The versatility that enables "one way to rule them all"... requires a devtool​

A unified approach to ETL processes centers around standardization without compromising flexibility. To achieve this, we need to be enabled to build and run custom code, bu also have helpers to enable us to standardise and simplify our work.

In the data space, we have a few custom code options, some of which portable. But what is needed to achieve universality and portability is more than just a code standard.

So what do we expect from such a tool?

• It should be created for our developers
• it should be easily pluggable into existing tools and workflows
• it should perform across a variety of hardware and environments.

Data teams don't speak Object Oriented Programming (OOP)​

Connectors are nice, but when don't exist or break, what do we do? We need to be able to build and maintain those connectors simply, as we work with the rest of our scripts.

The data person has a very mixed spectrum of activities and responsibilities, and programming is often a minor one. Thus, across a data team, while some members can read or even speak OOP, the team will not be able to do so without sacrificing other capabilities.

This means that in order to be able to cater to a data team as a dev team, we need to aknowledge a different abstraction is needed.

Goodbye OOP, hello @decorators!​

Data teams often navigate complex systems and workflows that prioritize functional clarity over object-oriented programming (OOP) principles. They require tools that simplify process definition, enabling quick, readable, and maintainable data transformation and movement. Decorators serve this purpose well, providing a straightforward way to extend functionality without the overhead of class hierarchies and inheritance.

Decorators in Python allow data teams to annotate functions with metadata and operational characteristics, effectively wrapping additional behavior around core logic. This approach aligns with the procedural mindset commonly found in data workflows, where the emphasis is on the transformation steps and data flow rather than the objects that encapsulate them.

By leveraging decorators, data engineers can focus on defining what each part of the ETL process does—extract, transform, load—without delving into the complexities of OOP. This simplification makes the code more accessible to professionals who may not be OOP experts but are deeply involved in the practicalities of data handling and analysis.

The ability to run embedded is more than just scalability​

Most traditional ETL frameworks are architected with the assumption of relatively abundant computational resources. This makes sense given the resource-intensive nature of ETL tasks when dealing with massive datasets.

However, this assumption often overlooks the potential for running these processes on smaller, more constrained infrastructures, such as directly embedded within an orchestrator or on edge devices.

The perspective that ETL processes necessarily require large-scale infrastructure is ripe for challenge. In fact, there is a compelling argument to be made for the efficiency and simplicity of executing ETL tasks, particularly web requests for data integration, on smaller systems. This approach can offer significant cost savings and agility, especially when dealing with less intensive data loads or when seeking to maintain a smaller digital footprint.

Small infrastructure ETL runs can be particularly efficient in situations where real-time data processing is not required, or where data volumes are modest. By utilizing the orchestrator's inherent scheduling and management capabilities, one can execute ETL jobs in a leaner, more cost-effective manner. This can be an excellent fit for organizations that have variable data processing needs, where the infrastructure can scale down to match lower demands, thereby avoiding the costs associated with maintaining larger, underutilized systems.

Running on small workers is easier than spinning up infra​

Running ETL processes directly on an orchestrator can simplify architecture by reducing the number of moving parts and dependencies. It allows data teams to quickly integrate new data sources and destinations with minimal overhead. This methodology promotes a more agile and responsive data architecture, enabling businesses to adapt more swiftly to changing data requirements.

It's important to recognize that this lean approach won't be suitable for all scenarios, particularly where data volumes are large or where the complexity of transformations requires the robust computational capabilities of larger systems. Nevertheless, for a significant subset of ETL tasks, particularly those involving straightforward data integrations via web requests, running on smaller infrastructures presents an appealing alternative that is both cost-effective and simplifies the overall data processing landscape.

Dealing with spiky loads is easier on highly parallel infras like serverless functions​

Serverless functions are particularly adept at managing spiky data loads due to their highly parallel and elastic nature. These platforms automatically scale up to handle bursts of data requests and scale down immediately after processing, ensuring that resources are utilized only when necessary. This dynamic scaling not only improves resource efficiency but also reduces costs, as billing is based on actual usage rather than reserved capacity.

The stateless design of serverless functions allows them to process multiple, independent tasks concurrently. This capability is crucial for handling simultaneous data streams during peak times, facilitating rapid data processing that aligns with sudden increases in load. Each function operates in isolation, mitigating the risk of one process impacting another, which enhances overall system reliability and performance.

Moreover, serverless architectures eliminate the need for ongoing server management and capacity planning. Data engineers can focus solely on the development of ETL logic without concerning themselves with underlying infrastructure issues. This shift away from operational overhead to pure development accelerates deployment cycles and fosters innovation.

Some examples of embedded portability with dlt​

Dagster's embedded ETL now supports dlt - enabling devs to do what they love - build.​

The "Stop Reinventing Orchestration: Embedded ELT in the Orchestrator" blog post by Pedram from Dagster Labs, introduces the concept of Embedded ELT within an orchestration framework, highlighting the transition in data engineering from bulky, complex systems towards more streamlined, embedded solutions that simplify data ingestion and management. This evolution is seen in the move away from heavy tools like Airbyte or Meltano towards utilizing lightweight, performant libraries which integrate seamlessly into existing orchestration platforms, reducing deployment complexity and operational overhead. This approach leverages the inherent capabilities of orchestration systems to handle concerns typical to data ingestion, such as state management, error handling, and observability, thereby enhancing efficiency and developer experience.

dlt was built for just such a scenario and we are happy to be adopted into it. Besides adding connectors, dlt adds a simple way to build custom pipelines.

Read more about it on Dagster blog post on dlt.

Dagworks' dlt + duckdb + ibis + Hamilton demo​

The DAGWorks Substack post introduces a highly portable pipeline of all libraries, and leverages a blend of open-source Python libraries: dlt, Ibis, and Hamilton. This integration exemplifies the trend towards modular, decentralized data systems, where each component specializes in a segment of the data handling process—dlt for extraction and loading, Ibis for transformation, and Hamilton for orchestrating complex data flows. These technologies are not just tools but represent a paradigm shift in data engineering, promoting agility, scalability, and cost-efficiency in deploying serverless microservices.

The post not only highlights the technical prowess of combining these libraries to solve practical problems like message retention and thread summarization on Slack but also delves into the meta aspects of such integrations. It reflects on the broader implications of adopting a lightweight stack that can operate within diverse infrastructures, from cloud environments to embedded systems, underscoring the shift towards interoperability and backend agnosticism in data engineering practices. This approach illustrates a shift in the data landscape, moving from monolithic systems to flexible, adaptive solutions that can meet specific organizational needs without heavy dependencies or extensive infrastructure.

Read more about it on Dagworks blog post on dlt.

Closing thoughts​

The concepts discussed here—portability, simplicity, and scalability—are central to modern data engineering practices. They reflect a shift towards tools that not only perform well but also integrate seamlessly across different environments, from high-powered servers to minimal infrastructures like edge devices. This shift emphasizes the importance of adaptability in tools used by data teams, catering to a broad spectrum of deployment scenarios without sacrificing performance.

In this landscape, dlt exemplifies the type of tool that embodies these principles. It's not just about being another platform; it's about providing a framework that supports the diverse needs of developers and engineers. dlt's design allows it to be embedded directly within various architectures, enabling teams to implement robust data processes with minimal overhead. This approach reduces complexity and fosters an environment where innovation is not hindered by the constraints of traditional data platforms.

We invite the community to engage with these concepts through dlt, contributing to its evolution and refinement. By participating in this collaborative effort, you can help ensure that the tool remains at the forefront of data engineering technology, providing effective solutions that address the real-world challenges of data management and integration.

Join the conversation and share your insights in our Slack community or contribute directly to the growing list of projects using us. Your expertise can drive the continuous improvement of dlt, shaping it into a tool that not only meets current demands but also anticipates future needs in the data engineering field.

Definitions of how I use the terms:

Data Governance: A system of oversight and guidance over the data, much like a government is a system of oversight and guidance for a country. The opposite of governance is anarchy, chaos, and entropy.

Data Democracy: A type of governance that ensures stakeholders are part of the governance.

Shift left: Assuming data flows from left to right, shift left represents a focus towards the origin.

Data Mesh: A decentralized data management strategy that treats data as a product, with domain-specific teams managing its quality, governance, and lifecycle.

Shift Left Data Democracy: From Access to Involvement​

In the traditional view, data democracy was largely about democratizing access—ensuring that everyone across the organization could easily retrieve and analyze data. This was a crucial step forward, breaking down silos and making information more available than ever before. However, as we've evolved, so too has our understanding of what true data democracy entails.

Shift left data democracy represents a more profound change. It's not just about making data accessible post-factum; it's about involving a broader spectrum of roles in the very processes of data ingestion, processing, and management. This approach extends the principles of democracy to the entire data lifecycle, beginning with data ingestion.

It's a shift from mere consumption to participation, where access is just the beginning.

Data mesh is the driver​

Just as the data mesh concept emerged to address the complexities of managing data in a distributed, domain-oriented environment, we now see a need for technology to evolve in parallel. The goal? To put data sources directly in the hands of the people who use them. This means giving teams the tools and autonomy to manage and process their data, ensuring governance and quality from the outset and throughout the data's lifecycle.

This shift left approach to data democracy aligns with the idea behind data mesh, recognizing that effective data management and governance are not centralized activities but distributed responsibilities. By involving more stakeholders from the very start of the data flows, we're not just democratizing access; we're democratizing the entire data flows.

Governance, from a power game, to a team sport; A brief history of how we got here​

Building a data warehouse is a beaten path - but how to go from technical solution to organisation-wide application?

Building a data warehouse for reporting on some business processes is a good start, but in order to leverage that data we need a culture to do so and the skills to do it correctly.

While a centralised solution enables a skilled team to deliver results, these results are often inflexible without hands on help - so how can the organisation be expected to become data driven? The process of tracking a goal, creating hypotheses, starting an experiment and then tracking outcomes is much more complex than that of tracking a metric in a dashboard.

Cue, the move to democratic data access.

From Monarchy to Democracy: Data access for the people!​

The move from a centralised system to a democratic system comes from the competitive use of data. In a centralised system where only management has access, data is used to keep track of goals. To enable people to use that data to do something about the goals, the user must have access and understanding of the data.

As with anything, the first step is obvious: Give people access - without it, there is no progress. However, once we do that, the reality rears its ugly head: Access is not enough!

Democratic access is great but as long as the data producers are not providing clean documented data , we don't have a democracy. Instead what we have is reality-adjusted communism - we all have plentiful access to the same black box or garbage that the big central team put in.

So, after democratizing data access, the next challenge was to answer the obvious question: So what does this data mean?

Turns out, the central team doesn't quite know either - it's rather the owner of the process we track, the data producer, that understands how the data they emit links to the real life process it tracks.

So how do we go from having data to understanding what it means?

From democratizing access to democratizing literacy though embedded analysts​

One easy way to help teams understand the data is to give them an analyst resource. And what better than someone who knows their domain?

Cue the embedded analysts. These folks are crucial in bridging the gap between data capabilities and domain-specific needs. By positioning data experts within specific business units, organizations can ensure that the insights generated are highly relevant and immediately applicable to the domain's unique challenges and opportunities.

This placement helps in several key ways:

• Domain expertise meets data munging: Embedded analysts develop a deep understanding of the specific challenges and workflows of the business unit they are part of, which enables them to tailor data models and analytics strategies effectively.
• Data literacy: These analysts act as champions of data within their teams, educating and training non-data savvy members on data-driven decision-making processes. This upskills the team and increases the overall data literacy within the unit.
• Faster response times: Being close to the operational realities of the business unit, embedded analysts can deliver faster, more targeted responses to data queries and needs, reducing the time from question to insight.

And just as we started, we solve another increment of the problem, which reveals the next.

Now that we can analyse the data, we need the data. But, it turns out the data we have is dirty, and we are missing some outright.

So let's solve the next problem: Data sources and quality.

The Advent of Data Mesh: Solving the data source problem​

Wow, well we went quite a way to get here, and a decade after talking about democratization, we are starting to recognize that governance is an activity, not a process. And democracy is more work than we originally thought.

The data mesh architecture marks a significant evolution in the data democratization journey. Data mesh advances the principles of embedded analysts by decentralizing data ownership entirely, promoting domain-specific control over data assets.

This architectural approach is based on the idea that data should not only be accessible but also actionable across various sections of an organization without bottlenecks.

And just like governments hire a lot of people, turns out, a governance system also needs people to work for it.

Data mesh tries to solve much of that by embracing domain-oriented decentralization. In this model, data is treated as a product with the domain teams as the product owners. These teams are responsible for ensuring their data's quality and relevance, significantly reducing the latency issues found in centralized systems by eliminating the lengthy processes of data cleansing and approval.

Further, data mesh empowers teams with the necessary tools and authority to manage their data effectively, fostering a culture where data is a valuable asset across all levels of the organization. This approach not only supports rapid decision-making and innovation within teams but also offers scalability and flexibility as organizational data needs evolve, allowing domains to independently expand their data operations without a comprehensive overhaul of the central data infrastructure.

Of course, at this point having a complete or partial data platform that offers some governance starts to become very important as we don't want individual business units to be burdened with responsibity but without proper tooling - or the outcome will be high entropy.

From retrofitting governance to applying it from the start: Shift left data democracy!​

Imagine a world where your company's data sources can just be picked and unpacked in the destination of your choice by analysts - not through an external saas tool, but via an internal service portal.

Shift-Left Data Democracy (SLDD) is a concept in data management that advocates for integrating data governance early in the data lifecycle. This approach shifts governance practices from being a retrospective or reactionary activity to an integral part of the initial design and development phases of data systems. By doing so, SLDD aims to embed governance, quality controls, and compliance measures at the point of data creation and throughout its subsequent handling.

By embedding governance early in the data lifecycle, SLDD eliminates the complex and costly process of retrofitting governance frameworks to mature datasets and systems. This proactive approach leads to streamlined operations, reducing both the complexity and the cost traditionally associated with late-stage governance implementation.

This early incorporation of governance enhances transparency throughout the entire process. Stakeholders gain a clear understanding of how data is managed and governed from the start, building trust and ensuring compliance.

What's revolutionary about SLDD is that a governed data source can easily be unfolded into a normalised or analytical model.

This "ad hoc data mart" can be used without central bottlenecks and easily customised to fit specific cases without having to reach modelling consensus with other teams. This built-in modularity avoids the creation of more bottlenecks downstream, enabling fast research and development where needed.

Further, a well-defined governance framework enables greater innovation within safe boundaries. Teams can explore and innovate knowing they are aligned with compliance and operational standards, which speeds up experimentation and development cycles. This environment encourages a more dynamic approach to data handling, where creativity is not stifled by fear of violating governance protocols. By treating governance as an integral part of the data management process rather than a hindrance, SLDD fosters a culture where data truly drives innovation.

Distinction between data mesh and shift-left data democracy​

While both concepts advocate for decentralized governance, they focus on different aspects of the data lifecycle. Data mesh architecture emphasizes the structural and operational decentralization of data management, granting autonomy to domain-specific teams. Shift-left data democracy, on the other hand, extends this decentralization to the very beginning of the data lifecycle, advocating for early involvement and broad stakeholder participation in governance processes.

The main difference is: Mesh is applied post-factum. For newly built systems, starting with governance as a technical universal standard is less complex. And while mesh grants autonomy, the entropy raises complexities and cost; on the other hand formalising and standardising responsibilities from the start of data production reduces entropy.

Practicing shift-left data democracy​

So how do we do it? Is this a future or can we already do it?

We asked ourselves the same and we are working towards fully supporting the standard.

Ensuring quality at the source​

Start with having quality control embedded in the source. Here's what I mean - start with a clear schema for your data, and ensure you have a strategy to adapt to change. One such strategy could be having data contracts, refusing and data that does not fit the defined schema. The other strategy, would be evolving the schema into a staging layer and notifying changes, so the engineering analyst can look into the data to understand what happened and correctly deal with the change.

At dlt we support schema evolution and data contracts. docs.

Metadata for full lineage​

Column and row level lineage are a basic necessity of development and traceability, so ensure each ingested package is annotated with source and time. Keep track of when columns are added to a source. Associate those schema changes with the corresponding load package to achieve column and row level lineage already from the ingestion layer, referring to a source defined as pipeline code, not table.

Besides data lineage, you want semantic metadata. What does a source actually represent as a business entity or process? To govern data semantically, we would need semantic tags at the source. This would enable us to know how to work with the data. For example, we could generate data vault, 3nf, star schema or activity schema models algorithmically starting from annotated json documents.

Besides business entities, domains or processes, semantic tags could also designate PII, security policies, or anything actionable. For example, PII tags could enable automatic lineage documentation and governance, while access tags could enable automatic access policies or automatic data modelling.

dlt currently supports column and row level lineage, as well as schema comments - which could be used as annotations.

The role of the Data platform engineer will grow​

In a shift left data democracy, the data platform engineer is a key character, as much as a CTO is in an organisation. By having a data platform engineer you ensure your data governance is done with automated tooling, to support implementation and compliance.

These data platform engineer becomes pivotal in empowering the democratization of data, providing the essential tooling and infrastructure that allow teams across the organization to manage their data autonomously.

Data platform engineers become enablers and facilitators, embedding governance and quality controls right from the start of the data lifecycle. Their work supports the organization by ensuring that data management practices are not only compliant and secure but also accessible and intuitive for non-specialists (democratic). This shift underlines a transition from centralized control to distributed empowerment, where data platform engineers support the broader goal of making data accessible, manageable, and governable across the entire spectrum of the organization.

The future of data management​

Are we heading towards semantically annotated data marts as code? Why not? We're in the age of serverless infrastructures, after all. Could data sociocracy become the future? Would we eventually encourage the entire organisation to annotate data sources with their learnings? Only time will tell.

Want to discuss?​

Join the dlt slack community to take part in the conversation.

The concept of simplicity and automation in a programming language is not new. Perl scripting language had the motto "Perl makes easy things easy and hard things possible".

The reason for this motto, was the difficulty of working with C, which requires more manual handling of resources and also a compilation step.

Perl scripts could be written and executed rapidly, making it ideal for tasks that needed quick development cycles. This ease of use and ability to handle complex tasks without cumbersome syntax made Perl incredibly popular in its heyday.

Perl was introduced as a scripting language that emphasized getting things done. It was created as a practical extraction and reporting tool, which quickly found its place in system administration, web development, and network programming.

History repeats, Python is a language for humans​

Python took the philosophy of making programming more accessible and human-friendly even further. Guido van Rossum created Python with the goal of removing the drudgery from coding, choosing to prioritize readability and simplicity. This design philosophy makes Python an intuitive language not just for seasoned developers but for beginners as well. Its syntax is clean and expressive, allowing developers to write fewer lines of code for tasks that would require more in Perl or other languages. Python's extensive standard library, along with its powerful data structures, contribute to its ability to handle complex applications with ease.

Python's widespread adoption across various domains, from web development to data science and machine learning, is largely attributed to its accessibility.

Its simple syntax resembles natural language, which lowers the barrier to entry for programming. Compared to Perl, Python offers an even more organized and readable approach to coding, making it an ideal teaching language that prepares new developers for future challenges in software development.

And just like perl, it's used for data extraction and visualisation - but now it's done by normie humans, not sysadmins or devs.

dlt makes easy things fast, and hard things accessible

Following the principles of Perl and Python, dlt aimed to simplify the data engineering process. dlt focuses on making the extraction and loading of data as straightforward as possible.

dlt makes easy things fast​

Starting from a simple abstraction like pipeline.run(data, table_name="table"), where data can be any iterable such as a generator or dataframe, dlt enables robust loading. Here is what the above function does, so you don't have to.

• It will (optionally) unpack nested lists into separate tables with generated join keys, and flatten nested dictionaries into a main row.
• If given a generator, it will consume it via microbatching, buffering to disk or external drives, never running out of memory (customisable).
• it will create "extract packages" of extracted data so if the downstream steps fail, it can resume/retry later.
• It will normalise the data into a shape that naturally fits the database (customisable).
• It will create "load packages" of normalised data so if the downstream steps fail, it can retry later.
• It infers and loads with the correct data types, for example from ISO timestamp strings (configurable).
• It can accept different types of write dispositions declaratively such as 'append', 'merge' and 'replace'.
• It will evolve the schema if we load a second time something with new columns, and it can alert the schema changes.
• It will even create type variant columns if data types change (and alert if desired).
• Or you can stop the schema from evolving and use the inferred schema or a modified one as a data contract
• It will report load packages associated with new columns, enabling passing down column level lineage

That's a lot of development and maintenance pain solved only at its simplest. You could say, the dlt loader doesn't break, as long as it encounters common data types. If an obscure type is in your data, it would need to be added to dlt or converted beforehand.

From robust loading to robust extraction​

Building on the simple loading abstraction, dlt is more than a tool for simple things.

The next step in dlt usage is to leverage it for extraction. dlt offers the concepts of 'source' and 'resource', A resource is the equivalent of a single data source, while a source is the group we put resources in to bundle them for usage.

For example, an API extractor from a single API with multiple endpoints, would be built as a source with multiple resources.

Resources enable you to easily configure how the data in that resource is loaded. You can create a resource by decorating a method with the '@resource' decorator, or you can generate them dynamically.

Examples of dynamic resources

• If we have an api with multiple endpoints, we can put the endpoints in a list and iterate over it to generate resources
• If we have an endpoint that gives us datapoints with different schemas, we could split them by a column in the data.
• Similarly, if we have a webhook that listens to multiple types of events, it can dispatch each event type to its own table based on the data.
• Or, if we want to shard a data stream into day-shards, we could append a date suffix in the resource name dynamically.

Once we group resources into a source, we can run them together (or, we could still run the resources independently)

Examples of reasons to group resources into sources.

• We want to run (load) them together on the same schedule
• We want to configure them together or keep their schemas together
• They represent a single API and we want to publish them in a coherent, easy to use way.

So what are the efforts you spare when using dlt here?

• A source can function similar to a class, but simpler, encouraging code reuse and simplicity.
• Resources offer more granular configuration options
• Resources can also be transformers, passing data between them in a microbatched way enabling patters like enrichments or list/detail endpoints.
• Source schemas can be configured with various options such as pushing down top level columns into nested structures
• dlt's requests replacement has built in retries for non-permanent error codes. This safeguards the progress of long extraction jobs that could otherwise break over and over (if retried as a whole) due to network or source api issues.

What else does dlt bring to the table?​

Beyond the ease of data extraction and loading, dlt introduces several advanced features that further simplify data engineering tasks:

Asynchronous operations: dlt harnesses the power of asynchronous programming to manage I/O-bound and network operations efficiently. This means faster data processing, better resource utilization, and more responsive applications, especially when dealing with high volumes of data or remote data sources.

Flexible destinations and reverse ETL: dlt isn't just about pulling data in; it's about sending it where it needs to go. Whether it's a SQL database, a data lake, or a cloud-based storage solution or a custom reverse etl destination, dlt provides the flexibility to integrate with various destinations.

Optional T in ETL: With dlt, transformations are not an afterthought but a core feature. You can define transformations as part of your data pipelines, ensuring that the data is not just moved but refined, enriched, and shaped to fit your analytical needs. This capability allows for more sophisticated data modeling and preparation tasks to be streamlined within your ELT processes.

Data quality and observability: dlt places a strong emphasis on data quality and observability. It includes features for schema evolution tracking, data type validation, and error handling, and data contracts, which are critical for maintaining the integrity of your data ecosystem. Observability tools integrated within dlt help monitor the health and performance of your pipelines, providing insights into data flows, bottlenecks, and potential issues before they escalate.

Community and ecosystem: One of the most significant advantages of dlt is its growing community and ecosystem. Similar to Python, dlt benefits from contributions that extend its capabilities, including connectors, plugins, and integrations. This collaborative environment ensures that dlt remains at the forefront of data engineering innovation, adapting to new challenges and opportunities.

In essence, dlt is not just a tool but a comprehensive one stop shop that addresses the end-to-end needs of modern data ingestion. By combining the simplicity of Python with the robustness of enterprise-grade tools, dlt democratizes data engineering, making it accessible to a broader audience. Whether you're a data scientist, analyst, or engineer, dlt empowers you to focus on what matters most: deriving insights and value from your data.

Conclusion​

As Perl and Python have made programming more accessible, dlt is set to transform data engineering by making sophisticated data operations accessible to all. This marks a significant shift towards the democratization of technology, enabling more individuals to contribute to and benefit from the digital landscape. dlt isn't just about making easy things fast and hard things accessible; it's about preparing a future where data engineering becomes an integral part of every data professional's toolkit.

Why Python is the right approach for doing Reverse ETL​

Reverse ETL is generally about putting data into a business application. This data would often come from a SQL database used as a middle layer for data integrations and calculations.

That’s fine - but nowadays most data people speak Python, and the types of things we want to put into an operational application don’t always come from a DB, they often come from other business applications, or from things like a dataframe on which we did some scoring, etc.

The full potential of Reverse ETL is in the flexibility of sources​

SQL databases are a good start, but in reality very often our data source is something else. More often than not, it’s a Python analyst’s implementation of some scoring or some business calculation.

Other times, it’s a business application - for example, we might have a form that sends the response data to a webhook, from where it could end up in Salesforce, DWH, and Slack as a notification. And of course, if this is done by a data person it will be done in Python.

Such, it follows that if we want to cater to the data crowd, we need to be Pythonic.

There’s synergy with ETL​

Reverse ETL is ultimately ETL. Data is extracted from a source, its transformed, and then loaded to a destination. The challenges are similar, the most notable difference being that pulling data from a strongly typed environment like a DB and converting it to weakly typed JSON is MUCH easier than the other way around. You can argue that Reverse ETL is simpler than ETL.

Flavors of Reverse ETL​

Just like we have ETL and ELT, we also have flavors of Reverse ETL

• Reverse ETL or TEL: Transform the data to a specification, read it from DB, and send it to an application.
• Tool Reverse ETL or ETL: Extract from DB, map fields to destination in the tool, load to destination.
• Pythonic Freestyle Reverse ETL: You extract data from wherever you want and put it anywhere except storage/DB. Transformations are optional.

Examples of Python reverse ETL

• Read data from Mongo, do anomaly detection, and notify anomalies to Slack.
• Read membership data from Stripe, calculate the chance to churn, and upload to CRM for account managers.
• Capture a form response with a webhook and send the information to CRM, DWH, and Slack.

Add python? - new skills unlocked!​

So why is it much better to do reverse ETL in Python?

• Live Streaming and Flexibility: Python's ability to handle live data streams and integrate with various APIs and services surpasses the capabilities of SQL-based data warehouses designed for batch processing.
• End-to-End Workflow: Employing Python from data extraction to operational integration facilitates a streamlined workflow, enabling data teams to maintain consistency and efficiency across the pipeline.
• Customization and Scalability: Python's versatility allows for tailored solutions that can scale with minimal overhead, reducing the reliance on additional tools and simplifying maintenance.
• Collaboration and Governance: By keeping the entire data workflow within Python, teams can ensure better governance, compliance, and collaboration, leveraging common tools and repositories.

Example: Building a Custom Destination and a pipeline in under 1h​

Documentation used: Building a destination: docs SQL source: docs In this example, you will see why it’s faster to build a custom destination than set up a separate tool.

dlt allows you to define custom destination functions. You'll write a function that extracts the relevant data from your dataframe and formats it for the Notion API.

This example assumes you have set up Google Sheets API access and obtained the necessary credentials to authenticate.

Step 1: Setting Up Google Sheets API (10min)​

1. Enable the Google Sheets API in the Google Developers Console.
2. Download the credentials JSON file.
3. Share the target Google Sheet with the email address found in your credentials JSON file.

Step 2: Define the Destination method in its own file sheets_destination.py (20min)​

Install the required package for the Google API client:

pip install --upgrade google-api-python-client google-auth-httplib2 google-auth-oauthlib

Here’s how to define a destination function to update a Google Sheet. In our case we wrote a slightly complex function that checks the headers and aligns the columns with the existing ones before inserting:

import dlt
from google.oauth2.service_account import Credentials
from googleapiclient.discovery import build


@dlt.destination(batch_size=100)
def google_sheets(items,
                            table_schema,
                            sheets_id: str = dlt.config.value,
                            credentials_json: dict = dlt.secrets.value,
                            range_name: str = 'Sheet1'):
    """
    Send data to a Google Sheet.
    :param items: Batch of items to send.
    :param table_schema: Schema of the table (unused in this example but required by dlt).
    :param sheets_id: ID of the Google Sheet, retrieved from config.
    :param credentials_json: Google Service Account credentials, retrieved from secrets.
    :param range_name: The specific range within the Sheet where data should be appended.
    """
    credentials = Credentials.from_service_account_info(credentials_json)
    service = build('sheets', 'v4', credentials=credentials)

    # Fetch existing headers from the sheet
    existing_headers_result = service.spreadsheets().values().get(
        spreadsheetId=sheets_id, range="Sheet1!A1:1"
    ).execute()
    existing_headers = existing_headers_result.get('values', [[]])[0] if existing_headers_result.get('values') else []

    # Determine new headers from items
    new_keys = set().union(*(d.keys() for d in items))
    # Identify headers that need to be added (not already existing)
    headers_to_add = [key for key in new_keys if key not in existing_headers]
    # New comprehensive headers list, preserving the order of existing headers and adding new ones at the end
    comprehensive_headers = existing_headers + headers_to_add

    # If there are headers to add, update the first row with comprehensive headers
    if headers_to_add:
        update_body = {'values': [comprehensive_headers]}
        service.spreadsheets().values().update(
            spreadsheetId=sheets_id, range="Sheet1!A1",
            valueInputOption='RAW', body=update_body
        ).execute()

    # Prepare the data rows according to the comprehensive headers list
    values = []
    for item in items:
        row = [item.get(header, "") for header in comprehensive_headers]  # Fill missing keys with empty string
        values.append(row)

    body = {'values': values}

    # Append the data rows
    if values:
        append_body = {'values': values}
        append_result = service.spreadsheets().values().append(
            spreadsheetId=sheets_id, range=range_name,
            valueInputOption='RAW', insertDataOption='INSERT_ROWS', body=append_body
        ).execute()
        print(f"{append_result.get('updates').get('updatedRows')} rows have been added to the sheet.")

Step 3: Configure secrets (5min)​

For the custom destination, you can follow this example. Configure the source as instructed in the source documentation.

secrets.toml

[destination.google_sheets]
credentials_json = '''
{
  "type": "service_account",
  "project_id": "your_project_id",
  "private_key_id": "your_private_key_id",
  ...
}
'''

config.toml

sheets_id = "1xj6APSKhepp8-sJIucbD9DDx7eyBt4UI2KlAYaQ9EKs"

Step 4: Running the pipeline in sheets_destination.py(10min)​

Now, assuming you have a source function dict_row_generator(), you can set up and run your pipeline as follows:

# ... destination code from above

# pass some destination arguments explicitly (`range_name`)
pipeline = dlt.pipeline("my_google_sheets_pipeline", destination=google_sheets(range_name="named_range"))

# Use the source function and specify the resource "people_report"
def dict_row_generator():
    yield {"row": 1, 'a': "a"}
    yield {"row": 2, 'b': "b"}
    yield {"row": 3, 'c': "c"}
    yield {"row": 1, 'a': 1}
    yield {"row": 2, 'b': 2}
    yield {"row": 3, 'c': 3}



# Now, run the pipeline with the specified source
info = pipeline.run(dict_row_generator)

In this setup, append_to_google_sheets acts as a custom destination within your dlt pipeline, pushing the fetched data to the specified Google Sheet. This method enables streamlined and secure data operations, fully utilizing Python's capabilities for Reverse ETL processes into Google Sheets.

What does dlt do for me here?​

Using dlt for reverse ETL instead of plain Python, especially with its @dlt.destination decorator, provides a structured framework that streamlines the process of data integrating into various destinations. Here’s how the dlt decorator specifically aids you compared to crafting everything from scratch in plain Python:

Faster time to Production grade pipelines​

The @dlt.destination decorator significantly reduces the need for custom boilerplate code. It provides a structured approach to manage batch processing, error handling, and retries, which would otherwise require complex custom implementations in plain Python. This built-in functionality ensures reliability and resilience in your data pipelines.

Focus on custom business logic and adding value​

The flexibility of creating custom destinations with dlt shifts the focus from the possibilities to the necessities of your specific use case. This empowers you to concentrate on implementing the best solutions for your unique business requirements.

Scalability and efficient resource use​

dlt facilitates efficient handling of large data loads through chunking and batching, allowing for optimal use of computing resources. This means even small worker machines can stream data effectively into your chosen endpoint instead of wasting a large machine waiting for the network. The library design supports easy scaling and adjustments. Making changes to batch sizes or configurations is straightforward, ensuring your data pipelines can grow and evolve with minimal effort. This approach simplifies maintenance and ensures that once a solution is implemented, it's broadly applicable across your projects.

In Conclusion​

Reverse ETL is just a piece of the ETL puzzle. It could be done cleaner and better when done in Python end to end.

Tools will always appeal to the non-technical folks. However, anyone with the ability to do Python pipelines can do Reverse ETL pipelines too, bringing typical benefits of code vs tool to a dev team - customisation, collaboration, best practices, etc.

So read more about how to built a dlt destination and consider giving it a try in your next reverse ETL pipeline.