Sure, I mean, I want to be clear—there's certainly a lot of inefficiency here. But my point is that the OP is talking about their data processing pipeline taking HOURS to handle queries for 50,000 IDs (50 parallel queries—1,000 IDs per query). When it comes down to it, I just don't believe that the inefficiency in serializing 1,000 numbers in python and then deserializing them in BigQuery has anything to do with the problems OP was experiencing. Remember—OP is probably going to be getting back 50,000 rows from the database anyway, the additional work here is linear with respect to the amount of rows retrieved. Sending 50,000 numbers over the wire to get back 50,000 rows worth of data is not an "hours long" processing cost, in the scheme of things.