Do we really need the pure zero-copy?

[ZhenshengLee]

Do we really need the pure zero-copy transport that avoid all the copys?

one of proposition views:
from ros-perception/image_common#212 (comment)

relies on populating the message via toImageMsg, which does a memcpy on the image data – for me, this is the bottleneck

one of oppostion views:
from https://github.com/continental/ecal/discussions/660#discussioncomment-2738856

Without knowing your use case I still want to say: Zero-Copy sounds awesome, but what you rather should optimize for is performance, independent from where it comes from. As long as your messages are not "that big", one additional copy will not take enough time to be noticeable and you can often achieve more effect by optimizing other parts of your code (e.g. the protobuf pitfalls I mentioned). Our applications ran just fine in the past without zero copy.

So, what's your opinion about the need and the possibilities of the zero-copy in robotics system?