Use orjson to de-/serialize json-rpc messages

[deathaxe]

After a recent discussion about json serialization performance, I was curious about impact orjson makes on LSP processing performance.

This PR showcases how orjson as transport de-/serializer.

During recent 2 weeks of daily use, I haven't realized any issues, but also my projects are small enough to not realize any significant impact on overall performance.

Maybe someone using heavy language servers such as typescript can try to benchmark it.

[predragnikolic]

It can be tested with this package that Rafal wrote
https://github.com/sublimelsp/LSP-performance-test

[deathaxe]

According to this plugin, orjson doesn't make any difference. Tests results in 0.189 - 0.192 with both json and orjson on my machine.

[predragnikolic]

This thread has some information of where bottle necks are located, and how they were solved
#2190

[deathaxe]

Yeah, and those include json serialization performance being an aspect. With promised performance of orjson I would have expected a little improvement with regards to deserialize messages from language servers. Maybe payload is just too small to outweight overhead for python->C/rust->python translations.

[rchl]

That test tests performance of our format_completion and not json deserialization. It would have to be modified.

[rchl]

I've added relevant tests in https://github.com/sublimelsp/LSP-performance-test

It does appear to be twice as fast:

orjson
Best of 3: 0.09664862500000027 usec. All times: [0.0971984999999993, 0.09785637499999922, 0.09664862500000027]

json:
Best of 3: 0.197854125000001 usec. All times: [0.2019498329999987, 0.19887404199999992, 0.197854125000001]

@deathaxe are you able to get this PR into a state that passes CI?

[rchl]

For the record, that was parsing of a 50MB payload.

Here are also results for parsing a tiny payload 10000 times:

json
Best of 3: 0.02308479100000227 usec. All times: [0.023518416999991132, 0.02308479100000227, 0.023143875000002367]

orjson
Best of 3: 0.0058567079999996 usec. All times: [0.006264167000001208, 0.006002708000000467, 0.0058567079999996]

Seems like orjson is still faster.

[deathaxe]

No idea how to satisfy pyright's limitations.

I would prefer an import-time decision about using orjson or json. But pyright doesn't get that right.

It doesn't even resolve orjson import.

[rchl]

Add orjson==3.10.7 in:

• LSP/.github/workflows/main.yml

  Line 44 in 603632e

  - run: pip3 install mypy==1.7.1 flake8==5.0.4 pyright==1.1.381 --user

• LSP/tox.ini

  Lines 22 to 25 in 603632e

  	deps =
  	mypy==1.7.1
  	flake8==5.0.4
  	pyright==1.1.381

[rchl]

I would prefer an import-time decision about using orjson or json. But pyright doesn't get that right.

Probably good enough as is. Getting the types right otherwise does indeed seem tricky.

[jwortmann]

orjson
Best of 3: 0.09664862500000027 usec. All times: [0.0971984999999993, 0.09785637499999922, 0.09664862500000027]

json:
Best of 3: 0.197854125000001 usec. All times: [0.2019498329999987, 0.19887404199999992, 0.197854125000001]

For the record, that was parsing of a 50MB payload.

Just for comparison, the time for a single frame on a 120 Hz display is 8333.3 µs, so the performance improvement of parsing a 50 MB payload would make around 0.0012 % of a single frame time on a 120 Hz display. Not sure if that is worth to introduce another dependency...

Also, such big payloads only seem to happen for completions, e.g. with LSP-typescript, where the server response for 36 MB already takes 2.9 seconds (see #2190). Then there is the time needed to create the CompletionItem objects and to pass them over the ST API, so we're probably at 3+ seconds. If the values from above are correct, we can now save 0.0000001 seconds from JSON parsing (for 50 MB). It doesn't look like to me that this would be the bottleneck...

Edit: #2190 (comment) says current JSON parsing takes 0.293 seconds, so there is probably an error in your reported times? Should it be seconds instead of usec?

---

The Python code could be simplified to single lines:

return orjson.loads(message) if orjson else json.loads(message.decode('utf-8'))

[deathaxe]

The Python code could be simplified to single lines:

That's just another flavor, but no real simplification. Actually such one-liners are often harder to read and don't execute faster by any means.

[rchl]

Edit: #2190 (comment) says current JSON parsing takes 0.293 seconds, so there is probably an error in your reported times? Should it be seconds instead of usec?

The output of the test is actually in seconds.
I guess I got confused because according to the documentation, the CLI variant of timeout outputs in usec.