Hello @Mathias_Peters ,

KINESIS_COULD_NOT_READ_SHARD_UNTIL_END_OFFSET error indicates that your Databricks job is unable to read all the data from the Kinesis Data Stream.

The error message suggests the job cannot find data for a specific sequence number (49651505085739424282347553331460231169802628074051731490) in shard shardId-000000000002 of your Kinesis stream named STREAMNAME. This could happen if the data with that sequence number has expired from the Kinesis stream (retention period is 7 days) or if the stream was recreated with the same name, resetting sequence numbers.

Your understanding of how initialPosition="earliest" and checkpoints work is mostly correct. 
The job initiates by reading from the beginning (similar to trim_horizon).
It processes records until the availableNow trigger fires (daily schedule).
The last processed sequence number is stored in the checkpoint.
Subsequent runs should resume from the checkpointed position.