The Data Catalog

Note: This documentation is based on Kedro 0.16.4, if you spot anything that is incorrect then please create an issue or pull request.

This section introduces catalog.yml, the project-shareable Data Catalog. The file is located in conf/base and is a registry of all data sources available for use by a project; it manages loading and saving of data.

All supported data connectors are available in kedro.extras.datasets.

Using the Data Catalog within Kedro configuration

Kedro uses configuration to make your code reproducible when it has to reference datasets in different locations and/or in different environments.

You can copy this file and reference additional locations for the same datasets. For instance, you can use the catalog.yml file in conf/base/ to register the locations of datasets that would run in production while copying and updating a second version of catalog.yml that can be placed in conf/local/ to register the locations of sample datasets on the local computer that you are using for prototyping your data pipeline.

There is built-in functionality for conf/local/ to overwrite conf/base/ detailed here. This means that a dataset called cars could exist in the catalog.yml files in conf/base/ and conf/local/. In code, in src, you would only call a dataset named cars and Kedro would detect which definition of cars dataset to use to run your pipeline - cars definition from conf/local/catalog.yml would take precedence in this case.

The Data Catalog also works with the credentials.yml in conf/local/, allowing you to specify usernames and passwords that are required to load certain datasets.

The are two ways of defining a Data Catalog through the use of YAML configuration, or programmatically using an API. Both methods allow you to specify:

  • Dataset name
  • Dataset type
  • Location of the dataset using fsspec, detailed in the next section
  • Credentials needed in order to access the dataset
  • Load and saving arguments
  • Whether or not you want a dataset or ML model to be versioned when you run your data pipeline

Specifying the location of the dataset

Kedro relies on fsspec for reading and saving data from a variety of data stores including local file systems, network file systems, cloud object stores, and Hadoop. When specifying a storage location in filepath:, a URL should be provided using the general form protocol://path/to/data. If no protocol is provided, the local file system is assumed (same as file://).

The following prepends are available:

  • Local or Network File System: file:// - the local file system is default in the absence of any protocol, it also permits relative paths.
  • Hadoop File System (HDFS): hdfs://user@server:port/path/to/data - Hadoop Distributed File System, for resilient, replicated files within a cluster.
  • Amazon S3: s3://my-bucket-name/path/to/data - Amazon S3 remote binary store, often used with Amazon EC2, using the library s3fs.
  • Google Cloud Storage: gcs:// - Google Cloud Storage, typically used with Google Compute resource using gcsfs (in development).
  • HTTP(s): http:// or https:// for reading data directly from HTTP web servers.

fsspec also provides other file systems that may be of interest to Kedro users, such as SSH, FTP and WebHDFS. See the documentation for more information.

Using the Data Catalog with the YAML API

The YAML API allows you to configure your datasets in a YAML configuration file, conf/base/catalog.yml or conf/local/catalog.yml.

Here is an example data config catalog.yml:

# Example 1: Loads / saves a CSV file from / to a local file system

  type: pandas.CSVDataSet
  filepath: data/01_raw/bikes.csv

# Example 2: Loads and saves a CSV on a local file system, using specified load and save arguments

  type: pandas.CSVDataSet
  filepath: data/01_raw/company/cars.csv
    sep: ','
    index: False
    date_format: '%Y-%m-%d %H:%M'
    decimal: .

# Example 3: Loads and saves a compressed CSV on a local file system

  type: pandas.CSVDataSet
  filepath: data/01_raw/company/boats.csv.gz
    sep: ','
    compression: 'gzip'
      mode: 'rb'

# Example 4: Loads a CSV file from a specific S3 bucket, using credentials and load arguments

  type: pandas.CSVDataSet
  filepath: s3://your_bucket/data/02_intermediate/company/motorbikes.csv
  credentials: dev_s3
    sep: ','
    skiprows: 5
    skipfooter: 1
    na_values: ['#NA', NA]

# Example 5: Loads / saves a pickle file from / to a local file system

  type: pickle.PickleDataSet
  filepath: data/06_models/airplanes.pkl
  backend: pickle

# Example 6: Loads an excel file from Google Cloud Storage

  type: pandas.ExcelDataSet
  filepath: gcs://your_bucket/data/02_intermediate/company/motorbikes.xlsx
    project: my-project
  credentials: my_gcp_credentials
    sheet_name: Sheet1

# Example 7: Save an image created with Matplotlib on Google Cloud Storage

  type: matplotlib.MatplotlibWriter
  filepath: gcs://your_bucket/data/08_results/plots/output_1.jpeg
    project: my-project
  credentials: my_gcp_credentials

# Example 8: Loads / saves an HDF file on local file system storage, using specified load and save arguments

  type: pandas.HDFDataSet
  filepath: data/02_intermediate/skateboards.hdf
  key: name
    columns: [brand, length]
    mode: w  # Overwrite even when the file already exists
    dropna: True

# Example 9: Loads / saves a parquet file on local file system storage, using specified load and save arguments

  type: pandas.ParquetDataSet
  filepath: data/02_intermediate/trucks.parquet
    columns: [name, gear, disp, wt]
    categories: list
    index: name
    compression: GZIP
    file_scheme: hive
    has_nulls: False
    partition_on: [name]

# Example 10: Load / saves a Spark table on S3, using specified load and save arguments

  type: spark.SparkDataSet
  filepath: s3a://your_bucket/data/01_raw/weather*
  credentials: dev_s3
  file_format: csv
    header: True
    inferSchema: True
    sep: '|'
    header: True

# Example 11: Loads / saves a SQL table using credentials, a database connection, using specified load and save arguments

  type: pandas.SQLTableDataSet
  credentials: scooters_credentials
  table_name: scooters
    index_col: [name]
    columns: [name, gear]
    if_exists: replace

# Example 12: Load a SQL table with credentials, a database connection, and applies a SQL query to the table

  type: pandas.SQLQueryDataSet
  credentials: scooters_credentials
  sql: select * from cars where gear=4
    index_col: [name]

# Example 13: Load data from an API endpoint, example US corn yield data from USDA

  type: api.APIDataSet
    key: SOME_TOKEN
    format: JSON
    commodity_desc: CORN
    statisticcat_des: YIELD
    agg_level_desc: STATE
    year: 2000
Note: When using pandas.SQLTableDataSet or pandas.SQLQueryDataSet you must provide a database connection string. In the example above we pass it using scooters_credentials key from the credentials (see the details in Feeding in credentials section below). scooters_credentials must have a top-level key con containing SQLAlchemy compatible connection string. Alternative to credentials would be to explicitly put con into load_args and save_args (pandas.SQLTableDataSet only).

Adding parameters

You can configure parameters for your project and reference them in your nodes. The way to do this is via add_feed_dict() method (Relevant API documentation: DataCatalog). You should be able to use this method to add any other entry / metadata you wish on the DataCatalog.

Feeding in credentials

Before instantiating the DataCatalog Kedro will first attempt to read the credentials from project configuration (see this section for more details). Resulting dictionary will then be passed into DataCatalog.from_config() as credentials argument.

Let’s assume that the project contains the file conf/local/credentials.yml with the following contents:

  key: token
  secret: key

  con: sqlite:///kedro.db

  id_token: key

In the example above catalog.yml contains references to credentials keys dev_s3 and scooters_credentials. It means that when instantiating motorbikes dataset, for example, the DataCatalog will attempt to read top-level key dev_s3 from the received credentials dictionary, and then will pass its values into the dataset __init__ as credentials argument. This is essentially equivalent to calling this:

    load_args=dict(sep=",", skiprows=5, skipfooter=1, na_values=["#NA", "NA"],),
    credentials=dict(key="token", secret="key"),

Loading multiple datasets that have similar configuration

You may encounter situations where your datasets use the same file format, load and save arguments, and are stored in the same folder. YAML has a built-in syntax for factorising parts of a YAML file, which means that you can decide what is generalisable across your datasets so that you do not have to spend time copying and pasting dataset configurations in catalog.yml.

You can see this in the following example:

_csv: &csv
  type: spark.SparkDataSet
  file_format: csv
    sep: ','
    na_values: ['#NA', NA]
    header: True
    inferSchema: False

  <<: *csv
  filepath: s3a://data/01_raw/cars.csv

  <<: *csv
  filepath: s3a://data/01_raw/trucks.csv

  <<: *csv
  filepath: s3a://data/01_raw/bikes.csv
    header: False

The syntax &csv names the following block csv and the syntax <<: *csv inserts the contents of the block named csv. Locally declared keys entirely override inserted ones as seen in bikes.

Note: It’s important that the name of the template entry starts with a _ so Kedro knows not to try and instantiate it as a dataset.

You can also nest reuseable YAML syntax:

_csv: &csv
  type: spark.SparkDataSet
  file_format: csv
  load_args: &csv_load_args
    header: True
    inferSchema: False

  <<: *csv
  filepath: s3a://data/01_raw/airplanes.csv
    <<: *csv_load_args
    sep: ;

In this example the default csv configuration is inserted into airplanes and then the load_args block is overridden. Normally that would replace the whole dictionary. In order to extend load_args the defaults for that block are then re-inserted.

Transcoding datasets

You may come across a situation where you would like to read the same file using two different dataset implementations. Use transcoding when you want to load and save the same file, via its specified filepath, using different DataSet implementations.

A typical example of transcoding

For instance, parquet files can not only be loaded via the ParquetDataSet using pandas, but also directly by SparkDataSet. This conversion is typical when coordinating a Spark to pandas workflow.

To enable transcoding, you will need to define two DataCatalog entries for the same dataset in a common format (Parquet, JSON, CSV, etc.) in your conf/base/catalog.yml:

  type: spark.SparkDataSet
  filepath: data/02_intermediate/data.parquet
  file_format: parquet

  type: pandas.ParquetDataSet
  filepath: data/02_intermediate/data.parquet

These entries will be used in the pipeline like this:

        node(func=my_func1, inputs="spark_input", outputs="my_dataframe@spark"),
        node(func=my_func2, inputs="my_dataframe@pandas", outputs="pipeline_output"),

How does transcoding work?

In this example, Kedro understands that my_dataframe is the same dataset in its spark.SparkDataSet and pandas.ParquetDataSet formats and helps resolve the node execution order.

In the pipeline, Kedro uses the spark.SparkDataSet implementation for saving and pandas.ParquetDataSet for loading, so the first node should output a pyspark.sql.DataFrame, while the second node would receive a pandas.Dataframe.

Transforming datasets

Transformers intercept the load and save operations on Kedro DataSets. Use cases that transformers enable include:

  • Performing data validation,
  • Tracking operation performance,
  • And, converting a data format (although we would recommend Transcoding for this).

Applying built-in transformers

The use case of tracking operation performance by applying built-in transformers for monitoring load and save operation latency will be covered.

Transformers are applied at the DataCatalog level. To apply the built-in ProfileTimeTransformer, you need to:

  1. Navigate to src/<package_name>/
  2. Apply ProfileTimeTransformer in the hook implementation TransformerHooks.after_catalog_created.
  3. Register the hook in your ProjectContext as follow:
from pathlib import Path
from typing import Dict

from kedro.extras.transformers import ProfileTimeTransformer # new import
from kedro.framework.context import KedroContext, load_package_context
from kedro.framework.hooks import hook_impl # new import
from import DataCatalog # new import

class TransformerHooks:
    def after_catalog_created(self, catalog: DataCatalog) -> None:

class ProjectContext(KedroContext):

    hooks = (TransformerHooks(),)

Once complete, rerun the pipeline from the terminal and you should see the following logging output:

$ kedro run

2019-11-13 15:09:01,784 - - INFO - Loading data from `companies` (CSVDataSet)...
2019-11-13 15:09:01,827 - ProfileTimeTransformer - INFO - Loading companies took 0.043 seconds
2019-11-13 15:09:01,828 - kedro.pipeline.node - INFO - Running node: preprocessing_companies: preprocess_companies([companies]) -> [preprocessed_companies]
2019-11-13 15:09:01,880 - kedro_tutorial.nodes.data_engineering - INFO - Running 'preprocess_companies' took 0.05 seconds
2019-11-13 15:09:01,880 - kedro_tutorial.nodes.data_engineering - INFO - Running 'preprocess_companies' took 0.05 seconds
2019-11-13 15:09:01,880 - - INFO - Saving data to `preprocessed_companies` (CSVDataSet)...
2019-11-13 15:09:02,112 - ProfileTimeTransformer - INFO - Saving preprocessed_companies took 0.232 seconds
2019-11-13 15:09:02,113 - kedro.runner.sequential_runner - INFO - Completed 1 out of 6 tasks

You can notice 2 new INFO level log messages from ProfileTimeTransformer, which report the corresponding dataset load and save operation latency.

Pro Tip: You can narrow down the application of the transformer by specifying an optional list of the datasets in add_transformer. For example, the command catalog.add_transformer(profile_time, ["dataset1", "dataset2"]) will apply profile_time transformer only to the datasets named dataset1 and dataset2. This may be useful when you need to apply a transformer only to a subset of datasets, rather than all of them.

Developing your own transformer

The use case of tracking operation performance by developing our own transformer for tracking memory consumption will be covered. A built-in memory profiler is supported, however, in this example you will learn how to create your own one.

You can profile memory using memory-profiler. The custom transformer should:

  1. Inherit the base class
  2. Implement the load and save method (as show in the example below)

Create src/<package_name>/ and then paste the following code into it:

import logging
from typing import Callable, Any

from import AbstractTransformer
from memory_profiler import memory_usage

def _normalise_mem_usage(mem_usage):
    # memory_profiler < 0.56.0 returns list instead of float
    return mem_usage[0] if isinstance(mem_usage, (list, tuple)) else mem_usage

class ProfileMemoryTransformer(AbstractTransformer):
    """ A transformer that logs the maximum memory consumption during load and save calls """

    def _logger(self):
        return logging.getLogger(self.__class__.__name__)

    def load(self, data_set_name: str, load: Callable[[], Any]) -> Any:
        mem_usage, data = memory_usage(
            (load, [], {}),
        # memory_profiler < 0.56.0 returns list instead of float
        mem_usage = _normalise_mem_usage(mem_usage)
            "Loading %s consumed %2.2fMiB memory at peak time", data_set_name, mem_usage
        return data

    def save(self, data_set_name: str, save: Callable[[Any], None], data: Any) -> None:
        mem_usage = memory_usage(
            (save, [data], {}),
        mem_usage = _normalise_mem_usage(mem_usage)
            "Saving %s consumed %2.2fMiB memory at peak time", data_set_name, mem_usage

Finally, you need to update TransformerHooks to apply your custom transformer:

from .memory_profile import ProfileMemoryTransformer # new import

class TransformerHooks:
    def after_catalog_created(self, catalog: DataCatalog) -> None:

        # as memory tracking is quite time-consuming, for demonstration purposes
        # let's apply profile_memory only to the master_table
        catalog.add_transformer(ProfileMemoryTransformer(), "master_table")

class ProjectContext(KedroContext):

    hooks = (TransformerHooks(),)

And re-run the pipeline:

$ kedro run

2019-11-13 15:55:01,674 - - INFO - Saving data to `master_table` (CSVDataSet)...
2019-11-13 15:55:12,322 - ProfileMemoryTransformer - INFO - Saving master_table consumed 606.98MiB memory at peak time
2019-11-13 15:55:12,322 - ProfileTimeTransformer - INFO - Saving master_table took 10.648 seconds
2019-11-13 15:55:12,357 - kedro.runner.sequential_runner - INFO - Completed 3 out of 6 tasks
2019-11-13 15:55:12,358 - - INFO - Loading data from `master_table` (CSVDataSet)...
2019-11-13 15:55:13,933 - ProfileMemoryTransformer - INFO - Loading master_table consumed 533.05MiB memory at peak time
2019-11-13 15:55:13,933 - ProfileTimeTransformer - INFO - Loading master_table took 1.576 seconds

Versioning datasets and ML models

Making a simple addition to your Data Catalog allows you to perform versioning of datasets and machine learning models.

Consider the following versioned dataset defined in the catalog.yml:

  type: pandas.CSVDataSet
  filepath: data/01_raw/company/cars.csv
  versioned: True

The DataCatalog will create a versioned CSVDataSet called cars.csv. The actual csv file location will look like data/01_raw/company/cars.csv/<version>/cars.csv, where <version> corresponds to a global save version string formatted as

You can run the pipeline with a particular versioned data set with --load-version flag as follows:

kedro run --load-version=""

where --load-version is dataset name and version timestamp separated by :.

This section shows just the very basics of versioning. You can learn more about how this feature can be used in Kedro IO.

Using the Data Catalog with the Code API

The code API allows you to configure data sources in code. This can also be used to operate the IO module within notebooks.

Configuring a Data Catalog

In a file like, you can construct a DataCatalog object programmatically. In the following, we are using a number of pre-built data loaders documented in the API reference documentation.

from import DataCatalog
from kedro.extras.datasets.pandas import (

io = DataCatalog(
        "bikes": CSVDataSet(filepath="../data/01_raw/bikes.csv"),
        "cars": CSVDataSet(
            filepath="../data/01_raw/cars.csv", load_args=dict(sep=",")
        "cars_table": SQLTableDataSet(
            table_name="cars", credentials=dict(con="sqlite:///kedro.db")
        "scooters_query": SQLQueryDataSet(
            sql="select * from cars where gear=4",
        "ranked": ParquetDataSet(filepath="ranked.parquet"),
Note: When using SQLTableDataSet or SQLQueryDataSet you must provide a con key containing SQLAlchemy compatible database connection string. In the example above we pass it as part of credentials argument. Alternative to credentials would be to put con into load_args and save_args (SQLTableDataSet only).

Loading datasets

Each dataset can be accessed by its name.

cars = io.load("cars")  # data is now loaded as a DataFrame in 'cars'
gear = cars["gear"].values

Behind the scenes

The following steps happened behind the scenes when load was called:

  • The value cars was located in the Data Catalog
  • The corresponding AbstractDataSet object was retrieved
  • The load method of this dataset was called
  • This load method delegated the loading to the underlying pandas read_csv function

Viewing the available data sources

If you forget what data was assigned, you can always review the DataCatalog.


Saving data

Saving data can be completed with a similar API.

Note: This use is not recommended unless you are prototyping in notebooks.

Saving data to memory

from import MemoryDataSet

memory = MemoryDataSet(data=None)
io.add("cars_cache", memory)"cars_cache", "Memory can store anything.")

Saving data to a SQL database for querying

At this point we may want to put the data in a SQLite database to run queries on it. Let’s use that to rank scooters by their mpg.

import os

# This cleans up the database in case it exists at this point
except FileNotFoundError:
    pass"cars_table", cars)
ranked = io.load("scooters_query")[["brand", "mpg"]]

Saving data in Parquet

Finally we can save the processed data in Parquet format."ranked", ranked)
Note: Saving None to a dataset is not allowed!