No, you don't need a GSI, but it might be more efficient depending on your circumstances.
Lets take your example of BorrowedTimestamp > 3days. Im going to assume this is for a particular user, so you have a userid to query.
You could do a query with a KeyConditionExpression of userid, then a FilterExpression of BorrowedTimestamp > 3days. Lets say the user has 10 books and 2 of them have a BorrowedTimestamp > 3days. This query will cost you 10 RCU (Read Capacity Units). That's because a FilterExpression just filters out items in your result set - DynamoDB actually found all 10 items in the query.
Now lets say you have a GSI where the partition key was userid and the range key was BorrowedTimestamp. Your KeyConditionExpression could specify both the parition key of the userid and the range key of BorrowedTimestamp > 3days. The result would be exactly the same. However this time it would only cost you 2 RCUs, and those RCUs would come from the index capacity not the table capacity.
Less RCUs sounds good, but remember you have to purchase throughput capacity for your primary index and GSI separately. This can be less efficient because you can't share purchased throughput between queries that use your primary key and GSI.
Finally if you didn't want to specify a userid at all you would use a scan. Scans sometimes don't scale well because they always evaluate every item in the table, but whether it works for you really depends on a lot of things (like how often you will use the scan, how many items you will have in the table etc).
