Hey @jrobison! Thanks for reaching out. The builds that are succeeding are being kicked off outside of builder (specifically via build schedules). This means they’re using the latest pipeline logic that successfully deployed. I think raw builds and build schedules will try to run on whatever input data is passed in from upstream. I would not be surprised at all if there is type coercion that’s happening behind the scenes. For pipeline builder itself, we ensure that new deployments of the pipeline are type compatible with the upstream data!
One quick workaround that you until we enable full data expectations on inputs is to create a dummy pipeline that outputs a dataset with data expectations on it, and then use the output from that as the input to your pipeline!