more efficient serialization of initial state for new worker

[Keksoj]

As pointed out in #1038, creating a new worker involves writing the config state on a file using serde_json::to_writer:

sozu/bin/src/worker.rs

Line 308 in 757a6f1

serde_json::to_writer(&mut state_file, state)

and conversely, using serde_json::from_reader to read the file again, from the new worker process:

sozu/bin/src/worker.rs

Line 220 in 757a6f1

let config_state: ConfigState = serde_json::from_reader(configuration_state_file)

which leads to performance issues, especially when the state is big.

This PR fixes this issue by using the same logic as in command sozu state save: convert the ConfigState to WorkerRequests and write them in a file, using a big buffer. Then parse it using nom to retrieve the requests.

This means changing the signature of functions that create workers. Instead of getting a ConfigState, they get Vec<WorkerRequest>.

Gain in performance

• intel i7-10510U
• 20000 clusters in the state
• 100 000 requests in the state file
• 4 workers
• upgrade command that restarts all 4 workers with the state file

Sōzu 0.15.17

Executed in   15.74 secs      fish           external
   usr time    7.04 millis    0.00 micros    7.04 millis
   sys time    4.32 millis  820.00 micros    3.50 millis

Sōzu 0.15.17 with the fix

Executed in    7.66 secs      fish           external
   usr time   29.88 millis   29.88 millis    0.00 millis
   sys time   14.20 millis    3.34 millis   10.85 millis

We do observe a doubling of performance, which is nice, but still improvable if we would serialize/parse the requests using protobuf (roughly and additionnal x2 increase in perf).

[Keksoj]

actually, optimizing worker creation with protobuf does not mean we have to merge this protobuf PR that changes channel behaviour and passes ServerConfig to the worker :

• Transmit config to new worker with a protobuf type + protobuf channels #1000

Pull request #1000 changes the type of Config that is passed to the worker, and passes it with a protobuf channel.

The fix of this PR here deals with a temporary file that is read only by the new worker process. We could write protobuf-serialized data in there.

The two pull requests are unrelated.

[Keksoj]

This fix here could be improved further by replacing JSON serialization with protobuf, but since it's a protobuf change, it could be done in the big protobuf PR later on.

[Wonshtrum]

Looks good to me!

[FlorentinDUBOIS]

Tested with a state that contains 350k requests and 15k certificates.
Before the patch, it tooks roughly 40s to start and after less than 2s