Capacity Planning for One Million Users

The direct recommendation: Do not size from registered users. Size each critical journey from peak requests, concurrent work, payloads, data growth, background jobs, recovery objectives, and an explicit service-level objective.

This article does not declare one provider universally best or invent a capacity or cost number. It gives you a reviewable method and requires measurement against the real workload. Primary sources were checked on 2026-08-28. Google Cloud comparisons are explicitly research-based, not presented as first-hand operating experience.

The decision in one minute

Separate registered, monthly active, daily active, and concurrently active users.
Model read, write, upload, search, report, and asynchronous-job paths independently.
Budget capacity for an availability-zone failure and deployment overlap, not only normal traffic.
Load-test the smallest realistic production topology before buying a larger one.

If the team cannot answer these points, the missing component is not another cloud service. The operating contract is incomplete. Write the assumptions, separate sourced facts from calculations, and identify what will be measured before purchase or release.

Reusable artifact

Area	Decision
Demand	DAU, sessions/user, actions/session, peak factor
Concurrency	Peak RPS x p95 service time
Data	Rows, objects, bytes/day, retention, versions
Reliability	SLO, RTO, RPO, zone failure headroom
Test	ramp, duration, cache state, pass/fail thresholds

Put this matrix in the architecture decision or implementation ticket. Any row without an owner or acceptance evidence is deferred risk.

How to use the artifact

1. Demand

The proposed decision is: DAU, sessions/user, actions/session, peak factor. Do not copy that value without context. Record its input source, environment, owner, and review date. Turn it into a repeatable check before and after release, including the wrong state the system must reject. Reopen the decision when data shape, region, traffic pattern, or ownership changes instead of assuming the previous setting remains safe.

2. Concurrency

The proposed decision is: Peak RPS x p95 service time. Do not copy that value without context. Record its input source, environment, owner, and review date. Turn it into a repeatable check before and after release, including the wrong state the system must reject. Reopen the decision when data shape, region, traffic pattern, or ownership changes instead of assuming the previous setting remains safe.

3. Data

The proposed decision is: Rows, objects, bytes/day, retention, versions. Do not copy that value without context. Record its input source, environment, owner, and review date. Turn it into a repeatable check before and after release, including the wrong state the system must reject. Reopen the decision when data shape, region, traffic pattern, or ownership changes instead of assuming the previous setting remains safe.

4. Reliability

The proposed decision is: SLO, RTO, RPO, zone failure headroom. Do not copy that value without context. Record its input source, environment, owner, and review date. Turn it into a repeatable check before and after release, including the wrong state the system must reject. Reopen the decision when data shape, region, traffic pattern, or ownership changes instead of assuming the previous setting remains safe.

5. Test

The proposed decision is: ramp, duration, cache state, pass/fail thresholds. Do not copy that value without context. Record its input source, environment, owner, and review date. Turn it into a repeatable check before and after release, including the wrong state the system must reject. Reopen the decision when data shape, region, traffic pattern, or ownership changes instead of assuming the previous setting remains safe.

Implementation sequence

Separate registered, monthly active, daily active, and concurrently active users. Start in a non-production environment and make step 1 produce a reviewable configuration, artifact, or metric.
Model read, write, upload, search, report, and asynchronous-job paths independently. Start in a non-production environment and make step 2 produce a reviewable configuration, artifact, or metric.
Budget capacity for an availability-zone failure and deployment overlap, not only normal traffic. Start in a non-production environment and make step 3 produce a reviewable configuration, artifact, or metric.
Load-test the smallest realistic production topology before buying a larger one. Start in a non-production environment and make step 4 produce a reviewable configuration, artifact, or metric.

Do not copy configuration between environments by hand. Version the policy, infrastructure, and release path where the service allows it, and keep production secrets out of the build. Another operator should be able to reproduce the path from the repository and runbook without hidden knowledge.

The DevOps and developer handoff contract

Handoff begins with an executable request: which commit or digest, which non-secret variables, which secret names and owners without copying values into the ticket, which health check, which migration, and what behavior is acceptable when a dependency fails? The developer supplies a local or test-environment run path, a safe configuration example, critical-journey tests, and the change’s data and backward-compatibility impact.

DevOps returns the environment address, permission boundary, log and dashboard locations, retention period, and deploy and rollback paths. Before approval, both sides name the release observer, decision window, and statistical threshold that aborts a canary or restores the previous artifact. After deployment, a message saying “done” is insufficient. Attach the deployed digest, timestamp, smoke result, public route, migration state, and any accepted deviation with a follow-up ticket. This contract prevents configuration failure from becoming an unowned gap between teams.

Failure modes to design tests around

Dividing daily requests by 86,400 and treating the average as the peak.

Do not approve this decision until an owner, environment, and pass threshold are named. Record the configuration, artifact digest, or log query that proves it, then add a deliberate failure test showing that the control rejects the wrong state. This turns advice into an operating contract shared by developers, DevOps, and security reviewers.

Scaling stateless compute while the database, queue consumer, connection pool, or third-party API saturates.

Calling a synthetic test a one-million-user proof without publishing inputs and failure thresholds.

Testing and release gates

Build a small matrix that links every risk to a test, owner, and evidence. Test the happy path, then test the expected denial or failure. Keep source success separate from build success, provider deployment separate from public-route response, and add data integrity and queue age to health checks for stateful systems.

Run deterministic checks in CI, environment and migration checks in CD, then verify the critical journey at the real deployed URL. Define an automatic abort threshold for error rate, p95 latency, or dependency failure. Retain the previous artifact and a data-recovery procedure. Success is readable evidence, not a green indicator by itself.

Cost without a marketing number

Model compute, storage, requests, transfer, logs, backups, keys, support, failure headroom, and engineering time. Use the actual region’s price on the decision date, then enable budget and anomaly alerts. A saving that removes failure capacity or extends the recovery objective needs explicit approval.

The safest optimization is deleting an unneeded resource or measurably reducing origin work. Reservations and long commitments come after the workload stabilizes. Give every line item a unit, review owner, and action when it crosses its threshold.

Field note from operations

The reusable lesson is that repository source, built artifact, provider state, browser response, and data integrity are different evidence states. Incidents often last longer because one green step is treated as proof that the whole chain succeeded. Capture timestamp, identity, route, and effective configuration before widening access or changing several layers at once.

This is an anonymized operating lesson, not a client outcome or benchmark. Its practical value is the procedure: change one reversible thing, watch the expected signal, and update the timeline.

Practical next steps

Copy the artifact above and fill it with system inputs and owners, not a sales forecast.
Select the two riskiest assumptions and run a failure and recovery test in non-production.
Record the decision, alternatives, cost model, acceptance evidence, and review date.
Continue through the DevOps and cloud series according to the next gap in your operating model.

If you need an evidence-led architecture or delivery review, start with a small scope through the web and application development service that names the risks and proof before a larger commitment.

Continue the series

Previous: The First 60 Minutes of a Critical Production Incident
Next: AWS or Google Cloud for Enterprise Apps? Choose by Workload, Not Logo
Related library article: A practical companion to this decision

Use these links as a learning path, but validate every decision against the actual system.

Sources and review date

Last reviewed: 2026-08-28. Recheck pricing, quotas, and regional feature availability immediately before implementation.