Insights · Platform Engineering
Platform engineering: run the platform like a product
A platform is an internal product built for named users; its success metric is adoption. Build it as a product and teams pull it.
The pain platform engineering answers is measurable, not fashionable. A 2026 empirical study of cloud-edge infrastructure complexity — 101 interviews across 86 organizations — found deployment complexity (38.6%) and onboarding difficulty (35.6%) were the top practitioner bottlenecks, and concluded that “complexity, not execution performance” is what limits delivery (Lertpongrujikorn, Nguyen & Salehi, arXiv). Every organization that grew past a handful of services knows the shape: each team assembling its own pipeline, its own Terraform conventions, its own secrets handling — all slightly different, all rediscovering the same answers.
The platform team, precisely
Team Topologies (Skelton & Pais) defines four team types — stream-aligned, enabling, complicated-subsystem, and platform — and gives the platform team the sharpest definition in the book: it offers “a compelling internal product to accelerate delivery by stream-aligned teams.” The motivation is cognitive load: “teams can only handle so much complexity before breaking down,” and overloaded teams “make poor decisions and move slowly.” A platform is how an organization spends architecture once, centrally, instead of spending every team’s attention forever.
The CNCF’s platforms white paper defines the artifact itself: “a platform for cloud-native computing is an integrated collection of capabilities defined and presented according to the needs of the platform’s users” — and insists on the operating model: platforms “should be designed and evolved … similar to any other software products,” with user research, product management, and — their word — internal marketing. Their warning is empirical, not rhetorical: platforms imposed top-down “breed resistance.”
Golden paths: the opinionated default
The golden path is the platform’s flagship product: a paved road from idea to production that bundles the organization’s answers — repo scaffold, CI/CD, observability, security controls, policy gates — so that the default route is also the compliant route. Two properties make it work:
- Self-service, not ticket. A developer scaffolds, deploys, and observes without a human intermediary. If the path runs through a queue, it is not a path; it is a toll booth.
- Escape is allowed, friction is honest. Teams may leave the paved road for legitimate needs — but they inherit the full operational bill of doing so, visibly. Adoption then measures whether the path is genuinely better, which is the feedback a product needs.
The product framing has a house style now: durable, ideate-build-run teams owning persistent problems rather than project teams that disband at go-live — product mode over project mode. A platform team that disbands after the launch will leave you with a scaffold and no roadmap.
Measure adoption with DORA, not vanity metrics
The four DORA metrics are the platform’s natural KPIs because they measure what the platform exists to improve: deployment frequency, change lead time, failed deployment recovery time, and change fail rate (DORA’s current names — recovery time has replaced the older MTTR vocabulary). A 2025 industry framework (MetricSynth, arXiv) describes aggregating exactly this pairing — DORA metrics plus domain KPIs — as platform telemetry. The platform dashboard answers one question: did the golden path make delivery faster, safer, or both, for the teams on it?
The evidence that platforms move those numbers exists and is worth quoting precisely. The 2024 DORA report — the research program describes itself as “the longest running academically rigorous research investigation of its kind” — found that “utilizing an internal developer platform improves individual productivity, team performance, and overall organizational performance.” The same report carries the caveat that separates platform engineering from platform theater: platforms “can also lead to decreased change stability and throughput” when done poorly. The platform is a product; products can fail in the market.
The platform is also a decision record
Every element of a golden path is a standing architectural decision: the CI system, the secrets approach, the observability stack, the deployment topology. Research on self-serve platforms (van Eijk et al., ICSA 2024) catalogs these as explicit design decisions — collected from 43 gray-literature sources and expert interviews — which is exactly the discipline to import: the platform’s own choices belong in the ADR catalogue, reviewable and supersable like any other architecture. On the E-Tax programme, the platform selections — API management, service mesh, CI/CD, observability — were governed as numbered records, and the golden path inherited them as defaults. The platform is where recorded decisions become someone’s paved road.
Four ways platforms fail
- Built without users. A platform nobody asked for becomes shelfware; the CNCF paper’s user-research insistence exists because this is the modal failure.
- Mandated instead of adopted. Compliance-by-decree produces compliance theater — workarounds, shadow pipelines, and the resistance the white paper predicts.
- No product ownership. Without a roadmap and an owner, the platform decays to the mean of its oldest choices; golden paths go unpaved.
- Measured on delivery, not adoption. If nobody tracks whether stream-aligned teams actually use the path — and whether their DORA numbers moved — the platform optimizes for demos.
Build your platform as a product
Sources
- Skelton & Pais, “Team Topologies — Key Concepts” (teamtopologies.com)
- CNCF TAG App Delivery, “Platforms for Cloud Native Computing” white paper
- DORA, “The four keys” — metric definitions
- DORA Research program
- DORA, “2024 Accelerate State of DevOps Report”
- Sriram Narayan, “Products Over Projects” (martinfowler.com, 2018)
- Lertpongrujikorn, Nguyen & Salehi, “Empirical Analysis of Cloud-Edge Infrastructure Complexity” (arXiv, 2026)
- Jain, Agrawal, Nigam & Patil, “MetricSynth: Aggregating DORA and KPI Metrics Across Multi-Platform Engineering” (arXiv, 2025)
- van Eijk et al., “Architectural Design Decisions for Self-Serve Data Platforms in Data Meshes” (ICSA 2024; arXiv)
- Helgesson, Appelquist & Runeson, “A Grounded Theory of Cognitive Load Drivers in Novice Agile Software Development Teams” (arXiv, 2021)