Data API#

After creating a Table and some column families, you are ready to store and retrieve data.

Cells vs. Columns vs. Column Families#

  • As explained in the table overview, tables can have many column families.
  • As described below, a table can also have many rows which are specified by row keys.
  • Within a row, data is stored in a cell. A cell simply has a value (as bytes) and a timestamp. The number of cells in each row can be different, depending on what was stored in each row.
  • Each cell lies in a column (not a column family). A column is really just a more specific modifier within a column family. A column can be present in every column family, in only one or anywhere in between.
  • Within a column family there can be many columns. For example within the column family foo we could have columns bar and baz. These would typically be represented as foo:bar and foo:baz.

Modifying Data#

Since data is stored in cells, which are stored in rows, we use the metaphor of a row in classes that are used to modify (write, update, delete) data in a Table.

Direct vs. Conditional vs. Append#

There are three ways to modify data in a table, described by the MutateRow, CheckAndMutateRow and ReadModifyWriteRow API methods.

  • The direct way is via MutateRow which involves simply adding, overwriting or deleting cells. The DirectRow class handles direct mutations.
  • The conditional way is via CheckAndMutateRow. This method first checks if some filter is matched in a a given row, then applies one of two sets of mutations, depending on if a match occurred or not. (These mutation sets are called the “true mutations” and “false mutations”.) The ConditionalRow class handles conditional mutations.
  • The append way is via ReadModifyWriteRow. This simply appends (as bytes) or increments (as an integer) data in a presumed existing cell in a row. The AppendRow class handles append mutations.

Row Factory#

A single factory can be used to create any of the three row types. To create a DirectRow:

row = table.row(row_key)

Unlike the previous string values we’ve used before, the row key must be bytes.

To create a ConditionalRow, first create a RowFilter and then

cond_row = table.row(row_key, filter_=filter_)

To create an AppendRow

append_row = table.row(row_key, append=True)

Building Up Mutations#

In all three cases, a set of mutations (or two sets) are built up on a row before they are sent of in a batch via


Direct Mutations#

Direct mutations can be added via one of four methods

  • set_cell() allows a single value to be written to a column

    row.set_cell(column_family_id, column, value,

    If the timestamp is omitted, the current time on the Google Cloud Bigtable server will be used when the cell is stored.

    The value can either by bytes or an integer (which will be converted to bytes as a signed 64-bit integer).

  • delete_cell() deletes all cells (i.e. for all timestamps) in a given column

    row.delete_cell(column_family_id, column)

    Remember, this only happens in the row we are using.

    If we only want to delete cells from a limited range of time, a TimestampRange can be used

    row.delete_cell(column_family_id, column,
  • delete_cells() does the same thing as delete_cell() but accepts a list of columns in a column family rather than a single one.

    row.delete_cells(column_family_id, [column1, column2],

    In addition, if we want to delete cells from every column in a column family, the special ALL_COLUMNS value can be used

    row.delete_cells(column_family_id, row.ALL_COLUMNS,
  • delete() will delete the entire row


Conditional Mutations#

Making conditional modifications is essentially identical to direct modifications: it uses the exact same methods to accumulate mutations.

However, each mutation added must specify a state: will the mutation be applied if the filter matches or if it fails to match.

For example:

cond_row.set_cell(column_family_id, column, value,
                  timestamp=timestamp, state=True)

will add to the set of true mutations.

Append Mutations#

Append mutations can be added via one of two methods

  • append_cell_value() appends a bytes value to an existing cell:

    append_row.append_cell_value(column_family_id, column, bytes_value)
  • increment_cell_value() increments an integer value in an existing cell:

    append_row.increment_cell_value(column_family_id, column, int_value)

    Since only bytes are stored in a cell, the cell value is decoded as a signed 64-bit integer before being incremented. (This happens on the Google Cloud Bigtable server, not in the library.)

Notice that no timestamp was specified. This is because append mutations operate on the latest value of the specified column.

If there are no cells in the specified column, then the empty string (bytes case) or zero (integer case) are the assumed values.

Starting Fresh#

If accumulated mutations need to be dropped, use


Reading Data#

Read Single Row from a Table#

To make a ReadRows API request for a single row key, use Table.read_row():

>>> row_data = table.read_row(row_key)
>>> row_data.cells
    u'fam1': {
        b'col1': [
            < at 0x7f80d150ef10>,
            < at 0x7f80d150ef10>,
        b'col2': [
            < at 0x7f80d150ef10>,
    u'fam2': {
        b'col3': [
            < at 0x7f80d150ef10>,
            < at 0x7f80d150ef10>,
            < at 0x7f80d150ef10>,
>>> cell = row_data.cells[u'fam1'][b'col1'][0]
>>> cell
< at 0x7f80d150ef10>
>>> cell.value
>>> cell.timestamp
datetime.datetime(2016, 2, 27, 3, 41, 18, 122823, tzinfo=<UTC>)

Rather than returning a DirectRow or similar class, this method returns a PartialRowData instance. This class is used for reading and parsing data rather than for modifying data (as DirectRow is).

A filter can also be applied to the results:

row_data = table.read_row(row_key, filter_=filter_val)

The allowable filter_ values are the same as those used for a ConditionalRow. For more information, see the Table.read_row() documentation.

Stream Many Rows from a Table#

To make a ReadRows API request for a stream of rows, use Table.read_rows():

row_data = table.read_rows()

Using gRPC over HTTP/2, a continual stream of responses will be delivered. In particular

See the PartialRowsData documentation for more information.

As with Table.read_row(), an optional filter_ can be applied. In addition a start_key and / or end_key can be supplied for the stream, a limit can be set and a boolean allow_row_interleaving can be specified to allow faster streamed results at the potential cost of non-sequential reads.

See the Table.read_rows() documentation for more information on the optional arguments.

Sample Keys in a Table#

Make a SampleRowKeys API request with Table.sample_row_keys():

keys_iterator = table.sample_row_keys()

The returned row keys will delimit contiguous sections of the table of approximately equal size, which can be used to break up the data for distributed tasks like mapreduces.

As with Table.read_rows(), the returned keys_iterator is connected to a cancellable HTTP/2 stream.

The next key in the result can be accessed via

next_key =

or all keys can be iterated over via

for curr_key in keys_iterator:

Just as with reading, the stream can be canceled: