How Uber tests payments in production

Overall reaction to the article

  • Many commenters find the piece “fluffy,” overlong, and light on concrete technical detail.
  • Others note that the core message is very simple: staging can’t catch all bugs; you must be prepared to detect and fix issues in production.
  • A few see value in the anecdotes and framing, especially for less-experienced engineers, but some replies treat that praise as satire.

Staging vs. production for payments

  • Broad agreement that some failures only surface in real production conditions: real banks, networks, device flows, fraud systems, edge cases.
  • Several engineers say they always test in staging first, then do at least one live payment in prod after each deployment.
  • Others argue good sandboxes (e.g. some modern providers) can be sufficient for most cases, with only a minimal “sanity check” in prod.

Quality of payment provider test environments

  • Many report test environments are unreliable or non-representative: different validation rules, broken features, stale data, or missing “special” account settings.
  • This has led some teams to largely abandon test endpoints and rely on real endpoints plus corporate cards or virtual test cards.
  • A minority report the opposite experience: for them, provider sandboxes (especially some well-known ones) closely match production and work well.

Using real cards and compliance issues

  • Common practice described: live “smoke tests” using corporate cards or prepaid/virtual cards, sometimes fully exercising refund/cancel flows.
  • Concerns raised:
    • Card network terms and payment provider agreements may forbid self-pay or repeated test transactions in live mode.
    • PCI/PA-DSS explicitly say real card data must not be used in non-production environments.
    • Some claim specific providers will terminate accounts for self-pay tests; others say they’ve done it for years without issue.
  • Legal/HR angle: pressuring employees to use personal cards is criticized; in some jurisdictions it may be illegal if not reimbursed and optional.

Rollout and “testing in production” strategies

  • Discussion of canary releases, region-by-region rollouts, and parallel “shadow” systems that replay real traffic and compare outputs.
  • Some note that a slow rollout isn’t always viable (e.g., urgent security fixes), but generally staged rollouts are considered best practice.