updown move enter open esc close matheesan.ca / index

    CASE 04 · GCP

    The prototype that lost

    A cheaper Cloud Run replacement for a high-volume iPaaS — and why I recommended against it.

    p95 401 ms37–45 ms warm~99% lower modelled cost
    GCPCloud RunBunbenchmarkingtechnical judgment

    problem

    A high-volume integration platform was a cost and control risk. The team needed to know whether a per-tenant Cloud Run service could replace it.

    approach

    Ported a six-flow production integration to per-tenant Cloud Run services: a Bun-compiled 107 MB container, preview deploys in about 7 seconds, metadata-only rollback, and verified cross-tenant isolation. Benchmarked cold start at p50/p95/p99 of 291/401/453 ms and warm requests at 37–45 ms.

    the hard part

    The target was 300 ms p95. The prototype missed it at 401 ms. Keeping an always-warm pool would have cost roughly 5.5× the alternative, which erased the headline savings. The honest answer was that the cheaper platform was the better platform for this workload.

    result

    Modelled hosting cost was about 99% lower, but I reported the trade-off where the prototype lost. That informed the organisation’s choice of the other platform. I also pushed PII redaction, alerting and PCI shared responsibility into the platform requirements.

    what I’d do differently

    Nothing. The point was to find out before migrating a high-volume platform — and the prototype did its job by losing on evidence.