5 ms·
There's nothing fundamental about protobuf itself that prevents streaming output. You just have to make a pass over the proto structure first to compute the si
by thedance 7y ago
There's nothing fundamental about protobuf itself that prevents streaming output. You just have to make a pass over the proto structure first to compute the sizes, then stream it out using the precomputed sizes. At no time do you necessarily need the entire representation of the output in memory. The C++ library, for example, offers this.
- neuland 7y agoYes, and that's what I did. In my case, I was writing to a file. And my languages were C, Python, and Lua. So, I wrote a small package (plus bindings for Python and Lua) that would write out the message length as a uint64, then a uint8 for which schema the message is (I was working with multiple different message types), and then the actual message bytes. So, it's do-able. But, I'm just saying that other formats/libraries support this natively out of the box.
- thedance 7y agoIt sounds a bit like you invented RecordIO[1]. I didn't realize you meant decoding from a stream containing several protobufs. In that case you do indeed need some kind of framing, because there is no end-of-message delimiter. The positive tradeoff is you can easily concatenate several protobufs into another, valid protobuf. That is impossible if there is a terminating symbol. 1: https://www.tensorflow.org/tutorials/load_data/tfrecord#tfrecords_format_details https://www.tensorflow.org/tutorials/load_data/tfrecord#tfre...
- neuland 7y agoOh neat. I'm not familiar with the ML/AI ecosystem. So, I didn't think to look there for formats; a bit of a shame I reinvented the wheel here, but it worked great for my specific use case. Yeah, it's all about the tradeoffs with formats. Protobuf is a good choice a lot of the time and I always think about it as one of the first/preferred options. gRPC is great too.