Data types

This section provides an overview of which data types used by the engine and how we can convert between data types using crandas.

The supported data types include:

  • int8, int16, int24, int32, int40, int48, int56, int64, int72, int80, int88, int96

  • fp8, fp16, fp24, fp32, fp40, fp48, fp56, fp64, fp72, fp80, fp88, fp96

  • uint8, uint16, uint24, uint32, uint40, uint48, uint56, uint64, uint72, uint80, uint88, uint96

  • vec_int (vector of integers)

  • varchar (text)

  • date

  • bytes (binary data)

  • bool (as integer)

Added in version 1.10: Fixed-point numbers from fp40 and bigger

Added in version 1.8: Integers from int40 and bigger and dates.

Added in version 1.7: Fixed-point numbers

For specific details about numeric values, see the next section.

For each data type, also a “nullable” variant is supported that can hold missing values. Nullable data data types are denoted with a question mark, e.g., int8?, bytes??. Nullable data types have some caveats; read on or go to their own section for details.

Warning

When boolean values are uploaded into the engine, they are transformed to integers and therefore take the values 0 and 1 instead of False and True respectively.

CDataFrame allow for missing values, but not by default, so they must be specified.

Data Type Conversion

Crandas provides a method for converting data types using the CSeries.astype() method. In the following example, we will show you how to convert a column of strings to a column of integers.

Note

This is the only type conversion currently supported.

import crandas as cd

# Create a crandas DataFrame with a string column
uploaded = cd.DataFrame({"vals": ["1","2","3","4"]})

# Convert the string column to a int column
uploaded = uploaded.assign(vals2=uploaded["vals"].astype(int))

The above example converts the string column vals to a new integer column called vals2.

For a more in depth look at specifying integer types, go to the next section.

Note

It is also possible to specify the desired type while uploading the data, using ctype={"val": "varchar[9]"}.

ctypes

Because the engine uses highly specialized algorithms to compute on secret data, it uses a specialized typing system that is more fine-grained than what pandas uses. Crandas implements this type system, which we call ctypes (similarly to pandas’ dtypes). In certain situations, it is important to specify the specific type of our data

import pandas as pd
from crandas import ctypes

# Specify data types for the DataFrame
table = cd.DataFrame(
    {"ints": [1, 2, 3], "strings": ["a", "bb", "ccc"]},
    ctype={"ints": "int8", "strings": "varchar[5]"},
)

or alternatively:

(...)
# Specify data types for the DataFrame
table = cd.DataFrame(
    {"ints": cd.Series([1, 2, 3], ctype="int8"),
    "strings": cd.Series(["a", "bb", "ccc"], ctype="varchar[5]")}
)

In the above example, we define the ints column with a NonNullableInteger data type (crandas.ctypes.NonNullableInteger()), and the strings column is defined with a varchar[5] data type (string with at most 5 characters).

Hint

If there are missing/null values in the column that can be specified by adding ? after the ctype (e.g. int8?, varchar?[5])

Crandas also supports other data types, such as byte arrays:

from uuid import uuid4

# Create a DataFrame with UUIDs stored as bytes
df = cd.DataFrame({"uuids": [uuid4().bytes for _ in range(5)]}, ctype="bytes")

You are also able to specify types through pandas’ typing, known as dtypes. Note that not all dtypes have an equivalent ctypes.

# Create a DataFrame with multiple data types
df = cd.DataFrame(
    {
        "strings": pd.Series(["test", "hoi", "ok"], dtype="string"),
        "int": pd.Series([1, 2, 3], dtype="int"),
        "int64": pd.Series([23, 11, 91238], dtype="int64"),
        "int32": pd.Series([12831, 1231, -1231], dtype="int32"),
    }
)

It is possible to retrieve the ctype of a crandas DataFrame or a column by using its .ctype attribute (see crandas.crandas.CDataFrame.ctype(), crandas.crandas.CSeriesColRef.ctype(), crandas.crandas.Col.ctype()), for example:

> cdf=cd.DataFrame({"a": [1], "b": "11"})
> print(cdf.ctype)
{'a': 'int[min=0,max=255]', 'b': 'varchar[2,ascii]'}
> print(cdf["a"].ctype)
int[min=0,max=255]

Using ctypes and schemas

As shown above, the ctype argument can be used to specify data types of individual columns for upload functions such as crandas.crandas.DataFrame(). Note that this ctype includes metadata provided by the user at upload as well as metadata (such as bounds; see Type detection from data) derived by the engine.

Instead of this, it is also possible to specify the data types of all columns at the same time, by using the schema argument. The schema specifies the order, names, and types of all columns to be uploaded, but it does not contain engine-derived metadata.

Note

The ctype and schema arguments are mutually exclusive. Although they take values of the same type, i.e. both are specified using a dictionary of a column name mapping to a crandas.ctype.Ctype, their use is different. A schema is a “full blueprint” containing an ordered dictionary of column names and types. The ctype argument is used when uploading a table that does not need to conform to a particular layout, but the user would like to specify some additional type information. Both allow mapping to a crandas.ctype.Ctype that does not have fully specified bounds: for example, in both the ctype and schema arguments a particular column can be specified as either int or int16. Obtaining the schema for an existing CDataFrame does however currently not specify any bounds, e.g. it will always return int even if the column was specified as int16; see the known limitation described below.

For example, the following specifies that the given CSV will be uploaded with the given column names in the specified order and with the specified types:

>>> cd.read_csv("titanic_scaled.csv", auto_bounds=True, schema={'Survived': 'int',
'Pclass': 'fp', 'Sex': 'int', 'Age': 'fp'})

When using script recording, schemas can be used to ensure that there is an exact correspondence between the (dummy) data when recording a script, and the (production) data used when executing the script. To this end, in both environments, the same schema needs to be specified as argument to upload functions (cdf=cd.DataFrame(..., schema=...); cdf=cd.read_csv(..., schema=...); cdf=cd.upload_pandas_dataframe(..., schema=...); etc), and/or to cd.get_table(..., schema=...).

To learn the schema of data available in pandas format without actually already uploading the data, the functions crandas.crandas.pandas_dataframe_schema() and crandas.crandas.read_csv_schema() can be used, for example:

>>> cd.read_csv_schema("titanic_scaled.csv", auto_bounds=True, ctype={"Survived": "int"})
{'Survived': 'int', 'Pclass': 'fp', 'Sex': 'int', 'Age': 'fp'}

It is also possible to retrieve the schema of an existing CDataFrame by using its .schema attribute (see crandas.crandas.CDataFrame.schema(), crandas.crandas.CSeriesColRef.schema(), crandas.crandas.Col.schema()). In this case, currently, a known limitation is that, when retrieving a CDataFrame from the engine, the fine-grained information from the original schema (for example, bit lengths) is lost. Consider for example:

>>> cdf = cd.DataFrame({"a": [1], "b": "11"}, ctype={"a": "int16"})
>>> print(cdf.schema)
{'a': 'int16', 'b': 'varchar'}
>>> cdf = cd.get_table(cdf.handle)
>>> print(cdf.schema) # the fact that `a` is a 16-bit number is lost
{'a': 'int', 'b': 'varchar'}

However, whether a column is nullable is part of the schema. For example:

>>> cdf = cd.DataFrame({"a": [1, pd.NA], "b": "11"}, ctype={"a": "int16?"})
>>> print(cdf.schema)
{'a': 'int16?', 'b': 'varchar'}
>>> cdf = cd.get_table(cdf.handle)
>>> print(cdf.schema)
{'a': 'int?', 'b': 'varchar'}

Given a schema, it is also possible to create a zero-row crandas dataframe matching the schema by using cdf = cd.DataFrame(schema=...) (without supplying any data).

It is also possible to manually create dummy data that adheres to a given schema, by creating columns that have ctypes corresponding to the respective schema ctypes. For example, given a schema {'a': 'int', 'b': 'varchar'}, the following creates a table of dummy data that matches the schema:

dummy_data = cd.DataFrame({
    "a": cd.Series([], ctype="int"),
    "b": cd.Series([], ctype="varchar"),
})

The dummy data can easily be instantiated by replacing the [] with actual values.

Finally, to convert an already existing crandas dataframe to a given schema (if this conversion is possible), cdf.astype(schema=...) can be used (see crandas.crandas.CDataFrame.astype()).

Type detection from data

For some types of column, crandas can derive the exact column type from the data being uploaded. This may lead to ColumnBoundDerivedWarning warnings, e.g.:

ColumnBoundDerivedWarning: Column type for column a (uint8) was automatically
derived from the input data, see User Guide->Data types->Automatic type detection
for more information.

This warning is given if no exact ctype (e.g., uint8, varchar[ascii]) is given for the column.

For example, for an integer column, if no size is given, the smallest integer type is selected that is valid for the column, e.g.:

cdf = cd.DataFrame({"vals": [1,2,3]})

In this example, the column vals will be derived to be of type uint8 (unsigned 8-bit integer) because all values lie in the range from 0 to 255 (inclusive). See Automatic type detection for integers.

Similarly, for varchar (text) columns, it will be derived from the data whether the column is of ASCII or unicode type. See String encoding and efficiency.

Note that, because the column type is derived from the data, this potentially leaks information. For example, if one party uploads a list of salaries that are all smaller than 65536 and another party uploads a list of salaries that contains higher salaries, then the first column will have column type uint16 and the second column will have column uint24. The first party can prevent this by explicitly assigning the ctype uint24 to the uploaded column.

When type information is detected from data, the user gets a warning about this, e.g., ColumnBoundDerivedWarning: Column "vals" was automatically derived to be of type uint8. These warnings can be suppressed with the standard Python mechanism, e.g.,

import warnings

# Suppress warnings about automatically derived column types
warnings.simplefilter("ignore", category=cd.ctypes.ColumnBoundDerivedWarning)

cdf = cd.DataFrame({"vals": [1,2,3]})

It is also possible to provide auto_bounds=True as argument to data uploading functions (see Query Arguments, or to set the configuration setting (see crandas.config) auto_bounds to True:

import crandas as cd
import crandas.config

# Suppress ColumnBoundDerivedWarning globally...
cd.config.settings.auto_bounds = True

# or for a single upload
cdf = cd.DataFrame({"vals": [1,2,3]}, auto_bounds=True)

Working with missing values

Crandas can work with null values, although this requires extra care. Columns do not allow null values by default but this can be achieved in multiple ways. Whenever a column with missing values is added, the engine will determine that such column can have null values. Additionally, it is possible to specify that a column will allow null values when uploading it, even if the column currently does not contain any such values.

Warning

When using a column with missing values in combination with script signing, it is advisable to explicitly specify that the column allows null values, by defining the ctype as discussed below. This way, there will not be a mismatch between the approved analysis and the performed analysis, even if the dummy or actual data does not contain nullable values. For example, the following code designates the column ints as allowing missing values, even if none of the uploaded values are missing.

from crandas import ctypes


table = cd.DataFrame(
    {"ints": [1, 2, 3], "strings": ["a", "bb", None]},
    ctype={"ints": "int32?"},
)

Hint

When specifying ctypes for columns with missing values you can use int32? or any other supported data type with ? (this indicates missing values).

Both columns created in this example allow for null values. The first one because it was explictly specified and the latter because it contains a null value. Crandas considers the same values to be null as pandas; in particular, this includes None, pandas.NA, and numpy.nan.

To turn a nullable column into a non-nullable one, the fillna function can be used. For example, the following code example replaces all missing values of a string column by the string empty:

import crandas as cd

cdf = cd.DataFrame({"a": ["test", None]})
cdf = cdf.assign(a=lambda x: x.a.fillna("empty"))

Numeric types have additional particularities that are important to know, both in the typing system and because we can do arithmetic operations over them. The next section deals with these types.