OpenLineage is a good start, but it's fundamentally a *post-execution* telemetry spec. It tells you what happened, not *how* it can be replayed or migrated. The rigidity for compliance demands a deterministic state machine, and that's the lock-in vector.
The real alternative isn't an export format, it's building on a portable runtime core. I've forced this by using Argo Workflows with WorkflowTemplate CRDs stored in Git. The state is just Kubernetes custom resource status, which, while verbose, is a known schema. The audit trail becomes the commit history plus the CR status logs. It moves the vendor from the orchestrator platform to the k8s distro, which is a marginally more open market.
But you're right about the pricing lever. Even with this, your operational metadata is held hostage by the platform's control plane if you use their hosted version. The only real mitigation is to run the control plane yourself, which brings back the transactional cache invalidation problem from earlier in the thread. There's no free lunch, just a choice of which SRE burden you prefer.
Boring is beautiful
Exactly. That 40% premium is for the sticker that says "AI-Ready". You're buying a marketing checkbox, not architectural changes.
The cron job scheduler analogy is spot on. They just swap `python script.py` for `docker run --env OPENAI_API_KEY`. The audit trail is still a bunch of JSON lines buried in Splunk that you can't query without their proprietary schema.
Real non-determinism means the orchestrator needs to handle retry logic that isn't just "task failed, try again". What about a step that yields three possible next steps and you need to log the LLM's reasoning for the branch it chose? Their "AI module" can't do that. It's just a wrapper.
-- old school