Most platform teams measure success by how much of the company uses what they built. Ricardo Castro, who leads platform and reliability work at FanDuel, thinks that number often measures the wrong thing.
He talks through the "ivory tower" trap of building toward a conference talk instead of your own data, the two warning signs that a platform is drifting, why mandated adoption can hide the real problem, and the single blunt question that surfaces more developer pain than any satisfaction survey: "what's the most annoying thing you do on a day-to-day basis?"
Chapters
- 0:32 — From computer science to platform engineering
- 2:57 — The ivory tower and field of dreams fallacy
- 6:46 — Two warning signs your platform team is drifting
If people don't want to go to your platform, that's a very big warning sign.
- 8:58 — Going to the gemba: sitting with developers
You can't tell whether a platform decision was right without the data to check it against.
- 14:50 — The single best question to ask developers
- 16:59 — Bridging engineering data to what leadership cares about
- 20:50 — Why mandated adoption can be a vanity metric
- 23:48 — Untangling DevOps, SRE, and platform engineering
DevOps is not a tool, is not a team, is not an engineer. It's more of a philosophy.
- 33:43 — Where AI actually helps SRE and platform teams
Key takeaways
- A platform with 80% adoption isn't necessarily a success story — if that adoption was mandated rather than chosen, the real number worth worrying about is the 20% still resisting, or worse, everyone secretly wishing they could leave.
- The single best question to surface real developer pain is blunt and specific: "what's the most annoying thing you do on a day-to-day basis?" — not a satisfaction survey.
- Dogfooding only works if platform teams remove their own shortcuts: keeping a backdoor console that skips the pipeline everyone else has to suffer through removes the incentive to ever fix it.
- DORA and DX metrics aren't proof a platform is succeeding — they're a precursor to success. Hitting the numbers doesn't guarantee anything; a team that stops there has mistaken the leading indicator for the goal.
- Rolling out a full platform stack at once — Kubernetes, Backstage, Argo CD — modelled on someone else's conference talk can make core metrics like change failure rate measurably worse, not better.
- "AI SRE" gets reduced to incident-response chatbots in most framing, but Castro sees more platform-wide value in using historical data to flag developing problems before they become incidents at all.
