| title |
|---|
Architecture |
This group of pages describes the internal mechanics of Datagrok. It is meant for more [advanced] users who would like to better understand how we [do X] to achieve excellent performance and provide features such as [1],[2],[3],[4],[5], and [6].
- Performance
- Architecture
- Grok script
Datagrok is designed with the following main objectives:
- Performance. This cannot be overstated. A modern system should be capable of working with tens of millions of datapoints, and every bit of performance counts. Memory consumption, latency, and network traffic fall in the same category.
- Security. We should be in complete control of the data.
- Usability. A system should be a joy to work with.
- Modularity. We should safely turn different features on or off without affecting the rest of the platform.
- Extensibility. We should be able to add new functionality easily.
- Interoperability. Other systems should be able to talk to us via open protocols, as well as seamlessly integrate with other platforms, environments, and programming languages.
- Maintainability. In the long run, it is indispensable to have the ability to navigate in the code easily, make changes, and reason about it. This is only possible when working with a clean, minimalistic, coherent codebase.
These goals heavily influenced the design. The platform appears quite traditional on the surface, as it consists of two parts - client and server. However, the way these parts work and connect is quite unusual.
A traditional approach to building such a system would be to commit to widely used technology such as JavaScript, select a bunch of widely used libraries and frameworks, including a UI framework, a visualization library, a serialization library, and a dozen of other ones, and plumb it all together. While the result would be somewhat predictable, it will also inevitably be mediocre due to all the compromises this path implies. You cannot build a fighter jet with a steam engine, random Lego parts, and a lot of duct tape.
Our firm belief, backed by 20 years of experience in developing analytical solutions and data analysis platforms, is that we need a clean break for such an ambitious project. Before building high-level features, we need to dive as deep as possible, build as an efficient foundation as possible, own every single detail of the implementation, and only then build our way up.
First of all, we built a fit-for-purpose data engine, which is essentially an in-memory columnar database. It is the heart of the system and is highly optimized for the routines typical for exploratory data analysis, i.e., fast sequential access for column data, efficient data storage, data transfer, data compression, filtering, aggregations, joins, caches, descriptive statistics, etc. It has an expressive and clean API.
On top of the data engine, we built several high-performance viewers that make heavy use of the data engine. The ability to use that engine unlocks unique possibilities for viewers, such as easy data aggregations or passing data to web workers for multithreaded rendering. All viewers access the same data (no copies are made), share specific statistics and cached calculations, have the same look and feel and usage patterns, and cooperate in particular tasks.
Also, on top of the data engine, we built a server. It is used for many purposes: data retrieval from different data sources, security, user management, metadata storage, multiple object repositories, user collaboration, running arbitrary scripts written in different languages, building and applying predictive models, and a lot more. The server utilizes extensible plugin architecture and currently has over 20 plugins. The server enables the client to do what is needed most efficiently, minimizing memory consumption and network traffic. It helps that both client and server are written in Dart and share a lot of common libraries.
Next, we built a data analytics application client that integrates our data engine, viewers, and server-based capabilities. The application can run entirely autonomously in the browser without making requests to the server. In that mode, it lets users import data of non-trivial sizes (tens of millions of rows) from local files in many popular data formats, clean and transform it using some tools and routines, interactively visualize it, and export the data. Using the server-based capabilities, it gets to another level by centrally managing data connections, enabling data governance and provenance, enabling scripting, providing data modeling capabilities, and dozens of other features outlined in our knowledge base.
In the results, Datagrok is an application that enables users to ingest, transform, analyze, visualize, model, and share data by combining powerful client and server capabilities. The application is highly flexible and extensible and allows creating new data pipelines, models dynamically and building viewers, plugins, and applications on top of it.
To extend already massive functionality, we created a compute engine for performing on-server computations. Compute engine lets the user easily invoke scripts written in languages such as R or Python and expose them to the Datagrok ecosystem as functions. Also, it utilizes a few different backends for predictive modeling while providing users the same look and feel. To create and share documents that contain live code, equations, visualizations, and narrative text, we added Jupyter Notebook features to compute server. Datagrok allows efficiently creating, editing, importing, linking, and applying Notebooks into tables. To scale computations, spin out multiple instances of some component, and our load balancer will take care of dispatching computations.
- Enterprise evaluation FAQ
- Infrastructure with detailed description of every component
- Deployment instruction to install and try Datagrok