Is Your Reporting Querying the Wrong Layer of Your Data Platform?

A data platform is built in layers. Raw data lands, gets cleaned and joined, and ends up in a form the business can read. Each layer is designed for a different job, with different engines, different performance characteristics, and very different costs per query.

Reporting is designed for one of those layers. In practice it often ends up querying another: the engineering tables, an acceleration engine, or a warehouse sized for transformation jobs. Nothing breaks. The reports still work. They are just slower than they should be, the bill is higher than it should be, and nobody can say exactly why.

This piece explains how that happens, how to tell whether it is happening to you, and what a serving layer built for reporting looks like on the common platforms.

How the layers are meant to work

Most modern platforms follow a version of the same pattern. Databricks popularized the names bronze, silver, and gold:

Alongside these, many estates add a query acceleration layer: materialized views, aggregation tables, or an engine such as Dremio with its reflections. These exist to make specific queries fast. They are tuned for the workloads they were built for.

In TOGAF terms, reporting is a business capability (management reporting, operational dashboards) with its own information requirements: how fresh, how detailed, how many readers, how often. The serving layer is the architecture building block that meets those requirements. When reports skip it, the requirements are being met by building blocks designed for something else.

How reporting ends up in the wrong place

It is rarely a decision. It is usually a shortcut that stayed:

Why it costs more

The mechanics differ by platform, but each one has a way of turning misplaced reporting into cost.

Power BI connection modes. In Import mode, data is loaded into the semantic model on a schedule, and report interactions read the model. Microsoft allows up to 8 scheduled refreshes a day on Pro and up to 48 on Premium or Fabric capacity. In DirectQuery mode, report interactions send queries to the source, so every filter click and page load becomes work on the source system. DirectQuery is the right choice for some cases. Chosen by default, it moves reporting load onto whatever engine sits underneath.

Fabric capacity. Everything in a Fabric capacity shares the same pool of capacity units. Fabric smooths usage over time, but when a capacity is overloaded it throttles in stages: new interactive operations are delayed by 20 seconds, then interactive requests are rejected, and eventually background work such as refreshes and pipelines is rejected too. A reporting pattern that consumes far more than it needs can slow the data engineering jobs on the same capacity, and the reverse. Direct Lake, Fabric’s mode for reading Delta tables directly, can also fall back to DirectQuery when a query exceeds the capacity’s guardrails or uses an unsupported feature, which can quietly change performance.

Snowflake warehouses. A warehouse is billed per second while running, with a 60-second minimum each time it resumes. Dashboards that fire small queries throughout the day can keep waking a warehouse, paying the minimum each time, or keep it running continuously. Snowflake caches query results for 24 hours, so identical queries against unchanged data are cheap. Queries generated by many slightly different filters often are not identical.

Databricks SQL. Serverless SQL warehouses bill DBUs while they run. A reporting workload pointed at a warehouse sized for engineering jobs pays engineering rates for dashboard queries.

Acceleration engines. Reflections and materialized views are maintained for the queries they were designed for. When reporting tools query them directly and broadly, they get rebuilt and expanded to keep up, and the acceleration layer starts to carry the reporting workload on its own.

How to tell whether it is happening

You can find out from the platform’s own records:

  1. Find where reporting queries land. Use Snowflake’s query history, Databricks query history in system tables, or the Fabric Capacity Metrics app. Group queries by the tool or service account that sent them. If your BI tool’s account appears against engineering warehouses or raw tables, reporting is in the wrong place.
  2. Compare refreshes to reads. Count how often each dataset refreshes and how often its reports are opened. A dataset refreshed 48 times a day and read twice is paying for 46 refreshes nobody uses.
  3. Check connection modes. List which semantic models use DirectQuery and why. Many will have no reason beyond the original default.
  4. Look for duplicate logic. If the same measure is defined in several reports, each one is computing it separately.
  5. Watch the shape of the load. On Fabric, look for interactive spikes at the start of the business day. On Snowflake and Databricks, look for warehouses that never suspend during working hours.

What a serving layer for reporting looks like

The fix is to give reporting a layer of its own, shaped to its requirements:

None of this is exotic. It is the serving layer most architectures intended to build, finished and enforced.

Keeping it that way

Reporting drifts back to the wrong layer the same way it got there: one convenient connection at a time. Treat it as architecture governance. Record where reporting may connect, review new reports and semantic models against that rule, and use TOGAF’s Architecture Change Management (Phase H) to revisit the serving layer when the business asks for something it cannot answer.

Where we come in

Misplaced reporting is an ownership problem as much as a design problem. It accumulates when nobody owns the serving layer after launch.

We take on that ownership: after-build ownership of your pipelines, serving layer, and the reporting that depends on them, with monitoring, cost control, and documentation. We run change management so new reports and models are reviewed against where they should connect. And when reports slow down or a bill jumps and nobody knows why, we step in urgently, find the cause, and leave it supportable.

If your reports are slower or more expensive than they should be, tell us about it.


Tell us what you're running.

Two quick steps. I'll read it personally and reply with an honest read on your situation and whether our team can help, usually within one business day.

  • I lead every engagement personally
  • No sales sequences, and your details are only used to reply

Prefer email? [email protected]

Step 1 of 2

Which sounds most like you?
What's prompting this? Pick any that apply
How many business-critical custom systems?
Timeline
Where should I reply?