At 10:07 every Monday, the executive report slows down. A pipeline is finishing, a notebook is still running and the finance team refreshes several large reports at once. By 10:30 the problem is gone.
A consultant can build a handsome Fabric demo without ever meeting that Monday. The consulting test is to recreate the contested half hour and explain every delay, capacity charge and recovery choice.
Preserve the Monday timeline
Choose one path from source to business use. It should be important enough to justify the exercise and bounded enough to finish.
Write the possible verdicts before the replay starts:
| Verdict | Meaning |
|---|---|
| Adopt | Fabric fits the workload and operating model; proceed to a defined next slice |
| Limit | Fabric fits part of the path; keep named components elsewhere |
| Defer | The platform may fit, but a prerequisite or commercial fact is unresolved |
| Reject | Evidence shows another path is better for this workload |
This prevents the replay from becoming a sales ritual. A consultant who cannot recommend “keep it where it is” is not evaluating fit.
The workload contract should include source volume and shape, arrival pattern, transformations, consumer, freshness target, concurrency, access roles, reconciliation, recovery expectation and current run cost. Mark estimates as estimates. Add a clock row for every operation active between 9:45 and 10:30.
Mark the collision on the current clock
Observe the current workload through an ordinary cycle and a busy cycle. Capture timestamps and identifiers from source arrival through the consuming report, API or model.
Record four kinds of evidence:
- Service: completion time, freshness, query response and failure rate.
- Quality: reconciliation, late records, duplicates, schema changes and correction flow.
- Operation: alerts, manual steps, support tickets, recovery and ownership.
- Economics: platform charges, licenses, engineering time and support burden tied to the path.
The timeline is not an argument for keeping the current stack. It is the standard the new path must beat or deliberately trade against. A slower result may still win if it removes an expensive operating failure. A faster query may lose if no one can explain the bill.
Assign each clock event to one Fabric responsibility
Microsoft describes Fabric as a SaaS analytics platform with workloads for ingestion, transformation, engineering, warehousing, real-time processing, data science and reporting over OneLake. That breadth is useful. It also makes over-design easy.
Create a component decision table:
| Need | Candidate Fabric component | Evidence for using it | Cost of adding it |
|---|---|---|---|
| Bring source data in | Pipeline, dataflow, shortcut, mirroring or another supported path | Source behavior and freshness case | New runtime, copy, permission and support surface |
| Transform data | SQL, notebook, Spark job, pipeline step or existing external engine | Transformation type, skill and scale | Compute, release path and debugging burden |
| Store and serve | Lakehouse, warehouse, event store or retained current system | Query and consumer pattern | Migration, duplication and reconciliation |
| Publish measures | Semantic model and Power BI report | Named decisions and access cases | Model ownership and refresh load |
OneLake is not a reason to move every source. A shortcut, copy, mirror or retained external system carries different consistency and failure behavior. The consultant should record which behavior the workload requires and what happens when the source or link is unavailable.
Recreate the same half hour in Fabric
Use the intended tenant, identity pattern, workspace separation and deployment route. A personal sandbox proves syntax. It does not prove operation.
The replay environment should include:
- one representative source with ordinary and awkward records
- the real transformation logic or a bounded version with the same complexity
- a governed data object and named owner
- the semantic or serving layer used by the consumer
- production-shaped volume and concurrency
- monitoring, reconciliation and a recovery procedure
Keep a decision log as the replay takes shape. Why did the team choose a lakehouse rather than a warehouse for this object? Why is a notebook acceptable in the release path? Which team owns a schema change? A future maintainer needs the trade, not only the configuration.
Read capacity as a debt ledger
Fabric capacity is shared compute. Microsoft documents the Fabric Capacity Metrics app as the place to monitor capacity consumption and inspect behavior such as throttling and query rejection. Use telemetry from the audition rather than a one-time calculator result.
Keep the replay pack with timestamps:
| Case | What to observe |
|---|---|
| Normal cycle | Completion, consumption, idle behavior and consumer response |
| Monday peak | Concurrent pipeline, notebook, model refresh and report demand |
| Backfill | Sustained processing and effect on interactive users |
| Bad transformation | Failure isolation, retry, duplicate prevention and cost |
| Late source | Freshness signal, consumer message and recovery |
Run the pack more than once. A single clean run can hide warming, caching or a workload that happened not to overlap.
Translate results into options: reschedule, optimize, isolate workloads, change capacity or leave a component outside Fabric. Do not quote a buyer-specific cost without the buyer’s current region, agreement, capacity choice and measured use.
Put each role on the same timeline
Create users that represent a platform administrator, workspace owner, developer, analyst and restricted consumer. Use the client’s intended identity groups, not one consultant account.
Test who can create workspaces, connect sources, view sensitive detail, publish items, share reports, change permissions and access underlying data through alternate tools. Deny a request and verify both the user experience and the audit evidence.
Microsoft’s Fabric administration guidance separates tenant, governance and operating responsibilities across the data estate. The engagement should translate those responsibilities into named client roles. “IT owns Fabric” is not an operating model.
For every governed object, record the data owner, technical owner, access approver, quality owner, support queue and deletion or retention rule. Platform labels can support these decisions. They cannot make them.
Disturb the clock and watch the consumer
Stop the source midway. Change a schema. Replay a file. Remove a permission. Fill the shared capacity with the peak pack. Publish a transformation that produces a wrong control total.
Observe five things: detection, containment, consumer signal, recovery and evidence. A retry that duplicates records is not recovery. A red pipeline with no on-call route is not detection in practice.
The acceptance record should show the failing input, run identifier, alert, action, corrected output and owner. Preserve enough detail for the client’s team to repeat the drill after the consultant leaves.
Move one change through the real release path
Move the thin slice from development through the same gates planned for production. Separate configuration and secrets from code. Review permissions at each stage. Confirm that a rollback or forward fix has an owner.
Do not let notebooks, SQL, pipelines and semantic models follow unrelated release customs without a reason. The artifacts can use different deployment mechanisms, but they need one change record and a test that proves the end-to-end workload still reconciles.
The first release is part of the replay. If the consultant must perform hidden portal work, record it and decide whether to automate, document or remove it.
Price every duplicate Monday during coexistence
Most migrations run two paths for a while. Include dual operation, reconciliation, user transition, historic reload, contract overlap and retirement work in the estimate.
Build the wave plan from dependencies rather than dashboard popularity. A report may look easy while relying on twelve upstream jobs and an undocumented spreadsheet. Another may move cleanly because its source and measure ownership are settled.
Classify each workload as move, redesign, retain or retire. Give retained workloads a reason and a recheck trigger. Fabric does not need to own the whole estate to be a useful choice.
Hand the Monday run to its future operator
Before the verdict, have the client’s operator run the Monday peak pack, diagnose a failed load, read the capacity view, correct a permission issue and release a small change. The consultant observes.
The handoff should include the component decisions, workload contract, annotated Monday timeline, capacity evidence, access model, release path, recovery records, known limits, cost assumptions and next review date.
Our data architecture consulting guide starts with a disputed number. A Fabric engagement can use the same number as its workload spine, then prove whether the platform improves the path without losing meaning or ownership.
The replay ends with a verdict and a bounded next move. If the only output is a demo workspace and a migration slide, the Monday problem is still waiting.