Like the similar resolved bug with the time-windowed forecaster's signal cache corrupting the input to the non-time-windowed forecaster, the eval snapshots may also corrupt the shared signal cache. We should move to using a separate cache, expanding the data that it downloads, or pre-downloading a larger portion of the data.
Like the similar resolved bug with the time-windowed forecaster's signal cache corrupting the input to the non-time-windowed forecaster, the eval snapshots may also corrupt the shared signal cache. We should move to using a separate cache, expanding the data that it downloads, or pre-downloading a larger portion of the data.