Use Cases - I want to…

Here follows a list of specific things that people may want to get done, and details of how Intake can help. The details of how to achieve each of these activities can be found in the rest of the detailed documentation.

Avoid copy&paste of blocks of code for accessing data

This is a very common pattern, if you want to load some specific data, to find someone, perhaps a colleague, who has accessed it before, and copy that code. Such a practice is extremely error prone, and cause a proliferation of copies of code, which may evolve over time, with various versions simultaneously in use.

Intake separates the concerns of data-source specification from code. The specs are stored separately, and all users can reference the one and only authoritative definition, whether in a shared file, a service visible to everyone or by using the Intake server. This spec can be updated so that everyone gets the current version instead of relying on outdated code.

Version control data sources

Version control (e.g., using git) is an essential practice in modern software engineering and data science. It ensures that the change history is recorded, with times, descriptions and authors along with the changes themselves.

When data is specified using a well-structured syntax such as YAML, it can be checked into a version controlled repository in the usual fashion. Thus, you can bring rigorous practices to your data as well as your code.

If using conda packages to distribute data specifications, these come with a natural internal version numbering system, such that users need only do conda update ... to get the latest version.

“Install” data

Often, finding and grabbing data is a major hurdle to productivity. People may be required to download artifacts from various places or search through storage systems to find the specific thing that they are after. One-line commands which can retrieve data-source specifications or the files themselves can be a massive time-saver. Furthermore, each data-set will typically need it’s own code to be able to access it, and probably additional software dependencies.

Intake allows you to build conda packages, which can include catalog files referencing online resources, or to include data files directly in tha package. Whether uploaded to anaconda.org or hosted on a private enterprise channel, getting the data becomes a single conda install ... command, whereafter it will appear as an entry in intake.cat. The conda package brings versioning and dependency declaration for free, and you can include any code that may be required for that specific data-set directly in the package too.

Update data specifications in-place

Individual data-sets often may be static, but commonly, the “best” data to get a job done changes with time as new facts emerge. Conversely, the very same data might be better stored in a different format which is, for instance, better-suited to parallel access in the cloud. In such situations, you really don’t want to force all the data scientists who rely on it to have their code temporarily broken and be forced to change this code.

By working with a catalog file/service in a fixed shared location, it is possible to update the data source specs in-place. When users now run their code, they will get the latest version. Because all Intake drivers have the same API, the code using the data will be identical and not need to be changed, even when the format has been updated to something more optimised.

Access data stored on cloud resources

Services such as AWS S3, GCS and Azure Datalake (or private enterprise variants of these) are increasingly popular locations to amass large amounts of data. Not only are they relatively cheap per GB, but they provide long-term resilience, metadata services, complex access control patterns and can have very large data throughput when accessed in parallel by machines on the same architecture.

Intake comes with integration to cloud-based storage out-of-the box for most of the file-based data formats, to be able to access the data directly in-place and in parallel. For the few remaining cases where direct access is not feasible, the caching system in Intake allows for download of files on first use, so that all further access is much faster.

Work with “Big Data”

The era of Big Data is here! The term means different things to different people, but certainly implies that an individual data-set is too large to fit into the memory of a typical workstation computer (>>10GB). Nevertheless, most data-loading examples available use functions in packages such as pandas and expect to be able to produce in-memory representations of the whole data. This is clearly a problem, and a more general answer should be available aside from “get more memory in your machine”.

Intake integrates with Dask and Spark, which both offer out-of-core computation (loading the data in chunks which fit in memory and aggregating result) or can spread their work over a cluster of machines, effectively making use of the shared memory resources of the whole cluster. Dask integration is built into the majority of the the drivers and exposed with the .to_dask() method, and Spark integration is available for a small number of drivers with a similar .to_spark() method, as well as directly with the intake-spark package.

Intake also integrates with many data services which themselves can perform big-data computations, only extracting the smaller aggregated data-sets that do fit into memory for further analysis. Services such as SQL systems, solr, elastic-search, splunk, accumulo and hbase all can distribute the work required to fulfill a query across many nodes of a cluster.

Find the right data-set

Browsing for the data-set which will solve a particular problem can be hard, even when the data have been curated and stored in a single, well-structured system. You do not want to rely on word-of-mouth to specify which data is right for which job.

Intake catalogs allow for self-description of data-sets, with simple text and arbitrary metadata, with a consistent access pattern. Not only can you list the data available to you, but you can find out what exactly that data represents, and the form the data would take if loaded (table versus list of items, for example). This extra metadata is also searchable: you can descend through a hierarchy of catalogs with a single search, and find all the entries containing some particular keywords.

You can use the Intake GUI to graphically browse through your available data-sets or point to catalogs available to you, look through the entries listed there and get information about each, or even show a sample of the data or quick-look plots. The GUI is also able to execute searches and browse file-systems to find data artifacts of interest. This same functionality is also available via a command-line interface or programmatically.

Work remotely

Interacting with cloud storage resources is very convenient, but you will not want to download large amounts of data to your laptop or workstation for analysis. Intake finds itself at home in the remote-execution world of jupyter and Anaconda Enterprise and other in-browser technologies. For instance, you can run the Intake GUI either as a stand-alone application for browsing data-sets or in a notebook for full analytics, and have all the runtime live on a remote machine, or perhaps a cluster which is co-located with the data storage. Together with cloud-optimised data formats such as parquet, this is an ideal set-up for processing data at web scale.

Transform data to efficient formats for sharing

A massive amount of data exists in human-readable formats such as JSON, XML and CSV, which are not very efficient in terms of space usage and need to be parsed on load to turn into arrays or tables. Much faster processing times can be had with modern compact, optimised formats, such as parquet.

Intake has a “persist” mechanism to transform any input data-source into the format most appropriate for that type of data, e.g., parquet for tabular data. The persisted data will be used in preference at analysis time, and the schedule for updating from the original source is configurable. The location of these persisted data-sets can be shared with others, so they can also gain the benefits, or the “export” variant can be used to produce an independent version in the same format, together with a spec to be reference it by - then you would share this spec with others.

Access data without leaking credentials

Security is important. Users’ identity and authority to view specific data should be established before handing over any sensitive bytes. It is, unfortunately, all too common for data scientists to include their username, passwords or other credentials directly in code, so that it can run automatically, thus presenting a potential security gap.

Intake does not manage credentials or user identities directly, but does provide hooks for fetching details from the environment or other service, and using the values in templating at the time of reading the data. Thus, the details are not included in the code, but every access still requires for them to be present.

In other cases, you may want to require the user to provide their credentials every time, rather that automatically establish them, and “user parameters” can be specified in Intake to cover this case.

Establish a data gateway

The Intake server protocol allows you fine-grained control over the set of data sources that are listed, and exactly what to return to a user when they want to read some of that data. This is an ideal opportunity to include authorisation checks, audit logging, and any more complicated access patterns, as required.

By streaming the data through a single channel on the server, rather than allowing users direct access to the data storage backend, you can log and verify all access to your data.

Clear distinction between data curator and analyst roles

It is desirable to separate out two tasks: the definition of data sources specifications, and accessing and using data. This is so that those who understand the origins of the data and the implications of various formats and other storage options (such as chunk-size) should make those decisions and encode what they have done into specs. It leaves the data users, e.g., data scientists, free to find and use the data-sets appropriate for their work and simply get on with their job - without having to learn about various storage formats and access APIs.

This separation is at the very base of what Intake was designed to do.

Users to be able to access data without learning every backend API

Data formats and services are a wide mess of many libraries and APIs. A large amount of time can be wasted in the life of a data scientist or engineer in finding out the details of the ones required by their work. Intake wraps these various libraries, REST APIs and similar, to provide a consistent experience for the data user. source.read() will simply get all of the data into memory in the container type for that source - no further parameters or knowledge required.

Even for the curator of data catalogs or data driver authors, the framework established by Intake provides a lot of convenience and simplification which allows each person to deal with only the specifics of their job.

Data sources to be self-describing

Having a bunch of files in some directory is a very common pattern for data storage in the wild. There may or may not be a README file co-located giving some information in a human-readable form, but generally not structured - such files are usually different in every case.

When a data source is encoded into a catalog, the spec offers a natural place to describe what that data is, along with the possibility to provide an arbitrary amount of structured metadata and to describe any parameters that are to be exposed for user choice. Furthermore, Intake data sources each have a particular container type, so that users know whether to expect a dataframe, array, etc., and simple introspection methods like describe and discover which return basic information about the data without having to load all of it into memory first.

A data source hierarchy for natural structuring

Usually, the set of data sources held by an organisation have relationships to one-another, and would be poorly served to be provided as a simple flat list of everything available. Intake allows catalogs to refer to other catalogs. This means, that you can group data sources by various facets (type, department, time…) and establish hierarchical data-source trees within which to find the particular data most likely to be of interest. Since the catalogs live outside and separate from the data files themselves, as many hierarchy structures as thought useful could be created.

For even more complicated data source meta-structures, it is possible to store all the details and even metadata in some external service (e.g., traditional SQL tables) with which Intake can interact to perform queries and return particular subsets of the available data sources.

Expose several data collections under a single system

There are already several catalog-like data services in existence in the world, and some organisation may have several of these in-house for various different purposes. For example, a SQL-server may hold details of customer lists and transactions, but historical time-series and reference data are held in archival data formats like parquet on a file-storage system; while real-time system monitoring is done by a totally unrelated system such as Splunk or elastic-search.

Of course, Intake can read from various file formats and data services. However, it can also interpret the internal conception of data catalogs that some data services may have. For example, all of the tables known to the SQL server, or all of the pre-defined queries in Splunk can be automatically included as catalogs in Intake, and take their place amongst the regular YAML-specified data sources, with exactly the same usage for all of them.

These data sources and their hierarchical structure can then be exposed via the graphical data browser, for searching, selecting and visualising data-sets.

Modern visualisations for all data-sets

Intake is integrated with the pyviz suite, particularly hvplot, to bring simple yet powerful data visualisations to any Intake data source by using just one single method for everything. These plots are interactive, and can include server-side dynamic aggregation of very large data-sets to display more data points than the browser can handle.

You can specify specific plot types right in the data source definition, as have these customised visualisations available to the user as simple one-liners know to well-describe the content of the data, or even view the same visuals right in the graphical data source browser application. Thus, Intake is already an all-in-one data investigation and dashboarding app.

Update data specifications in real time

Intake data catalogs are not limited to reading static specification from files. They can also execute queries on remote data services and return lists of data sources dynamically at runtime. New data sources may appear, for example, as directories of data files are pushed to a storage service, or new tables are created within a SQL server.

Distribute data in a custom format

Sometimes, the well-know data formats are just not right for a give data-set, and a custom-built format is required. In such cases, the code to read the data may not exist in any library. Intake allows for code to be distributed along with data source specs/catalogs or even files in a single conda package. That encapsulates everything needed to describe and use that particular data, and can then be distributed as a single entity, and installed with a one-liner.

Furthermore, should the few builtin container types (sequence, array, dataframe) not be sufficient, you can supply your own, and then build drivers that use it. This was done, for example, for xarray-type data, where multiple related N-D arrays share a coordinate system and metadata. By creating this container, a whole world of scientific and engineering data was opened up to Intake. Creating new containers is not hard, though, and we foresee more coming, such as machine-learning models and streaming/real-time data.