Loaded Crossplane's full doc set into MiniMax M3's 1M context window to speed up our evaluation
Been evaluating Crossplane for 8 weeks. Team lead got sold on it at KubeCon, came back like "we're moving off Terraform by Q1." Nobody on the team has touched Crossplane before and guess who gets to do the evaluation because I "have the most Kubernetes experience." Thanks.
The doc situation is brutal. Official docs, three provider doc sets, composition examples scattered across GitHub, XRD references that contradict the tutorials half the time. Spent an entire Saturday on composition patterns and by 4pm I was LESS confident than when I started. Not a great sign.
Did something kind of lazy. Downloaded everything, all provider docs, maybe 50 composition yamls from examples repo. Converted to plaintext, roughly 650k tokens. Loaded it into MiniMax M3 (handles 1M natively) and started asking the questions I'd been going in circles on.
Nested compositions referencing outputs without a separate Claim. This had been driving me nuts. Answer was spread across three doc pages AND a Feburary GitHub discussion I somehow missed. It nailed it, pulled the right section from each source.
Also needed ammo against my teammate who keeps insisting GCP networking is "production ready." M3 flagged several CRDs as still v1beta1. Confirmed in the provider repo. Sent him the screenshot at 11pm on Sunday. Maybe petty but whatever.
Where it ate it: ArgoCD integration. Mixed up the official GitOps guide with a 2024 community blog post. Config snippet would've broken our sync flow. Caught it because namespace didn't match staging. If I hadn't checked that would've been a fun incident.
Whole "250 pages across 5 repos" problem feels universal for infra tooling evaluation though. Anyone found a less painful workflow?