inside the fragment, similarly to the sparse case. Note that here the sparse format is adopted, i.e., no empty cell is materialized. This is in contrast to the padding technique explained earlier in this tutorial for dense fragments. In other words, when writing sparse cells to dense arrays, TileDB creates a sparse fragment, which is treated as a sparse array snapshot! We provide more information on fragments in a later tutorial.

Writing in global layout

Warning

Currently writing in global layout is not supported in the Python API.

TileDB allows you to write in global order similar to the case of sparse arrays. This generally leads to better performance, but comes at the expense of extra usage complexity. The limitation with this layout is that you need to always write integral tiles, i.e., the subarray you set to the query (and write into) must not partially intersect tiles, but instead encompass them entirely.

The following code writes to subarray [1,4], [1,2] (see example writing_dense_global). Observe that, similar to the case of sparse arrays, we can submit the same query multiple times, effectively appending to the same fragment (list the contents of the resulting array to verify this). The difference here is that we update the contents of the buffer we already set, without needing to reset the buffer to the query in the second write. Either way works here (resetting a new buffer, or updating the contents of the set buffer without resetting). Important: do not forget to finalize the query when you are done writing/appending in global order and before closing the array.

The resulting array is depicted below. Note that the subarray we wrote into contains exactly two tiles. Any attempt to write to partial tiles in global order would have failed (and the behavior can be unexpected).

Writing in global order mode must be done with extra care in case some tile extent does not divide the respective dimension domain. As we have explained in an earlier tutorial, this results in internal domain expansion. Moreover, TileDB does not allow you to write outside your defined domain. Therefore, if your domain contains partial tiles, you will not be able to write to them in global order.

We illustrate with an example. Consider you have a 4x3 array with 2x2 space tiling, as shown in the figure below. The domain contains two entrire tiles (upper left and lower left) and two partial tiles (upper right and lower right). In this case, you can write in global order in subarray [1,4], [1,2], but choose another layout (e.g., row-major) for [1,4], [3,3]. This is done in code example writing_dense_global_expanison in two writes, which are also shown in the figure below.

Writing and performance

The writing performance can be affected by various factors, such as the tiling, compression and query layout. See the performance/introduction tutorial for more information about the TileDB performance.