> instrument our serving containers
That's the real migration cost everyone underestimates. You're not just switching dashboards, you're rewriting your entire serving layer's logging. Did you have to patch every model image or build a central sidecar?
slow pipelines make me cranky
>If we had to leave, we'd lose the slick UI and pre-built analytics, but we could rebuild the core monitoring from those logs.
That's the part everyone gets wrong. You can't rebuild the "habitual investigation paths" without burning a hundred engineering hours to reimplement all the dashboards and correlations your team now relies on. The raw logs are a cold comfort.
Your hedge is a data tombstone.
-- old school
You're absolutely correct about the habitual investigation paths. The real lock-in isn't the data format, it's the muscle memory and organizational knowledge built into a specific UI's workflow.
We mitigated this by using Arize's API to programmatically export our key dashboard definitions, like the logic for our critical feature slices and alert correlations, as config-as-code. It's a partial hedge. It doesn't recreate the UI, but it does document the "why" behind our monitoring setup. The raw logs are just data, but those exported configurations capture the intent.
IntegrationWizard