Extreme performance issue because JsonProvider is not cached

[erik-wramner]

Describe the bug
Common code for manual JSON uses Json.createObjectBuilder(), Json.createValue(...) and friends. These methods all delegate to the provider:

    public static JsonNumber createValue(Number value) {
        return JsonProvider.provider().createValue(value);
    }

Unfortunately, the provider is not cached and the provider() method is extremely expensive. When we replaced Json with AppJson, the only difference being that the provider is cached as a singleton, a job that took 14 minutes could be completed in 6 seconds. This really makes a difference.

To Reproduce
Use typical code that calls Json.whatever() and measure execution time for a large JSON document.

Expected behavior
The provider is cached so that the expensive initialization is not done every time a new JsonValue is created.

System information:

• OS: Windows, Linux
• Java Version: 17, 21
• Yasson Version: 3.0.4

Additional context
Yasson is pulled in by Quarkus, but that should not make a difference.

[KyleAure]

The creation of the JsonProvider and the static methods you discussed on the Json factory class are all part of the Json Processing API and not part of the Json Binding layer. I don't think there is anything we can do in Yasson or Json Binding to fix this performance issue.

I did some digging and looks like this issue has already been reported several times in the Json Processing project:
jakartaee/jsonp-api#26
jakartaee/jsonp-api#154
jakartaee/jsonp-api#346

[dasTanmoy0096]

While I am not too familiar with the Jakarta EE world, it basically means Yasson can't do anything because it's mandated by the spec?
Came here because a single Json response while using Helidon, takes about a second or so. And that's the time even without any database operations involved.

[erik-wramner]

That's really bad. If Yasson/Json Binding has more traction, perhaps you can push this with Json Processing? It is embarrassing that something as serious as this has been allowed to remain for years like that when the issue is known.

When JavaOne was still a thing, I would have raised it there, but it is harder to navigate now.

[KyleAure]

I did raise this this issue with a colleague who works on the JSON Processing side. 🤞🏼
The only thing we could do on the Yasson/JSON-B side would be to avoid using these methods until they are fixed, and we already do that.

[KyleAure]

Closing this issue as this must be fixed in the Jakarta JSON Processing API.