Just to expand on the other answers here, I would point out that the 64Mb limit is documented here: https://milvus.io/docs/limitations.md#Input-and-Output-per-RPC

Note that this is the size of the gRPC call which will contain your data, presumably with some additional metadata appended, serialised into protobuf format. The implication of this is that it is not easy for your python program to determine a priori whether a payload is going to exceed the limit.

I think that the only way to address this is just to be really defensive as you load data into a milvus collection. In other words, do the insert in batches and if you see this error attempt to retry it with a smaller batch size. Just make sure that you are being very specific with the errors you handle and take care to ensure that you won't leave a collection partially updated if it fails midway through.