The whole “zero overhead” thing is IMO a red herring. I care about a few things: stability across versions and languages, space efficiency (sometimes) and performance. I do not care about “overhead” — performance trumps overhead every time.

Your deserializer is probably running on a CPU, and that CPU probably has a very fast L1 cache and might be targeted by a compiler that can do scalar replacement of aggregates and such. A non-zero-overhead deserializer can run very quickly and result in the output being streamed efficiently from its source and ending up hot in L1 in a useful format. A zero-overhead deserializer might do messy reads in a bad order without streaming hints and run much slower.

And then to get very very large records, as in the OP, where getting a good on-disk layout may require thought. And, frequently, the right layout isn’t even array-of-structs, which is why there are so many tools designed to query column stores like Parquet efficiently.