Case 03 Shipped

Retailer launches as plan‑first workflows

Onboarding changes for three retailers across 24 service repositories, an access module that finds missing repository access before a launch step runs, and a tenant migration that cut its running cost by about 80%.

Context
Multi-tenant e-commerce platform, 2026
My part
I led the DevOps team, authored the onboarding changes for three retailers and extended a teammate-built provisioning platform with an access-governance module and a deploy-wave workflow step.
Team
I led the DevOps team. A teammate built the core provisioning platform, and my modules plug into it.
Status
Delivered to production

Stack: Python · React · TypeScript · REST APIs · GitLab · Slack · pytest · Terraform · BigQuery

  • 3 retailers onboarded, all three in production Shipped
  • 79 repositories checked for access before execution Shipped
  • ~80% lower running cost for one migrated tenant Measured
On this page

Context

A multi-tenant e-commerce platform serves many retailers from one codebase. Onboarding three retailers touched configuration and application code in 24 service repositories: single sign-on, retailer and banner settings, fulfilment tooling, carrier and charges configuration, service configuration delivered through S3, BigQuery and Terraform resources, themes, and CI deployment paths.

Problem

A launch followed a checklist of 205 items. Work stalled when someone found, halfway through, that they lacked access to a repository. Manual steps were easy to repeat or to skip.

Constraints

  • The automation lives in a shared Python and React platform that a teammate built.
  • Every step must be safe to run twice.
  • Nothing changes before a reviewer has seen the plan.
  • Changes to tenant CI go through the normal merge-request flow.

Options

The team could keep the checklist as a document and script the worst steps, or model the checklist itself as workflow steps with a plan and an audit trail. The platform already took the second path; my work added the access audit and generated CI to it.

Engineering decisions

  1. Plan first. Every workflow shows what it will change before it changes anything.
  2. Audit access before execution, and route each gap to the owner of the repository.
  3. Idempotent steps, so a rerun converges and never duplicates work.
  4. Generated CI. Tenant deploy-wave jobs come from data; the step validates the YAML and opens a merge request through the platform’s Git flow.
  1. Plan
    • Launch plan checklist items as workflow steps, nothing runs yet
  2. Audit
    • Repository access audit 79 repositories checked
  3. Route
    • Owners notified through Slack
    • Scheduled CI report open gaps stay visible
  4. Execute
    • Idempotent steps safe to run twice
    • Deploy-wave jobs generated, YAML validated
  5. Review
    • Merge request reviewed before merge
A retailer launch as a plan-first workflow. Access gaps reach their owners before any step changes a repository.Abstracted view. Names and counts are illustrative, not the real topology.

Implementation

The repository-access governance module is about 3,400 lines across Python, React and TypeScript, REST APIs, GitLab, Slack and scheduled CI reporting. Its audit checked 79 repositories and grouped the gaps into 13 owner routes before a launch began. It is careful by default: messages start as a dry run, posts have a cap, repeats wait 24 hours, the scheduled job only reports, and owners still approve every grant.

The deploy-wave step generates a tenant’s CI jobs, validates the YAML, then opens a merge request through the platform’s Git flow. It edits only the target YAML blocks, so merge requests stay small, and tests prove that a rerun changes nothing.

In the operator interface of the 166-step workflow, I added filters by actor type and a metadata search.

I planned and ran a tenant migration between platform environments that cut that tenant’s running cost by about 80%.

When a tenant moved between platform environments, I replicated its feature-flag state, 64 flags across four banners, and read every flag back to confirm parity.

Validation

  • pytest on the deploy-wave step, including a rerun that must change nothing.
  • YAML validation before any merge request opens.
  • Readback of every migrated feature flag, comparing source and target.

Outcome

  • ShippedOnboarding changes for three retailers, through pre-production to production; all three reached production.
  • ShippedAn access audit across 79 repositories that runs before execution.
  • MeasuredA tenant migration that cut that tenant’s running cost by about 80%.
  • ShippedA tenant migrated between platform environments with its feature flags read back as matching.
  • ShippedA retailer onboarded into an SLA analytics service (BigQuery configuration for development, integration and production), a 55-file retailer web theme with visual-regression baselines, and rewritten SSO and theming runbooks.