Prepare an architecture consultation around a real decision
Prepare the decision, one representative workload, the relevant evidence, and the people who can act on the result. Separate observed behavior from missing measurements, provide proportionate access, and ask for a decision record with a testable next step.
Inside this article 6 sections +
Prepare for an architecture consultation by defining the decision you need, selecting one representative workload, and assembling enough evidence to explain its behavior. A complete documentation program is unnecessary before the first conversation. A vague request to “review everything” is equally unhelpful because it hides the choice the consultation should support.
The preparation should make productive questions possible. For example: should a slow customer export move to a separate worker, does the database need a different access pattern, or is the real delay caused by a partner API? Each possibility implies a different intervention and a different way to judge success.
Write the decision in one paragraph
Describe the customer or business problem, the observed symptom, the options currently being considered, and the constraint that makes the decision important. Include the consequence of waiting. A team facing a recurring timeout before a customer launch needs a different consultation from one exploring a long-term platform simplification.
A useful brief might say: “Account exports above a certain size miss the product's completion objective. We are considering separate workers or a database redesign, but have not isolated queue wait from query time. We need a reversible next step before expanding the export audience.” This is an illustrative brief, not a claim about a client system.
Identify who can make the resulting decision. A consultant can explain tradeoffs, but the product owner must confirm acceptable behavior and the operating team must understand the responsibility they will inherit. Invite the people needed for that decision rather than assembling a large meeting solely by seniority.
Assemble a small evidence packet
The request path can be a simple diagram with ownership and data movement. Show which system authorizes the request, which record represents accepted work, and which component declares completion. Include queues and scheduled work; a synchronous HTTP diagram can omit most of the time the customer waits.
For behavior evidence, include ordinary and problematic examples. A single worst-case trace may be important, but it does not establish the normal workload distribution. Record the observation interval, relevant sample count, and whether failures or timeouts are excluded from the available metrics.
Acknowledge missing evidence explicitly. If queue wait is not instrumented, say so. If the reported latency comes only from successful requests, identify that boundary. Unknowns are useful inputs to a consultation because they can justify a small measurement task before a larger architectural commitment.
Provide one trace that can be explained
Choose a synthetic or appropriately sanitized operation with a durable identifier. The example below is an original evidence-envelope shape. Its times are invented to illustrate separation of stages, not measured performance.
{
"operation_id": "export-demo-42",
"workload": {"rows": 12000, "format": "csv"},
"durations_ms": {
"acceptance": 80,
"queue_wait": 7200,
"database_read": 450,
"format_and_upload": 900
},
"outcome": "completed",
"unknowns": ["connection acquisition measured separately"]
}
For this invented operation, the listed stages total 8,630 milliseconds and queue wait dominates. That does not prove the database is healthy or that more workers are safe. It directs the next questions toward worker availability, admission policy, competing jobs, and downstream capacity before assuming a database rewrite is necessary.
Preserve the distinction between an illustration and an observation when sharing evidence. Actual traces should identify the environment and configuration, while examples used to explain a desired measurement should be labeled as examples. Mixing them can make a recommendation appear empirically supported when it is still a hypothesis.
Make access proportional to the question
Begin with architecture context, sanitized payloads, selected metrics, and relevant code excerpts. If deeper inspection is necessary, provide scoped access appropriate to the agreed question. A first consultation rarely needs unrestricted production administration or a full customer database export.
Prepare a reproducible local or staging path when possible. A fixture that triggers the important behavior is more useful than asking an external reviewer to infer everything from a screenshot. Document which dependencies are real, stubbed, or omitted so conclusions do not extend beyond the environment being inspected.
Arrange access before the session and verify it with a bounded check. Time spent discovering that a dashboard requires another administrator's approval does not improve the architectural decision. Keep access ownership and expiry visible so the temporary review does not become an unmanaged permanent account.
Choose the depth of consultation the decision needs
A focused consultation works well when the team can explain the system and needs help resolving one tradeoff. Bring a current diagram, one representative trace, the alternatives under consideration, and the business constraint that makes the choice difficult. Ask for a decision record with conditions and follow-up actions. The output should clarify what can be decided from the available evidence and what cannot.
A codebase review is more appropriate when the explanation itself is uncertain or competing teams describe different causes. It needs access and time to trace the relevant behavior. A performance or integration experiment is appropriate when a measurable unknown separates the options. Agree whether the engagement includes running that experiment or only designing it; those are materially different deliverables and should not be hidden inside a vague consulting scope.
For an initial discussion with Zero Gravity Solutions, a sanitized decision brief is enough to establish the next useful step. Include the current constraint, desired outcome, existing ownership, and any deadline that changes the options. Do not send credentials in the brief. We can then agree on the necessary access, the review boundary, and the evidence your team should receive before further implementation is commissioned.
Agree on an output that supports the next action
Ask for a decision record containing the observed constraints, plausible explanations, options, tradeoffs, recommended next step, and remaining uncertainty. A broad technology recommendation without a connection to the evidence is difficult to evaluate and even harder to implement responsibly.
The next step may be a measurement change, a limited experiment, a configuration adjustment, or a scoped implementation. Define what would confirm or reject its hypothesis. For the export example, a useful experiment might bound worker concurrency while measuring queue age and database occupancy under a representative workload.
Keep decisions and tasks distinct. “Investigate database limits” is a task; “retain the existing data model until a controlled test shows it prevents the required completion behavior” is a decision with a review condition. Both may belong in the output, but they serve different purposes.
The architecture review guide covers the broader assessment, and asynchronous tracing helps build useful execution evidence. Our technical consulting service focuses the conversation on decisions your team can assess and act on.