For the modern database professional, the “maintenance trap” is a pervasive reality that stifles career growth and business impact. When your day is consumed by patching, manual tuning, and reactive troubleshooting, you aren’t architecting the future—you’re just keeping the lights on. The numbers confirm this stagnation: 72% of IT budgets are currently swallowed by generic maintenance rather than innovation.
However, we have reached a tipping point where the value scale is tilting. AI is not a replacement for the database expert; it is the long-awaited engine of liberation. Through the convergence of Retrieval Augmented Generation (RAG) and Autonomous systems, the traditional DBA is being reimagined as a hybrid strategist. This shift allows you to stop querying rows and start querying reason, moving from a technician of records to an architect of intelligence.
You’re Already 80% of a Data Scientist (Without Realizing It)
There is a persistent myth that database professionals must start from zero to enter the world of machine learning. The reality is far more empowering: you have already mastered the most difficult phase of the discipline. Industry data reveals that most data scientists spend 80% of their time finding, cleaning, and reorganizing data—a process known as Data Wrangling.
As a database expert, you are already an elite “wrangler.” The strategic pivot now is shifting these intensive tasks to the database itself. By transforming the database into a hybrid data management + machine learning platform, the professional evolves into a high-value AI Engineer or Data Engineer. You are the ideal candidate for these roles because you understand the underlying data structures better than anyone else.
“Most data scientists spend 80 percent of their time on tasks other than analysis, which is a massive inefficiency. Shifting these tasks to the database provides freedom from drudgery and allows the professional to focus on high-impact strategy.”
The “Self-Driving” Database is the Ultimate Career Insurance
The rise of the Autonomous Database is the ultimate insurance policy for your career. By automating the mechanical aspects of data management, these systems utilize three critical pillars:
Self-Driving: Automatically handles provisioning, monitoring, and tuning.
Self-Securing: Provides active protection against external attacks and malicious internal actors.
Self-Repairing: Maximizes uptime by protecting against planned and unplanned maintenance.
The business imperative is undeniable. Database downtime costs an average of $7,900 per minute, and 91% of organizations experience unplanned data center outages. Furthermore, 85% of security breaches occur after a CVE has already been published. By offloading these high-stakes, repetitive tasks to an autonomous system, you reclaim the bandwidth to focus on Architecture, planning, and data modeling. You aren’t losing your job; you are losing the tasks that make your job tedious.
SQL to JSON: The Secret Bridge to Large Language Models
As organizations race to implement Retrieval Augmented Generation (RAG), the database professional becomes the critical link in the AI supply chain. RAG enables Large Language Models (LLMs) to reason over private, enterprise data, but this requires a specialized technical bridge.
The surprising key to this architecture is the conversion of structured SQL results into JSON format. Because LLMs require context in a semi-structured format, the database professional now acts as the guardian of schema context. You are responsible for retrieving specific data and packaging it as a private, structured context that prevents the “hallucinations” common in generic AI. These Augmented Prompts—which combine precise user instructions with retrieved database context—are rapidly becoming the “stored procedures” of the AI era.
Move the Algorithms, Not the Data
The traditional “Data Lake” approach of moving massive datasets to external analytical tools is increasingly obsolete. Our new mantra is: “Move the Algorithms, Not the Data!” By utilizing In-database machine learning (OML), you can execute complex models directly where the data lives.
This shift enables unprecedented scale. For instance, using SPARC M8-2 hardware and the Airline On-Time dataset, systems have demonstrated the ability to process 640 million rows in-memory. Modern database professionals can now perform Feature Engineering—creating derived attributes that reflect domain knowledge—and execute models for Clustering, Anomaly Detection, Time Series Forecasting, and Regression using simple SQL syntax. This eliminates the security risks of data movement and brings Analytical Maturity to the core of the data center.
The Six-Week Transformation Roadmap
The transition from a Database Developer to a Data Scientist is a structured evolution, not a leap into the unknown. This six-week roadmap aligns your existing skills with the Analytical Maturity model:
Week 1: Business Understanding – Identify the core organizational problem.
Week 2: Data Understanding – Explore and profile available data assets.
Week 3: Data Preparation – Leverage your Data Wrangling expertise as the primary driver of project success.
Week 4: Modeling – Apply in-database ML algorithms.
Week 5: Evaluation – Rigorously test the accuracy of insights.
Week 6: Deployment – Move from Diagnostic Analysis (“What happened?”) to ML-Enabled Applications (“What will happen?”).
By following this path, you move beyond simple reporting and begin building Automated ML Applications that provide predictive value to the business.
Conclusion: The Choice to Innovate
We are entering the age of the “Thinking Database.”The industry is moving toward a future where the heavy lifting of maintenance is handled by the system itself, while the innovation is handled by you. Tools like OML Notebooks and Apache Zeppelin are now standard, accessible through the languages you already speak: SQL, Python, and R.
The choice for the database professional is clear. As the “Self-Driving” era takes hold, your value will no longer be measured by how well you maintain the engine, but by where you choose to drive the vehicle. When the database starts managing itself, will you use your new freedom to build the next generation of intelligent applications, or will you keep looking for a better wrench?
For years, enterprise AI has lived in a protected sandbox. It was the era of the “pilot,” a time defined by low-stakes experimentation and “innovation at any cost.” But as we enter 2026, that era is officially dead. The transition to autonomous, agent-driven systems has hit a hard ceiling: the realization that innovation without control is a structural liability.
The “data chaos” that once served as mere operational friction has mutated into a fundamental threat to the business. Organizations are discovering that the velocity of their AI is capped by the integrity of their data foundations. We have shifted from a post-GDPR world of reactive compliance to a high-stakes environment where Accountability is the only currency that matters.
This transformation is driven by a convergence of maturing technologies and a heavy-handed regulatory reality. Enterprises are no longer asking if they canbuild it; they are asking if they can prove its origin, quality, and safety. In 2026, the competitive edge belongs to those who stopped chasing “more data” and started building a governed foundation for the age of autonomy.
2. Governance is No Longer a Burden—It’s the Engine
August 2026 marks the first major enforcement cycle of the EU AI Act, and the shockwaves are being felt globally. Under Article 10, high-risk AI systems must meet rigorous quality criteria for training, validation, and testing datasets. Governance has evolved from a “reactive defense” tax into a “proactive competitive edge.”
A crucial strategic shift within Article 10 is the newly “legalized” use of sensitive data for the sake of fairness. Paragraph 5 allows providers to process special categories of personal data strictly for bias detection and correction, provided they meet stringent safeguards. This marks a pivot toward using governance as a tool for engineering social and technical trust.
To manage this, enterprises are establishing AI Governance Officers and adopting frameworks like ISO/IEC 42001 and the NIST AI RMF. These roles oversee model inventories and risk assessments, ensuring that intelligence is not just powerful, but sustainable and audit-ready.
“True intelligence must be portable, open, and sovereign—because your ability to move, scale, and adapt is what determines your competitive edge.” — Brett Sheppard
3. The Unstructured Data Goldmine: From Messy Files to Vector Reality
While 90% of enterprise data is unstructured—think images, video, and billions of PDFs—less than 1% was utilized for GenAI just two years ago. In 2026, the goldmine is finally open. The key has been the rise of Unstructured Data Integration (UDI) and Unstructured Data Governance (UDG).
This isn’t just about file storage; it’s about making legacy documents “agent-ready.” UDI pipelines now automate text chunking, embedding generation, and vectorization, allowing messy inputs to be ingested directly into vector databases. This enables Retrieval-Augmented Generation (RAG) at a scale that was previously impossible.
By unlocking these assets, companies are powering a new wave of Agentic AI capable of real-time risk detection and sophisticated document analysis. The goal is no longer just “search”—it is the conversion of raw organizational knowledge into actionable intelligence.
4. The Great Rapprochement: The Hybrid “Meshy Fabric”
The architectural civil war between Data Fabric and Data Mesh has ended in a hybrid marriage. Organizations that fell into the “velocity trap”—focusing on decentralization (Mesh) without automated infrastructure (Fabric)—found themselves buried in inconsistency. The most successful 2026 enterprises use a Data Fabric to automate intelligence while using a Data Mesh to enforce domain-led ownership.
Architectural Pivot
Data Fabric (Automation Layer)
Data Mesh (People/Process)
Strategic Driver
Unifying distributed systems via active metadata.
Managing data as a product with domain accountability.
Implementation
Technology-centric; automated integration.
Organizational-centric; domain-owned governance.
Key Enabler
Augmented data catalogs and AI-driven mapping.
Self-serve platforms and federated standards.
This “meshy fabric” ensures that the Data Fabricprovides the intelligent connective tissue, while the Data Mesh ensures the human domain experts are accountable for the quality of the data products being fed into AI agents.
5. Synthetic Data: The “Privacy-First” Training Hack
The “Privacy Paradox”—the friction between the need for massive datasets and the legal mandates of the GDPR—has been bypassed via Privacy Enhancing Technology (PET). Synthetic data, which mirrors the statistical patterns of real-world datasets without copying individual identities, has moved into the mainstream.
Beyond privacy, synthetic data is now a primary tool for bias mitigation. It allows developers to fill “data gaps” and create “edge cases” that real-world datasets often ignore. In sectors like healthcare and finance, this mimics the statistical properties required for high-utility models without the risk of re-identification or regulatory exposure.
“Synthetic data can be defined as data that has been generated from real data and that has the same statistical properties as the real data.” — Dr. Khaled El Emam
6. “Agent-Ready” Data and the Science of Model Provenance
As AI evolves toward Agentic AI—systems that act autonomously in procurement or IT operations—the demand for Accountability has reached a fever pitch. For an agent to execute a contract, it must have “agent-ready” data: information that is traceable, high-quality, and context-rich.
Simultaneously, the industry is moving from heuristic fingerprinting to mathematical proof. Using the Model Provenance Set (MPS), a sequential test-and-exclusion procedure, organizations can now achieve a provable asymptotic guarantee of a model’s lineage.
This isn’t just a tool; it’s a statistical proof. It allows enterprises to detect unauthorized reuse and protect intellectual property by identifying related models in complex derivation chains. In 2026, you don’t just “verify” a model; you prove its provenance.
7. Sovereignty is the New Architecture
Cloud strategy has shifted from a matter of IT efficiency to a compliance and risk management obligation. Driven by the EU Data Act, organizations are pivoting toward Sovereign Multicloud Architectures. This isn’t just about local hosting; it’s about the legal mandate of “fair cloud switching” and “vendor neutrality.”
The EU Data Act has fundamentally changed data sharing by mandating new rights for data access and portability. This has forced a mass redesign of data-sharing processes and vendor contracts. In 2026, the question of “where your data sits” is a matter of sovereignty.
Public sector and finance leaders are leading this charge, moving critical workloads to certified sovereign environments. They recognize that in the age of autonomous AI, control over the underlying infrastructure is the only way to mitigate the risk of vendor lock-in and geopolitical friction.
8. Conclusion: The Trust Dividend
The digital economy of the next decade is being built on the foundations we lay today. By 2026, the convergence of Governance, Sovereignty, and Automation has created a “Trust Dividend.” Those who invested in making their data agent-ready and audit-proof are now scaling autonomous systems with a level of confidence their competitors can’t match.
As we look toward an increasingly autonomous future, the question for every technical leader has shifted:
Is your data estate merely a collection of assets, or is it a governed foundation ready for the age of autonomy?
Deploying agentic workflows is no longer a luxury for the modern creator; it is the baseline for survival in a field that moves faster than most can read. As a Senior Technical Content Strategist, I focus on systems that actually perform. I’m Ira Warren Whiteside, and my perspective on AI and Agentic AI isn’t theoretical—it’s built into my daily architecture. This shift toward high-efficiency workflows became a necessity during a recent recovery period. While my throat was healing from extreme weight lee (loss), I had to ensure my output remained high-fidelity without the luxury of manual, exhaustive research sessions.
The challenge is the “Creator’s Dilemma”: how to manage research-heavy technical projects while staying at the cutting edge of a relentless industry. The solution lies in treating AI not as a ghostwriter, but as a sophisticated research and synthesis layer that bridges the gap between deep technical archives and publication-ready insights.
1. Speed as a Competitive Advantage
In a technical ecosystem, speed is the ultimate competitive advantage. NotebookLM serves as a powerful catalyst for this, functioning as a specialized engine for rapid synthesis. By offloading the heavy lifting of initial research and document correlation, the platform allows a strategist to bypass the friction of manual data sorting.
Reducing the time spent on manual synthesis shifts the focus where it belongs: on high-level strategy and technical exploration. When you aren’t bogged down in the mechanics of organization, you are free to find the narrative within the data. As my recent workflow proves, this approach:
“speeds up research… saves time… excellent creators workflow.”
2. Turning Your Archives into a Discovery Engine
Generic AI models provide generic results. To produce truly authoritative content, you must mine your own intellectual property. This workflow uses the tool as a mirror, bringing out new discoveries based specifically on my own writings, ideas, and targeted prompts. It creates a closed-loop feedback system where past logic informs future innovation.
This is far more valuable than a standard LLM query; it ensures the output is grounded in a unique perspective rather than a homogenized dataset. It allows the creator to see patterns in their own thinking that might otherwise remain buried in thousands of lines of documentation.
Exploration through Variety: The system produces a wide variety of outputs—from summaries to deep-dive briefings—enabling a more comprehensive exploration of complex technical topics.
3. Bridging the Gap: From AI to RDBMS
For a Technical Insider, a workflow must handle more than just prose. It must integrate seamlessly with structured engineering data. My process bridges the gap between creative synthesis and the world of RDBMS STATISTICS, T-SQL SCRIPTS, AND SERVICES FROM METADATA MECHANICS.
This isn’t just about storing scripts; it’s about using AI to interpret technical metadata. It’s the ability to turn a raw T-SQL execution plan or a complex database schema into a high-level architectural narrative. By processing these technical artifacts through an intelligent workflow, I can generate documentation and insights that are as functionally accurate as they are readable.
METADATA MECHANICS represents the intersection of structured data and narrative strategy. This “clean aesthetic” in data management allows me to move from raw database statistics to polished technical blogging without losing the underlying technical rigor.
4. Grounding Insights in Reality
The primary risk of AI-integrated writing is the “hallucination”—the confident assertion of a technical falsehood. In technical blogging, credibility is the only currency that matters. This workflow mitigates that risk by ensuring that “references are included” for every generated insight.
Direct citations back to the source context are the essential antidote to AI errors. When writing about complex RDBMS behaviors or specific T-SQL implementations, having a clickable path back to the source material ensures that every claim is verified. This grounding transforms an AI tool from a creative assistant into a reliable technical partner.
The Future of the Intelligent Workflow
Integrating a tech-focused AI workflow allows a creator to explore and keep up with new technology while maintaining a rigorous publishing cadence. By leveraging these agentic systems, we move beyond simple content creation and into the realm of intellectual discovery.
As you evaluate your own technical output, ask yourself: how are you integrating your own METADATA MECHANICS into your creative process? The goal is to move past the manual synthesis bottleneck and begin gaining deeper, data-driven insights from the archives you’ve already built.
Beyond the Dashboard: 5 Surprising Truths About the New Era of Analytics Engineering
1. The Death of Artisanal Data and the Industrial Revolution
The world of 1974, when the first relational database was defined, moved at the speed of a mail-order catalog. You posted a check and waited weeks for delivery. For decades, data processing mirrored this “artisanal” cadence—bespoke, slow, and manual. Today, that world is gone. We live in a “data-in-motion” reality where software talks to other software 24/7, generating an unrelenting stream of events.
The wall between the isolated analyst and the siloed engineer has been demolished by necessity. We are witnessing the end of “cowboy coding”—the era of unchecked manual scripts and fragile pipelines. In its place, analytics has evolved into a high-stakes engineering discipline. While our tools have transitioned from manual entries to industrialized pipelines, the fundamental need for rigorous data modeling remains the core of this revolution. To survive the modern era, organizations must stop treating data as a collection of one-off projects and start treating it as a precision manufacturing process.
2. The “Stark-Holmes” Hybrid: Why Deduction and Engineering Must Merge
The modern Analytics Engineer is a rare hybrid, blending two seemingly disparate archetypes: the meticulous investigator Sherlock Holmes and the genius engineer Tony Stark.
Success in this field requires the deductive reasoning of Holmes—using keen observation to identify the core of a business challenge before a single line of code is written—fused with Stark’s software engineering mastery. This role isn’t just about moving data; it’s about applying the foundational strengths of software engineering to the pursuit of knowledge.
“Analytics engineering is more than just technology: it’s a management tool that will be successful only if it’s aligned with your organization’s strategies and goals.” — Rui Machado & Hélder Russa, Analytics Engineering with SQL and dbt
By adopting this mindset, the Analytics Engineer ensures the data value chain is resilient, turning raw data into the “original facts” that illuminate the current state of the business.
3. Pragmatic SQL: Why “Sloppy” Code is Smarter at Scale
In the traditional world, query correctness was binary: you were either right or you were wrong. In the era of LLM-driven interfaces and “Text-to-Big SQL,” we must embrace the counter-intuitive reality of partial correctness.
When running queries on engines like Amazon Athena or BigQuery, the traditional obsession with “clean” SQL is a cost-center. If an LLM-generated query includes “superfluous columns,” it is often more cost-effective to drop those columns in a downstream tool like Spark than to pay for a full re-execution on a massive dataset. To measure this, we use the VES* (Valid Efficiency Score) and VCES (Valid Cost-Efficiency Score).
Crucially, VES* accounts for the total end-to-end time (Te2e), which includes the back-and-forth interactions between the LLM and the agent. Our research shows that “Both Ends Count”—generation and execution. For example, while models like Opus 4.6 achieve perfect accuracy, they can take 92.37% longer to return a result than GPT-4o. In interactive analytics, “fast” often beats “perfect.”
The Scale Factor:
Small Scale (SF10): Agent reasoning and tool interaction dominate the latency.
Large Scale (SF1000): Physical query execution on the engine becomes the bottleneck. At this scale, even a 10% accuracy gap becomes a massive financial liability, as failed queries at SF1000 are exponentially more expensive than at SF10.
4. “A Car Needs Brakes to Go Fast”: The Paradox of DataOps
There is a persistent myth that testing is a bottleneck. In reality, it is your greatest accelerator. As Harvinder Atwal famously noted, “A car needs brakes to go fast.”Without the “brakes” of a rigorous testing framework, teams are forced to move slowly to avoid breaking production.
Industrializing the data chain requires a radical shift in resource allocation. While traditional teams typically devote only 20% of their effort to quality, modern DataOps teams devote 50% of their code and staffto testing and development velocity. To move from “Cowboy” to “Industrial,” you must implement three essential test types:
Input Tests: Verifying counts, conformity (e.g., Zip codes), and consistency before data enters a pipeline node.
Business Logic Tests: Validating that data matches business assumptions (e.g., ensuring every customer exists in a dimension table).
Output Tests: Checking the results of operations (e.g., ensuring row counts are within expected ranges after a cross-product join).
5. SQL’s Second Act: Tables Only Tell Half the Story
The industry is shifting from “data-passive” to “data-active” architectures. Traditionally, SQL was designed for data at rest (Tables), but the future belongs to data in motion (Streams).
The distinction is fundamental: Streams tell the story of how we got here, while Tables only tell the current state of the world. This shift transforms our query complexity from being a function of the data size to a function of the data’s velocity.
Pull Queries (Traditional)
Push Queries (Modern Streaming)
Termination: Terminate once a bounded result is returned.
Persistence: Run forever until explicitly terminated.
Execution: Requires full table scans or index lookups.
Incremental: Computes “deltas” and incremental updates.
Latency: Client must re-submit query to see changes.
Real-time: Results are “pushed” to the client immediately.
Complexity: Linear cost based on table size: O(N).
Complexity: Linear cost based on update frequency: O(rate).
6. The dbt Revolution: Enabling the Data Mesh
The shift from warehouses to data lakes allowed data to land before transformation, creating a desperate need for a self-service platform where analysts could model raw data. dbt (data build tool) has emerged as the primary “Data Mesh Enabler,” allowing teams to focus on value delivery rather than architectural maintenance.
To build meaningful models at scale, we use the Medallion Architecture:
Bronze (Raw): Landing zone for raw data.
Silver (Transformed): Cleaned, filtered, and joined data ready for analysis.
Gold (Curated): Highly polished, business-ready datasets optimized for consumption.
As leading architects Jacob Frackson and Michal Kolacek suggest:
“If your team is struggling with inefficient views, tangled stored procedures, or low analytics adoption… this book will help you see a new way forward.”
7. Conclusion: Introspection Over Extravagance
Modern analytics is defined by mindset, not by the complexity of your Python scripts. The goal is to solve business problems with precision and pragmatism. As you build your infrastructure, remember this final warning:
“Avoid building an extravagant aircraft when a humble bicycle would suffice.”
Let the complexity of the problem guide your efforts, not the lure of the latest algorithm. Is your organization still relying on “hope as a strategy,” or are you ready to industrialize your data value chain?
Beyond the Dashboard: 5 Surprising Truths About the New Era of Analytics Engineering
1. The Death of Artisanal Data and the Industrial Revolution
The world of 1974, when the first relational database was defined, moved at the speed of a mail-order catalog. You posted a check and waited weeks for delivery. For decades, data processing mirrored this “artisanal” cadence—bespoke, slow, and manual. Today, that world is gone. We live in a “data-in-motion” reality where software talks to other software 24/7, generating an unrelenting stream of events.
The wall between the isolated analyst and the siloed engineer has been demolished by necessity. We are witnessing the end of “cowboy coding”—the era of unchecked manual scripts and fragile pipelines. In its place, analytics has evolved into a high-stakes engineering discipline. While our tools have transitioned from manual entries to industrialized pipelines, the fundamental need for rigorous data modeling remains the core of this revolution. To survive the modern era, organizations must stop treating data as a collection of one-off projects and start treating it as a precision manufacturing process.
2. The “Stark-Holmes” Hybrid: Why Deduction and Engineering Must Merge
The modern Analytics Engineer is a rare hybrid, blending two seemingly disparate archetypes: the meticulous investigator Sherlock Holmes and the genius engineer Tony Stark.
Success in this field requires the deductive reasoning of Holmes—using keen observation to identify the core of a business challenge before a single line of code is written—fused with Stark’s software engineering mastery. This role isn’t just about moving data; it’s about applying the foundational strengths of software engineering to the pursuit of knowledge.
“Analytics engineering is more than just technology: it’s a management tool that will be successful only if it’s aligned with your organization’s strategies and goals.” — Rui Machado & Hélder Russa, Analytics Engineering with SQL and dbt
By adopting this mindset, the Analytics Engineer ensures the data value chain is resilient, turning raw data into the “original facts” that illuminate the current state of the business.
3. Pragmatic SQL: Why “Sloppy” Code is Smarter at Scale
In the traditional world, query correctness was binary: you were either right or you were wrong. In the era of LLM-driven interfaces and “Text-to-Big SQL,” we must embrace the counter-intuitive reality of partial correctness.
When running queries on engines like Amazon Athena or BigQuery, the traditional obsession with “clean” SQL is a cost-center. If an LLM-generated query includes “superfluous columns,” it is often more cost-effective to drop those columns in a downstream tool like Spark than to pay for a full re-execution on a massive dataset. To measure this, we use the VES* (Valid Efficiency Score) and VCES (Valid Cost-Efficiency Score).
Crucially, VES* accounts for the total end-to-end time (Te2e), which includes the back-and-forth interactions between the LLM and the agent. Our research shows that “Both Ends Count”—generation and execution. For example, while models like Opus 4.6 achieve perfect accuracy, they can take 92.37% longer to return a result than GPT-4o. In interactive analytics, “fast” often beats “perfect.”
The Scale Factor:
Small Scale (SF10): Agent reasoning and tool interaction dominate the latency.
Large Scale (SF1000): Physical query execution on the engine becomes the bottleneck. At this scale, even a 10% accuracy gap becomes a massive financial liability, as failed queries at SF1000 are exponentially more expensive than at SF10.
4. “A Car Needs Brakes to Go Fast”: The Paradox of DataOps
There is a persistent myth that testing is a bottleneck. In reality, it is your greatest accelerator. As Harvinder Atwal famously noted, “A car needs brakes to go fast.”Without the “brakes” of a rigorous testing framework, teams are forced to move slowly to avoid breaking production.
Industrializing the data chain requires a radical shift in resource allocation. While traditional teams typically devote only 20% of their effort to quality, modern DataOps teams devote 50% of their code and staffto testing and development velocity. To move from “Cowboy” to “Industrial,” you must implement three essential test types:
Input Tests: Verifying counts, conformity (e.g., Zip codes), and consistency before data enters a pipeline node.
Business Logic Tests: Validating that data matches business assumptions (e.g., ensuring every customer exists in a dimension table).
Output Tests: Checking the results of operations (e.g., ensuring row counts are within expected ranges after a cross-product join).
5. SQL’s Second Act: Tables Only Tell Half the Story
The industry is shifting from “data-passive” to “data-active” architectures. Traditionally, SQL was designed for data at rest (Tables), but the future belongs to data in motion (Streams).
The distinction is fundamental: Streams tell the story of how we got here, while Tables only tell the current state of the world. This shift transforms our query complexity from being a function of the data size to a function of the data’s velocity.
Pull Queries (Traditional)
Push Queries (Modern Streaming)
Termination: Terminate once a bounded result is returned.
Persistence: Run forever until explicitly terminated.
Execution: Requires full table scans or index lookups.
Incremental: Computes “deltas” and incremental updates.
Latency: Client must re-submit query to see changes.
Real-time: Results are “pushed” to the client immediately.
Complexity: Linear cost based on table size: O(N).
Complexity: Linear cost based on update frequency: O(rate).
6. The dbt Revolution: Enabling the Data Mesh
The shift from warehouses to data lakes allowed data to land before transformation, creating a desperate need for a self-service platform where analysts could model raw data. dbt (data build tool) has emerged as the primary “Data Mesh Enabler,” allowing teams to focus on value delivery rather than architectural maintenance.
To build meaningful models at scale, we use the Medallion Architecture:
Bronze (Raw): Landing zone for raw data.
Silver (Transformed): Cleaned, filtered, and joined data ready for analysis.
Gold (Curated): Highly polished, business-ready datasets optimized for consumption.
As leading architects Jacob Frackson and Michal Kolacek suggest:
“If your team is struggling with inefficient views, tangled stored procedures, or low analytics adoption… this book will help you see a new way forward.”
7. Conclusion: Introspection Over Extravagance
Modern analytics is defined by mindset, not by the complexity of your Python scripts. The goal is to solve business problems with precision and pragmatism. As you build your infrastructure, remember this final warning:
“Avoid building an extravagant aircraft when a humble bicycle would suffice.”
Let the complexity of the problem guide your efforts, not the lure of the latest algorithm. Is your organization still relying on “hope as a strategy,” or are you ready to industrialize your data value chain?
of Analytics Engineering
1. The Death of Artisanal Data and the Industrial Revolution
The world of 1974, when the first relational database was defined, moved at the speed of a mail-order catalog. You posted a check and waited weeks for delivery. For decades, data processing mirrored this “artisanal” cadence—bespoke, slow, and manual. Today, that world is gone. We live in a “data-in-motion” reality where software talks to other software 24/7, generating an unrelenting stream of events.
The wall between the isolated analyst and the siloed engineer has been demolished by necessity. We are witnessing the end of “cowboy coding”—the era of unchecked manual scripts and fragile pipelines. In its place, analytics has evolved into a high-stakes engineering discipline. While our tools have transitioned from manual entries to industrialized pipelines, the fundamental need for rigorous data modeling remains the core of this revolution. To survive the modern era, organizations must stop treating data as a collection of one-off projects and start treating it as a precision manufacturing process.
2. The “Stark-Holmes” Hybrid: Why Deduction and Engineering Must Merge
The modern Analytics Engineer is a rare hybrid, blending two seemingly disparate archetypes: the meticulous investigator Sherlock Holmes and the genius engineer Tony Stark.
Success in this field requires the deductive reasoning of Holmes—using keen observation to identify the core of a business challenge before a single line of code is written—fused with Stark’s software engineering mastery. This role isn’t just about moving data; it’s about applying the foundational strengths of software engineering to the pursuit of knowledge.
“Analytics engineering is more than just technology: it’s a management tool that will be successful only if it’s aligned with your organization’s strategies and goals.” — Rui Machado & Hélder Russa, Analytics Engineering with SQL and dbt
By adopting this mindset, the Analytics Engineer ensures the data value chain is resilient, turning raw data into the “original facts” that illuminate the current state of the business.
3. Pragmatic SQL: Why “Sloppy” Code is Smarter at Scale
In the traditional world, query correctness was binary: you were either right or you were wrong. In the era of LLM-driven interfaces and “Text-to-Big SQL,” we must embrace the counter-intuitive reality of partial correctness.
When running queries on engines like Amazon Athena or BigQuery, the traditional obsession with “clean” SQL is a cost-center. If an LLM-generated query includes “superfluous columns,” it is often more cost-effective to drop those columns in a downstream tool like Spark than to pay for a full re-execution on a massive dataset. To measure this, we use the VES* (Valid Efficiency Score) and VCES (Valid Cost-Efficiency Score).
Crucially, VES* accounts for the total end-to-end time (Te2e), which includes the back-and-forth interactions between the LLM and the agent. Our research shows that “Both Ends Count”—generation and execution. For example, while models like Opus 4.6 achieve perfect accuracy, they can take 92.37% longer to return a result than GPT-4o. In interactive analytics, “fast” often beats “perfect.”
The Scale Factor:
Small Scale (SF10): Agent reasoning and tool interaction dominate the latency.
Large Scale (SF1000): Physical query execution on the engine becomes the bottleneck. At this scale, even a 10% accuracy gap becomes a massive financial liability, as failed queries at SF1000 are exponentially more expensive than at SF10.
4. “A Car Needs Brakes to Go Fast”: The Paradox of DataOps
There is a persistent myth that testing is a bottleneck. In reality, it is your greatest accelerator. As Harvinder Atwal famously noted, “A car needs brakes to go fast.”Without the “brakes” of a rigorous testing framework, teams are forced to move slowly to avoid breaking production.
Industrializing the data chain requires a radical shift in resource allocation. While traditional teams typically devote only 20% of their effort to quality, modern DataOps teams devote 50% of their code and staffto testing and development velocity. To move from “Cowboy” to “Industrial,” you must implement three essential test types:
Input Tests: Verifying counts, conformity (e.g., Zip codes), and consistency before data enters a pipeline node.
Business Logic Tests: Validating that data matches business assumptions (e.g., ensuring every customer exists in a dimension table).
Output Tests: Checking the results of operations (e.g., ensuring row counts are within expected ranges after a cross-product join).
5. SQL’s Second Act: Tables Only Tell Half the Story
The industry is shifting from “data-passive” to “data-active” architectures. Traditionally, SQL was designed for data at rest (Tables), but the future belongs to data in motion (Streams).
The distinction is fundamental: Streams tell the story of how we got here, while Tables only tell the current state of the world. This shift transforms our query complexity from being a function of the data size to a function of the data’s velocity.
Pull Queries (Traditional)
Push Queries (Modern Streaming)
Termination: Terminate once a bounded result is returned.
Persistence: Run forever until explicitly terminated.
Execution: Requires full table scans or index lookups.
Incremental: Computes “deltas” and incremental updates.
Latency: Client must re-submit query to see changes.
Real-time: Results are “pushed” to the client immediately.
Complexity: Linear cost based on table size: O(N).
Complexity: Linear cost based on update frequency: O(rate).
6. The dbt Revolution: Enabling the Data Mesh
The shift from warehouses to data lakes allowed data to land before transformation, creating a desperate need for a self-service platform where analysts could model raw data. dbt (data build tool) has emerged as the primary “Data Mesh Enabler,” allowing teams to focus on value delivery rather than architectural maintenance.
To build meaningful models at scale, we use the Medallion Architecture:
Bronze (Raw): Landing zone for raw data.
Silver (Transformed): Cleaned, filtered, and joined data ready for analysis.
Gold (Curated): Highly polished, business-ready datasets optimized for consumption.
As leading architects Jacob Frackson and Michal Kolacek suggest:
“If your team is struggling with inefficient views, tangled stored procedures, or low analytics adoption… this book will help you see a new way forward.”
7. Conclusion: Introspection Over Extravagance
Modern analytics is defined by mindset, not by the complexity of your Python scripts. The goal is to solve business problems with precision and pragmatism. As you build your infrastructure, remember this final warning:
“Avoid building an extravagant aircraft when a humble bicycle would suffice.”
Let the complexity of the problem guide your efforts, not the lure of the latest algorithm. Is your organization still relying on “hope as a strategy,” or are you ready to industrialize your data value chain?
The software development landscape is undergoing its most dramatic transformation since the shift from assembly to high-level languages. By 2026, projections suggest that 90% of all code will be AI-generated. This reality has sparked a wave of anxiety, but the data tells a more nuanced story of bifurcation rather than obsolescence.
While entry-level tech hiring decreased by 25% year-over-year in 2024 and employment for developers aged 22–25 declined nearly 20%, the demand for senior talent capable of managing AI systems has reached a fever pitch. We are witnessing the death of the “Syntax Memorizer”—the 2022-style developer whose primary value was handwriting functional lines. In their place emerges the System Orchestrator: an engineer who leverages AI to deliver the output once expected from a team of ten.
Underneath the hype, a new layer of engineering work has emerged. This isn’t research or model training; it is product engineering where AI is a system component. If you are a full-stack architect looking to future-proof your career, the transition to becoming an AI engineer requires a deliberate evolution of your technical stack and mindset.
1. Prompting is Now “Table Stakes” (Master Context Engineering)
Many developers remain fixated on the surface layer: perfecting prompts or chasing the latest “hacks.” While prompt engineering was the buzzy role of 2023, it has rapidly become a standard capability, much like using an IDE or keyboard shortcuts.
The professional differentiator is no longer just the prompt; it is Context Engineering. This is the rigorous discipline of managing the non-prompt elements supplied to a model—metadata, API tool definitions, and token budgeting—to ensure reliability and provenance. Your value is shifting from a “Code Writer” to an architect of the environment in which the AI operates.
As Andrew Ng points out, you cannot simply “vibe code” your way to production-grade systems:
“Without understanding how computers work, you can’t just ‘vibe code’ your way to greatness. Fundamentals are still important, and for those who additionally understand AI, job opportunities are numerous!”
2. RAG is the Single Most Critical Skill (The Undervalued Infrastructure)
If you commit to one technical skill this year, make it Retrieval-Augmented Generation (RAG). While social media is captivated by flashy autonomous agents, RAG is the “undervalued infrastructure layer” that startups and enterprises are actually paying for.
RAG is the process of providing a Large Language Model (LLM) with proprietary data at the right time to prevent hallucinations. In practice, this involves:
Converting documents into embeddings(numerical vectors).
Managing vector databases like Pinecone or Qdrant for high-dimensional storage.
Designing semantic retrieval systems that allow models to interact with live, changing data.
This is the foundation of useful AI products. For example, when a DoorDash driver asks how to handle spilled pickle juice, a RAG system retrieves the specific internal protocol for vehicle maintenance to provide an accurate, human-readable answer. Similarly, Spotify uses these patterns to find songs with semantically similar lyrics. Mastering the “boring” plumbing of data flow is what separates a hobbyist from a $350k IC.
3. Workflows Over Agents (The “Deterministic” Advantage)
The term “AI Agent” is dangerously overloaded. In a hype-driven market, non-technical CEOs often demand “autonomous agents” that run until a task is done. In reality, these uncontrolled agentic loopsoften lead to exploding token costs and non-deterministic failures.
The superior architectural pattern is the controlled workflow. As an engineer, your job is to create deterministic outcomes in a non-deterministic world. This requires:
Human-in-the-loop patterns: Designing checkpoints for critical decisions.
Orchestration: Utilizing patterns like “ReAct” or “Orchestrator” to classify and route tasks programmatically.
FinOps Mindset: Implementing observability tools like Helicone or LangSmith to monitor token consumption and latency.
Having a technical opinion on workflows vs. agents is a superpower. Most companies are operating on “social media vibes”; the AI engineer provides the strategic direction and cost control necessary for enterprise scale.
4. The Return of the “CS Fundamentalist”
There is a persistent myth that AI makes Computer Science degrees obsolete. The reality is that as the cost of generating code drops to zero, the cost of the friction created by bad code—security flaws, technical debt, and architectural rot—skyrockets.
Andrew Ng notes that while 30% of traditional CS knowledge (like memorizing syntax) is fading, the remaining 70% is more vital than ever. You cannot verify or supervise AI-generated code if you do not understand the Critical Fundamentals:
Concurrency and Parallelism: Essential for managing asynchronous AI API calls and system throughput.
Memory and Performance Complexity: Vital for optimizing token usage and high-dimensional vector searches.
Networking Basics: Crucial for managing the distributed nature of modern AI services.
Deep technical knowledge is what builds the “design taste” required to know when to introduce an architectural principle and when to push back against a model’s suggestion.
5. Testing isn’t Dead—It Just Got a “Black Box” Problem
Traditional unit testing is insufficient for non-deterministic AI services. Because LLMs are “black boxes,” they require a new testing paradigm focused on Evals (evaluation sets).
Instead of testing for a specific string output, professional AI engineers utilize the LLM-as-a-judgepattern. By creating a “Gold Set” of ideal responses, you can use one LLM to score another’s output on a scale of 1 to 10. This allows you to:
Detect model drift or prompt regressions before they reach the user.
Safely upgrade or downgrade models (e.g., GPT-4o to a smaller, faster model) without breaking functionality.
Ensure that a minor prompt change by a teammate hasn’t compromised system logic.
Flying blind with non-deterministic services is a recipe for losing customer trust. A rigorous testing mindset is now the primary differentiator between an “AI Bro” and a professional engineer.
Conclusion: Crossing the 3-Month Gap
The transition from a standard full-stack developer to a high-earning AI Engineer is a marathon, but the initial competency gap can be bridged in roughly one to three months by following a structured roadmap:
Phase 1: Integrate & Accelerate (Month 1): Adopt AI pair programmers (Cursor, Copilot) and agentic review tools. Focus on moving from simple comments to structured context engineering.
Phase 2: Architect & Orchestrate (Months 2-3):Build a RAG-based application. Store proprietary data in a vector database and implement a controlled workflow using a framework like LangGraph or a manual “human-in-the-loop” pattern.
Phase 3: Strategize & Lead (Ongoing): Develop a quality framework using Evals and LLM-as-a-judge. Quantify your impact on team velocity and begin managing the technical debt that AI code inevitably generates.
In tech-forward hubs like San Francisco, senior individual contributors who master this orchestration are commanding salaries between $200,000 and $350,000.
The question is no longer whether AI will change your job, but how you will respond to the shift. Do you want to be the developer struggling to compete with AI-generated syntax, or the orchestrator designing the systems that command it?
1. Introduction: The Unseen Mechanics of the AI Revolution
Large Language Models (LLMs) have successfully transitioned from laboratory curiosities to ubiquitous enterprise tools. To the casual observer, the progress looks like a linear march toward increasingly “smarter” chatbots. However, the technical reality is far more nuanced. Behind the curtain of viral interfaces, the most impactful breakthroughs are no longer just about increasing parameter counts or ingestion volume. As a Research Strategist, I observe that the real frontier has shifted toward “unseen mechanics”—the sophisticated methods researchers use to steer, optimize, and ground these models to transform them from unpredictable black boxes into high-precision, reliable instruments.
2. The Operational Safety Gap: Why Your Agent “Enters the Wrong Chat”
A critical challenge for enterprise deployment is “operational safety.” While global discourse often focuses on preventing generic harms (e.g., assisting in illegal acts), operational safety addresses a model’s ability to remain faithful to its intended purpose. Recent research, specifically the OffTopicEvalbenchmark, reveals a startling reality: LLMs are prone to “entering the wrong chat.”
When tasked with a professional role—such as an AI bank teller—models frequently fail to refuse out-of-domain (OOD) queries, straying into discussions about poetry or travel advice. The data shows that even top-tier models struggle; Llama-3 and Gemma collapsed to accuracy levels of 23.84% and 39.53% respectively in agentic scenarios. Even GPT-4 plateaus in the 62–73% range. Interestingly, the benchmark identifies Mistral (24B) at 79.96% and Qwen-3 (235B) at 77.77% as the current leaders in operational reliability.
To suppress these failures without the overhead of retraining, researchers are utilizing prompt-based steering. Techniques like Query Grounding (Q-ground) provide consistent gains of up to 23%, while System-Prompt Grounding (P-ground) delivered a massive 41% boost to Llama-3.3 (70B).
“To suppress these failures, we propose prompt-based steering methods: query grounding (Q-ground) and system-prompt grounding (P-ground), which substantially improve OOD refusal. Q-ground provides consistent gains of up to 23%, while P-ground delivers even larger boosts.”
3. Surgical Alignment: Steering the “Brain” Without Retraining
A major obstacle in fine-tuning is the “superposition” problem: LLM neurons are semantically entangled, often responding to multiple unrelated factors. This makes standard fine-tuning messy, as adjusting one behavior (like bias) often accidentally degrades linguistic fluency.
The Sparse Representation Steering (SRS)framework offers a “surgical” alternative. Using Sparse Autoencoders (SAEs), SRS projects dense activations (n) into a significantly higher-dimensional sparse feature space (m>n). This allows researchers to disentangle activations into millions of monosemantic features. To identify exactly which features to “turn up or down,” SRS utilizes bidirectional KL divergencebetween contrastive prompt distributions to quantify per-feature sensitivity.
This level of precision, often characterized by the L0 norm (the number of non-zero elements), allows developers to modulate specific attributes like truthfulness or safety at inference time with minimal side effects on overall quality.
“Due to the semantically entangled nature of LLM’s representation, where even minor interventions may inadvertently influence unrelated semantics, existing representation engineering methods still suffer from… content quality degradation.”
4. The 20% Rule: Efficiency via the “Heavy Hitter Oracle”
Deploying LLMs at scale is hindered by the KV Cache bottleneck. Because the cache scales linearly with sequence length, long conversations eventually overwhelm GPU memory. However, the Heavy Hitter Oracle (H2O) discovery has revealed a counter-intuitive efficiency: LLMs only need a fraction of their “memory” to maintain performance.
Researchers found that a small portion of tokens—Heavy Hitters (H2)—contribute the vast majority of value to attention scores. These tokens correlate with frequent co-occurrences in the text. By formulating KV Cache eviction as a dynamic submodular problem, the H2O framework retains only the most critical 20% of tokens. This results in up to a 29x improvement in throughput. This breakthrough democratizes AI, allowing massive models to run on smaller, cheaper hardware while retaining full contextual awareness.
5. The “Tool-Maker” Evolution: From Passive Solvers to Software Engineers
We are witnessing a fundamental shift from LLMs as “Tool Users” to LLMs as “Tool Makers” (LATM). Frameworks like LATM and CREATOR allow models to recognize when their inherent capabilities are insufficient—such as for complex symbolic logic—and respond by writing their own reusable Python functions.
This enables a cost-effective “division of labor.” An expensive, high-reasoning model (like GPT-4) acts as the Tool Maker, crafting a sophisticated utility function. A lightweight, cheaper model then acts as the Tool User, applying that function to thousands of requests. This allows models to solve problems they were never originally trained for by essentially creating their own specialized software on the fly.
6. The Semantic Shift: Moving Beyond the “Library Card Catalog”
Search technology is evolving from traditional Lexical Search to Semantic Search, fundamentally changing how information is retrieved.
Lexical Search acts like a literal “card catalog.” It relies on exact keyword matching. Searching for “affordable electric vehicles” might miss a document about a “Tesla Model 3” if those specific words are absent.
Semantic Search functions like a “knowledgeable librarian.” Using Dense Embeddings and Natural Language Processing (NLP), it maps queries into a vector space where similar concepts are mathematically grouped. It understands that “budget” and “affordable” are conceptually linked.
By leveraging Vector Databases (such as Milvus or Qdrant), modern systems now utilize a Hybridapproach. This combines the literal precision and speed of lexical search with the deep conceptual “brain” of semantic search, ensuring that intent is captured even when language is misaligned.
7. Conclusion: The Dawn of the “Interpretable” Era
The advancements moving through the AI frontier—from sparse steering and heavy-hitter optimization to autonomous tool-making—signal the end of the “black box” era. We are entering a phase where LLMs are becoming modular, efficient, and, most importantly, interpretable. By moving toward surgical control over internal representations, we move closer to systems we can truly understand and govern.
As we look forward, a vital question remains for the industry: Does the future of AI rely on building ever-larger models, or is the true path to intelligence found in making our control over them more modular and precise?
The challenge for businesses is to seek answers to questions, they do this with Metrics (KPI’s) and know the relationships of the data, organized by logical categories(dimensions) that make up the result or answer to the question. This is what constitutes the Information Value Chain
Navigation
Let’s assume that you have a business problem, a business question that needs answers and you need to know the details of the data related to the business question.
Information Value Chain
Information Value Chain
Business is based on Concepts.
People thinks in terms of Concepts.
Concepts come from Knowledge.
Knowledge comes from Information.
Information comes from Formulas.
Formulas determine Information relationships based on quantities.
Quantities come from Data.
Data physically exist.
In today’s fast-paced high-tech business world this basic navigation (drill thru) business concept is fundamental and seems to be overlooked, in the zeal to embrace modern technology
In our quest to embrace fresh technological capabilities, a business must realize you can only truly discover new insights when you can validate them against your business model or your businesses Information Value Chain, that is currently creating your information or results.
Today data needs to be deciphered into information in order to apply formulas to determine relationships and validate concepts, in real time.
We are inundated with technical innovations and concepts it’s important to note that business is driving these changes not necessarily technology
Business is constantly striving for a better insights, better information and increased automation as well as the lower cost while doing these things several of these were examined and John Thuma’s‘ latest article
Historically though these changes were few and far between however innovation in hardware storage(technology) as well as software and compute innovations have led to a rapid unveiling of newer concepts as well as new technologies
Demystifying the path forward.
In this article we’re going to review the basic principles of information governance required for a business measure their performance. As well as explore some of the connections to some of these new technological concepts for lowering cost
To a large degree I think we’re going to find that why we do things has not changed significantly it’s just how, we know have different ways to do them.
It’s important while embracing new technology to keep in mind that some of the basic concepts, ideas, goals on how to properly structure and run a business have not changed even though many more insights and much more information and data is now available.
My point is in the implementing these technological advances could be worthless to the business and maybe even destructive, unless they are associated with a actual set of Business Information Goals(Measurements KPI’s) and they are linked directly with understandable Business deliverables.
And moreover prior to even considering or engaging a data science or attempt data mining you should organize your datasets capturing the relationships and apply a “scoring” or “ranking” process and be able to relate them to your business information model or Information Value Chain, with the concept of quality applied real time.
The foundation for a business to navigate their Information Value Chain is an underlying Information Architecture. An Information Architecture typically, involves a model or concept of information that is used and applied to activities which require explicit details of complex information systems.
Subsequently a data management and databases are required, they form the foundation of your Information Value Chain, to bring this back to the Business Goal. Let’s take a quick look at the difference between relational database technology and graph technology as a part of emerging big data capabilities.
However, considering the timeframe for database technology evolution, has is introduced a cultural aspect of implementing new technology changes, basically resistance to change. Business that are running there current operations with technology and people form the 80s and 90s have a different perception of a solution then folks from the 2000s.
Therefore, in this case regarding a technical solution “perception is not reality”, awarement is. Business need to find ways to bridge the knowledge gap and increase awarement that simply embracing new technology will not fundamentally change the why a business is operates , however it will affect how.
Relational databases were introduced in 1970, and graph database technology was introduced in the mid to 2000
There are many topics included in the current Big Data concept to analyze, however the foundation is the Information Architecture, and the databases utilized to implement it.
There were some other advancements in database technology in between also however let’s focus on these two
History
1970
In a 1970s relational database, Based on mathematical Set theory, you could pre-define the relationship of tabular (tables) , implement them in a hardened structure, then query them by manually joining the tables thru physically naming attributes and gain much better insight than previous database technology however if you needed a new relationship it would require manual effort and then migration of old to new , In addition your answer it was only good as the hard coding query created
2020
In mid-2000’s the graph database was introduced , based on graph theory, that defines the relationships as tuples containing nodes and edges. Graphs represent things and relationships events describes connections between things, which makes it an ideal fit for a navigating relationship. Unlike conventional table-oriented databases, graph databases (for example Neo4J, Neptune) represent entities and relationships between them. New relationships can be discovered and added easily and without migration, basically much less manual effort.
Nodes and Edges
Graphs are made up of ‘nodes’ and ‘edges’. A node represents a ‘thing’ and an edge represents a connection between two ‘things’. The ‘thing’ in question might be a tangible object, such as an instance of an article, or a concept such as a subject area. A node can have properties (e.g. title, publication date). An edge can have a type, for example to indicate what kind of relationship the edge represents.
Takeaway.
The takeaway there are many spokes on the cultural wheel, in a business today, encompassing business acumen, technology acumen and information relationships and raw data knowledge and while they are all equally critical to success, the absolute critical step is that the logical business model defined as the Information Value Chain is maintained and enhanced.
It is a given that all business desire to lower cost and gain insight into information, it is imperative that a business maintain and improve their ability to provide accurate information that can be audited and traceable and navigate the Information Value Chain Data Science can only be achieved after a business fully understand their existing Information Architecture and strive to maintain it.
Note as I stated above an Information Architecture is not your Enterprise Architecture. Information architecture is the structural design of shared information environments; the art and science of organizing and labelling websites, intranets, online communities and software to support usability and findability; and an emerging community of practice focused on bringing principles of design, architecture and information science to the digital landscape. Typically, it involves a model or concept of information that is used and applied to activities which require explicit details of complex information systems.
In essence, a business needs a Rosetta stone in order translate past, current and future results.
In future articles we’re going to explore and dive into how these new technologies can be utilized and more importantly how they relate to all the technologies.
I was heavily involved in business intelligence, data warehousing and data governance as of several years ago and recently have had many chaotic personal challenges, upon returning to professional practice I have discovered things have not changed that much in 10 yearsagovernance The methodologies and approaches are still relatively consistent however the tools and techniques have changed and In my opinion not for the better, without focusing on specific tools I’ve observed that the core to data or MDM is enabling and providing a capability for classifying data into business categories or nomenclature.. and it has really not improved.
This basic traditional approach has not changed, in essence man AI model predicst a Metric and is wholly based on the integrity of its features or Dimensions.
Therefore I decided, to update some of the techniques and code patterns, I’ve used in the past regarding the information value chain and or record linkage , and we are going to make the results available with associated business and code examples initially with SQL Server and data bricks plus python
My good friend, Jordan Martz of DataMartz fame has greatly contrinuted to this old mans BigData enlightenment as well as Craig Campbell in updating some of the basic classification capabilities required and critical for data governance. If you would like a more detailed version of the source as well as the test data, please send me an email at iwhiteside@msn.com. Stay tuned for more update and soon we will add Neural Network capability for additional automation of “Governance Type” automated classification and confidence monitoring.
Before we focus on functionality let’s focus on methodology
Initially understand key metrics to be measured/KPI‘s their formulas and of course teh businesse’s expectation of their calculations
Immediately gather file sources and complete profiling as specified in my original article found here
Implementing the processes in my meta-data mart article would provide numerous statistics regarding integers or float field however there are some special considerations for text fields or smart codes
Before beginning classification you would employ similarity matching or fuzzy matching as described here
As I said I posted the code for this process on SQL Server Central 10 years ago here is s Python Version.
databricks-logo Roll You Own – Python Jaro_Winkler(Python)
databricks-logoroll You Own – Python Jaro_Winkler(Python) Import Notebook Step 1a – import pandas import pandas
Step 2 – Import Libraries
libraries from pyspark.sql.functions import input_file_name
from pyspark.sql.types import * import datetime, time, re, os, pandas
ML Libraires
from pyspark.ml.feature import RegexTokenizer, StopWordsRemover, NGram, HashingTF, IDF, Word2Vec, Normalizer, Imputer, VectorAssembler from pyspark.ml import Pipeline
import mlflow
from mlflow.tracking import MLFlowClient
from sklearn.cluster import KMeans import numpy as np Step 3 – Test JaroWinkler
while (iCounter <= len_str2): df=pandas.DataFrame([[iCounter,0]], columns=column_names) df_temp_table2=pandas.concat([df_temp_table2,df], ignore_index=True) iCounter=iCounter+1
iCounter=1 m=round((max_len/2)-1,0) i=1
while(i <= len_str1): a1=str1_in[i-1]
if m >= i:
f=1
z=i+m
else:
f=i-m
z=i+m
if z > max_len:
z=max_len
while (f <= z):
a2=str2_in[int(f-1)]
if(a2==a1 and df_temp_table2.loc[f-1].at['FStatus']==0):
common = common + 1
df_temp_table1.at[i-1,'FStatus']=1
df_temp_table2.at[f-1,'FStatus']=1
break
f=f+1
i=i+1
%sql DROP TABLE IF EXISTS NameAssociative; CREATE TABLE NameAssociative;
SELECT Name ,NameInput
,sha2(replace ( NameLookput,’%[^a-Z0-9, ]%’,’ ‘) , 256) as NameLookupCeaned ,a.NameLookupKey ,sha2(replace( NameInput,’%[^a-Z0-9, ]%’,’ ‘) , 256) as NameInput,b.NameInputKey ,JaroWinkler(a.NameLookup, b.NameInput) MatchScor ,RANK() OVER (Partition by a.DetailedBUMaster ORDER BY JaroWinkler(a.NameLookupCleande, b.NameInputCleaned) DESC) NameLookup,b.NameLookupKey FROM NameInput as a CROSS JOIN NameLookup as b
We want to build an Enterprise Analytical capability by integrating the concepts for building a Metadata Mart with the facilities for the Semantic Web
Metadata Mart Source
(Metadata Mart as is) Source Profiling(Column, Domain & Relationship)
+
(Metadata Mart Plus Vocabulary(Metadata Vocabulary)) Stored as Triples(subject-predicate-object) (SSIS Text Mining)
+
(Metadata Mart Plus)Create Metadata Vocabulary following RDFa applied to Metadata Mart Triple(SSIS Text Mining+ Fuzzy (SPARGL maybe))
+
Bridge to RDFa – JSON-LD via Schema.org
Master data Vocabulary with lineage (Metadata Vocabulary + Master Vocabulary) mapped to MetaContent Statements)) based on person.schema.org
Creates link to legacy data in data warehouse
+RDFa applied to web pages
+JSON-LD applied to
+ any Triples from any source
Semantic Self Service BI
Metadata Mart Source + Bridge to RDFa
I have spent some time in this for quite a while now and I believe there is a quite a bit of merit in approaching the collection of domain data and column profile data, in regards to the meta-data mart, and organize them in a triple’s fashion
The basis for JSON-LD and RDFa is the collection of data as a triple. Delving into said deeper
I believe with the proper mapping for the object reference and deriving of the appropriate predicates in the collection of the value we could gain some of the same benefits as well as bringing the web data being “collected, there by linking to source data.
Metadata (metacontent), or more correctly, the vocabularies used to assemble metadata (metacontent) statements, are typically structured according to a standardized concept using a well-defined metadata scheme, including: metadata standards and metadata models. Tools such as controlled vocabularies, taxonomies, thesauri, data dictionaries, and metadata registries can be used to apply further standardization to the metadata. Structural metadata commonality is also of paramount importance in data model development and in database design.
Metadata (metacontent) syntax refers to the rules created to structure the fields or elements of metadata (metacontent).[11] A single metadata scheme may be expressed in a number of different markup or programming languages, each of which requires a different syntax. For example, Dublin Core may be expressed in plain text, HTML, XML, and RDF.[12]
A common example of (guide) metacontent is the bibliographic classification, the subject, the Dewey Decimal class number. There is always an implied statement in any “classification” of some object. To classify an object as, for example, Dewey class number 514 (Topology) (i.e. books having the number 514 on their spine) the implied statement is: “<book><subject heading><514>. This is a subject-predicate-object triple, or more importantly, a class-attribute-value triple. The first two elements of the triple (class, attribute) are pieces of some structural metadata having a defined semantic. The third element is a value, preferably from some controlled vocabulary, some reference (master) data. The combination of the metadata and master data elements results in a statement which is a metacontent statement i.e. “metacontent = metadata + master data”. All these elements can be thought of as “vocabulary”. Both metadata and master data are vocabularies which can be assembled into metacontent statements. “
The MetadataMart serve as the source for both metadata vocabulary and MDM for the Master Data Vocabulary.
For the Master Data Vocabulary consider schema.org which defines most of the schemas we need. Consider the following schema.org Persons Properties of Objects and Predicates:
The key is to link source data in the Enterprise via a Business Vocabulary from MDM to the Source Data Metadata Vocabulary from a Metadata Mart to conform the triples collected internally and externally.
In essence information from the web applications can be integrated with the dimensional metadata mart, MDM Model and existing Data Warehouses providing lineage for selected raw data from web to Enterprise conformed Dimensions that have gone thru Data Quality processes.
With Data Profiling can apply the age old management adage “You get what you inspect, not what you expect” Readers Digest.
This article will describe how to implement a data profiling dashboard in Excel and a metadata repository as well as the TSQL required to load and update the repository.
Future articles will explore the data model as well as column and table relationship analysis using the Domain profiling results.
Data Profiling is essential to properly determine inconsistencies as well data transformation requirements for integration efforts.
It is also important to be able to communicate the general data quality for the datasets or tables you will be processing.
With the assistance over the years of a few friends(Joe Novella, Scott Morgan and Michael Capes) as well as the work of Stephan DeBlois, I have created a set of TSQL Scripts that will create a set of tables that will provide the statistics to present your clients with a Data Quality Scorecard comparable to existing vendor tools such as Informatica , Data Flux , Data Stage and the SSIS Data Profiling Task.
This article contains a Complete Data Profiling Kit – Download (Code, Excel Dashboards and Samples) providing capabilities similar to leading vendor tools, such as Informatica, Data Stage, Dataflux etc…
The primary difference is the repository is open and the code is also open and available and customizable. The profiling process has 4 steps as follows:
Create Table Statistics – Total count of records.
Create Column Statistics – Specific set of statistics for each column(i.e… minimum value, maximum value , distinct count, mode pattern, blank count, null count etc…)
Here is a sample of the Column Statistics Dashboard: There are three panels show one worksheet for a sample “Customers” file.
Complete Column Dashboard:
DatabaseName – Source Data Base Name
SchemaName – Source Schema Name
TableName – Source Table Name
ColumnName – Source Column Name within this Table
RecordCount – Table record count
DistinctDomainCount – The number of distinct values within entire table.
UniqueDomainRatio – The ration of unique records to total records.
MostPopularDomain – The most frequently occurring value for this column.
MostPopularDomainCount – The total for the most popular domain.
MostPopularDomainRatio – The ration for the most popular value to total number of values.
MinDomain – The lowest value with in this column.
First indicator for valid values violations.
MaxDomain – The highest value with in this column.
First indicator for valid values violations.
MinDomainLength – Length in characters for the minimum domain.
MaxDomainLength – Length in characters for the maximum domain.
NullDomainCount – The number of nulls within this column.
NullDomainRatio – The ration of nulls to total records.
BlankDomainCount – The number of blanks within this column.
BlankDomainRatio – The ration of blanks to total records.
DistinctPatternCount – The number of Distant values within entire table.
MostPopularPattern – The most frequently occurring pattern for this column.
MostPopularPatternCount – The total for the most popular domain.
MostPopularPatternRatio – The ration for the most popular pattern to total number of patterns.
InferredDataType – The data inferred from the values.
Complete Dashboard:
Column Profiling Dashboard 1-3:
Column Profiling Dashboard 2-3:
Column Profiling Dashboard 3-3:
Domain Analysis:
In the example below you see an Excel Worksheet that contains a pivot table allowing you to examine a columns patterns , in this cse Zip code, and subsequently drill into the actual values related to one of the patterns. Notice the Zip code example, we will review the pattern “9999”, or Zip code with only 4 numeric digits. When you click on the pattern og “9999” the actual value is revealed is
Domain Analysis for ZipCode
Domain Analysis for Phone1
Running the Profiling Scripts Manually
Perquisites:
The scripts support two databases. One is the MetadataMart for storing the profiling results, the other is the source for your profiling.
There are four scripts , simple run them in the following order:
0_Create Profilng Objects – Create all the Data Profiling Tables and Views
1_Load_TableStat – This script will load records into the TableStat profiling table
2_Load ColumnStat – This script will load records into the ColumnStat profiling table. Specify Database, Schema and Table name filters as needed. Example, to profile every table names starting with “Dim” then change the table filter to SET @TABLE_FILTER = ‘Dim%. Specify the Database where the Data Profiling tables reside DECLARE @MetadataDB VARCHAR(256) SET @MetadataDB = ‘ODS_METADATA’ Specify Database, Schema and Table name filters as needed. Example, to profile every table names starting with “Dim” then change the table filter to SET @TABLE_FILTER = ‘Dim%
SET @DATABASE_FILTER = ‘CustomerDB’
SET @SCHEMA_FILTER = ‘dbo’
SET @TABLE_FILTER = ‘%Customer%’
SET @COLUMN_FILTER = ‘%’
3_load DomainStat – This script will load records into the DomainStat profiling table. Specify the Database where the Data Profiling tables reside DECLARE @MetadataDB VARCHAR(256) SET @MetadataDB = ‘ODS_METADATA’ Specify Database, Schema and Table name filters as needed. Example, to profile every table names starting with “Dim” then change the table filter to SET @TABLE_FILTER = ‘Dim%
SET @DATABASE_FILTER = ‘CustomerDB’
SET @SCHEMA_FILTER = ‘dbo’
SET @TABLE_FILTER = ‘%Customer%’
SET @COLUMN_FILTER = ‘%’
-1_DataProfiling – Restart – Deletes and recreates all profiling tables
Please feel free to contact me with any questions?
