# Description
This helps to ensure data produced on a stream is immediately available
to the consumer of the stream. The BufWriter introduced for performance
reasons in 0.93 exposed the behavior that data messages wouldn't make it
to the other side until they filled the buffer in @cablehead's
[`nu_plugin_from_sse`](https://github.com/cablehead/nu_plugin_from_sse).
I had originally not flushed on every `Data` message because I figured
that it isn't really critical that the other side sees those messages
immediately, since they're not used for control and they are flushed
when waiting for acknowledgement or when the buffer is too full anyway.
Increasing the amount of data that can be sent with a single underlying
write increases performance, but this interferes with some plugins that
want to use streams in a more real-time way. In the future I would like
to make this configurable, maybe even per-command, so that a command can
decide what the priority is. But for now I think this is reasonable.
In the worst case, this decreases performance by about 40%, when sending
very small values (just numbers). But for larger values, this PR
actually increases performance by about 20%, because I've increased the
buffer size about 2x to 16,384 bytes. The previous value of 8,192 bytes
was too small to fit a full buffer coming from an external command, so
doubling it makes sense, and now a write of a buffer from an external
command can be done in exactly one write call, which I think makes
sense. I'm doing this at the same time because flushing each data
message would make it very likely that each individual data message from
an external stream would require exactly two writes rather than
approximately one (amortized).
Again, hopefully the tradeoff isn't too bad, and if it is I'll just make
it configurable.
# User-Facing Changes
- Performance of plugin streams will be a bit different
- Plugins that expect to send streams in real-time will work again
# Tests + Formatting
- 🟢 `toolkit fmt`
- 🟢 `toolkit clippy`
- 🟢 `toolkit test`
- 🟢 `toolkit test stdlib`
# Description
This breaks `nu-plugin` up into four crates:
- `nu-plugin-protocol`: just the type definitions for the protocol, no
I/O. If someone wanted to wire up something more bare metal, maybe for
async I/O, they could use this.
- `nu-plugin-core`: the shared stuff between engine/plugin. Less stable
interface.
- `nu-plugin-engine`: everything required for the engine to talk to
plugins. Less stable interface.
- `nu-plugin`: everything required for the plugin to talk to the engine,
what plugin developers use. Should be the most stable interface.
No changes are made to the interface exposed by `nu-plugin` - it should
all still be there. Re-exports from `nu-plugin-protocol` or
`nu-plugin-core` are used as required. Plugins shouldn't ever have to
use those crates directly.
This should be somewhat faster to compile as `nu-plugin-engine` and
`nu-plugin` can compile in parallel, and the engine doesn't need
`nu-plugin` and plugins don't need `nu-plugin-engine` (except for test
support), so that should reduce what needs to be compiled too.
The only significant change here other than splitting stuff up was to
break the `source` out of `PluginCustomValue` and create a new
`PluginCustomValueWithSource` type that contains that instead. One bonus
of that is we get rid of the option and it's now more type-safe, but it
also means that the logic for that stuff (actually running the plugin
for custom value ops) can live entirely within the `nu-plugin-engine`
crate.
# User-Facing Changes
- New crates.
- Added `local-socket` feature for `nu` to try to make it possible to
compile without that support if needed.
# Tests + Formatting
- 🟢 `toolkit fmt`
- 🟢 `toolkit clippy`
- 🟢 `toolkit test`
- 🟢 `toolkit test stdlib`
