Re:Flow-IPC: Open-Source Toolkit for Low-Latency Inter-Proce...
by fwsgonzo on Saturday April 13, 2024 @01:39PM
(#40024655)
I tried to migrate to capn-proto but it just doesn't build on MinGW so I have no choice but to wait. Like you say, it gets worse the more I wait. But, if the APIs are somewhat sane, they should hopefully also be somewhat similar: Able to switch case on oneofs, movable data structures etc.
I don't like that protobuf has recently started linking with abseil, which despite being a good framework, I can't use it if it doesn't build absolutely everywhere I need it to. So, maybe I'll be forced over to CapnProto one of these days?
Re:Flow-IPC: Open-Source Toolkit for Low-Latency Inter-Proce...
by Arelius on Saturday April 13, 2024 @07:35PM
(#40027198)
Honestly I like capn-proto. But it's community is a bit light. Even Windows which is "supported" has been a bit fix-it-yourself.
If MinGW isn't already supported, if probably expect to have to take that over yourself.
Re:Flow-IPC: Open-Source Toolkit for Low-Latency Inter-Proce...
by sgtnoodle on Saturday April 13, 2024 @06:25PM
(#40026720)
I've also developed a strikingly similar low latency real-time IPC message bus for work. It also uses sockets with transparent shared memory optimization. In my case, it's the backbone for an autonomous aircraft's avionics. I made everything agnostic to the message scheme, though, and most of the tooling supports an in-house schema, protobuf, JSON, YAML, etc. There's also clients implemented in C++, Rust, Python and Julia.
What troubles has protobuf caused you?
Re:Flow-IPC: Open-Source Toolkit for Low-Latency Inter-Proce...
by signa11 on Monday April 15, 2024 @07:07AM
(#40038937)
> What troubles has protobuf caused you?
we are primarily a c++ shop, so all comments are from that vantage point. in no particular order following:
1. extremely bloated generated code.
2. wire format is not optimal, specifically for large nested messages which trashes the cache on a already severely constrained h/w.
3. not easy to perform shared-memory exchange for large messages, for example, if you have 'string' types in your messages, those cannot be allocated on arenas.
4. <this space is for rent>
we _unfortunately_ gravitated towards protobuf's despite my fervent appeal to go with capn-proto. that has caused a cascade of troubles / missed opportunities for optimizations etc. etc.