51 posts total
Open table formats like Iceberg and Delta turn the file layout into the API. Here's where that pays off, where it just adds operational surface, and how to tell which side you're on.
Four org models for data teams — centralized, embedded, hybrid, and platform-team-of-experts — with the failure modes of each and a maturity trajectory by company stage.
A practical privacy posture for mid-market analytics: minimize at ingest, mask at the warehouse, enforce role-based access, keep an honest audit trail. No over-engineering, no compliance theater.
Dashboard sprawl is rarely killed by its first-derivative cost. The damage that matters is second-derivative: time-to-decision degrades, executive trust erodes, and the analytics function loses its seat at the table. Here is how to inventory the surface area and rationalize it.
A working analyst's read on where LLMs actually move the needle in 2026 — SQL drafting, code review, narrative writing, exploration scaffolds — and the four areas where they still don't.
A 30/60/90-day plan for a data quality program that leads with measurement, banks visible wins before remediation spend, and earns budget by showing dollars rather than asking for them.
Data contracts only look like a schema problem. Their real value is the conversation they force at sprint planning — before a breaking change ships, not after a dashboard goes red.
Pipeline uptime is the wrong metric. The signals that actually predict bad decisions are freshness, completeness, accuracy, schema drift, and lineage — tiered to business criticality.
Feature stores solve real engineering problems — online/offline parity and point-in-time correctness — but most mid-market data teams don't have those problems yet. A decision framework, plus the cheaper alternatives that work for 80% of use cases.
A practical comparison of dbt and SQLMesh across materialization semantics, virtual data environments, contract enforcement, ecosystem maturity, and hiring market — with criteria-driven recommendations by team size.
Most executive dashboards are dead by week six. The cause is rarely the chart library or data model — it is that the dashboard was never wired into a decision, an owner, or a recurring meeting. What separates dashboards that survive from those that quietly stop loading.
Most warehouse overspend is not a pricing problem. It is a query-shape problem, and a small number of patterns generate the overwhelming majority of the bill. Here is how to find them, fix them, and do it inside six weeks of disciplined attention.
Hub-and-spoke or federated? The right answer depends on which failures you're seeing today. Here are the signals that tell you which way to move.
Six recurring dbt anti-patterns we untangle on client engagements: monolithic marts, misleading reusable macros, ambient tests, over-incremental models, ref-graph spaghetti, and inconsistent surrogate keys.
Source models, metric definitions, and consumption views are three different jobs. Collapse them and your warehouse turns into a graveyard of contradictory dashboards.
Centralized metric definitions solve a specific class of problem at a specific scale. Below that scale, a metric store is overhead pretending to be governance.
An honest read on enterprise AI agents in 2026 — the narrow vertical workflows where they earn their keep, and the open-ended, cross-system, messy-state problems where they still fall apart.
Three pattern families — routing, evaluation, and human-in-loop — that separate agentic analytics workflows that survive production from demos that quietly die in staging.
Four categories of analytics return — revenue lift, cost reduction, risk mitigation, decision speed — sized in dollars at $100M+ revenue. No vanity metrics, no aspirational case studies.
A criteria-driven framework for choosing a BI vendor in 2026 — semantic layer fit, embedded analytics, governance, three-year total cost, hiring market, and the substance behind the AI features.
Shadow data — the spreadsheets, Notion tables, and Airtable bases operating teams actually run on — is usually framed as a governance failure. It is not. It signals where the central data function has under-invested. The response is graduated absorption: formalize, leave alone, or retire.
Four questions to ask before any data catalog demo — what ingestion actually breaks on, how lineage is computed under the hood, real PII auto-tagging precision, and the user-onboarding cost vendors quietly externalize.
A pragmatic incident response model for data teams: severity tiers, communication templates, and postmortem rules that survive contact with reality — without cosplaying as an SRE org.
A decision framework for mid-market data leaders weighing Snowflake against Databricks across workload fit, governance, ML maturity, cost, hiring, and lock-in.
A practical playbook for the first-time data leader: how to map stakeholders, set communication cadence, push back without burning capital, and use narrative as a strategic tool.
Most data teams measure ticket throughput and pipeline uptime and call it a day. Those numbers describe the plumbing, not the value. Here are five operational KPIs that actually predict whether your data function is earning its budget — with concrete measurement notes for each.
When chargeback is worth introducing for a central data function, the failure modes that wreck most rollouts, and a set of allocation rules that survive contact with a real finance org.
What's converged, what's still a mess, and what's hype in the modern data stack circa 2026 — with a working theory of why the same three problems keep eating client budgets.
Pragmatic mid-ground patterns for analytics modeling at $100M+ data volumes — where Kimball still wins, where activity schema earns its keep, and how to combine them without breaking your warehouse bill.
How to migrate from Teradata or Oracle to Snowflake, BigQuery, or Databricks without breaking trust — discovery, dual-run, and the criteria that earn the right to decommission the old system.
A 5-tier maturity model for data governance at $100M+ organizations — from ad-hoc to optimized — with the observable signals and the specific investments that move you up a tier.
Four stage gates — discovery, scope, build, ship — with explicit exit criteria that stop data projects from drifting into open-ended consulting engagements or shelfware dashboards.
Most teams treating the open table format decision as load-bearing are choosing between things that won't differentiate their workload. Here's where the choice actually matters, and where it's decorative.
A working comparison of Cube, Lightdash, dbt Semantic Layer, and Looker as semantic layers in 2026 — what each is actually optimized for, when it earns its keep, and when it's overkill for the problem you have.
Most dashboards die from scope drift, not bad SQL. A scope contract — purpose, decisions, owner, review cadence, retirement criteria — turns a dashboard into a managed asset instead of a slowly accumulating liability.
Text-to-SQL agents promise natural-language access to the warehouse. The architectural realities — schema breadth, semantic context, verification, and cost — explain why production deployments narrow fast.
Seven questions that separate data consultants who deliver from those who bill — covering team composition, knowledge transfer, exit criteria, and the economics behind their staffing model.
Late-arriving facts are not an exception. They are the default. Modeling them as edge cases produces silent data drift; modeling them as the norm produces a warehouse you can trust.
A pragmatic migration plan for moving an organization off ad-hoc SQL and copy-pasted report logic onto a governed semantic layer, including how to retire the legacy reports without political damage.
Most data team rituals are inherited from a previous org chart and quietly stop working when the org chart changes. The ones that survive a reorg share a common shape: they are tied to outcomes the business already cares about, not to the team's internal cadence.
Most data teams justify their budget with output metrics — dashboards shipped, models deployed, queries served. None of those tell you whether the team is worth what it costs. Cost per decision does.
A concrete week-by-week onboarding plan for new data hires that ships measurable output by day 30, owns a domain by day 60, and drives a strategic project by day 90 — without the usual six-month ramp tax.
CDC is the default for non-trivial data pipelines, and that is often the wrong default. The right choice depends on latency needs, source system tolerance, and the engineering capacity to operate a streaming substrate — not on which pattern sounds more modern.
You don't need a metrics platform to get a semantic layer. You need three folders, a CI check, and the discipline to stop writing business logic in dashboards.
Vendor pricing anchors analytics TCO on compute and seats. The costs that actually compound — egress, model rebuilds, reverse ETL, observability, and the human time to keep it coherent — sit outside the quote and decide whether the platform earns its place.
A data strategy describes the destination. A roadmap describes the route. Most data orgs have both and still flounder because the third doc — the operating contract — does not exist.
Self-serve BI fails because executives buy it as a tool. The orgs that make it work treat it as a redistribution of analytical labor with explicit rules.
Treating data as a product means contracts, SLAs, and a named owner accountable for outcomes — not a Confluence page calling a table a product.
Pie charts get a bad rap, but the real failure mode is asking the consumer to guess at proportions. Visualization is the last mile of the entire data investment.
Power BI empowers teams to make data-driven decisions in meetings instead of after. Here's how it's changing the way decisions actually get made.
24 topics
Stay in touch
We don't run a newsletter yet — but email us and we'll add you to the short list that gets new posts when they go up.