body content sometimes (!) missing in async content handler about nginx-clojure

[xfeep]

Hi Andreas,
Thank you very much. You provided very useful information.
Now I can confirm that it will be a problem if we read body out of the sope of handler.invoke(req) because body will be closed at the end of handler.invoke(req).
So if your thread runs quickly before the end of handler.invoke(req) you can read the body successfully otherwise you will get an closed inputstream.
In the futerue version I will consider the tradeoff between resource leak risk and flexibility.

[xfeep]

Hi, Filler.
Could you please tell us what value is bodyInputStream.getClass() when the bodyInputStream is missing?
There're three kinds of bodyInputStream. The first is FileInputStream when Nginx uses tmp body file. The second is NativeInputStream and the last is SequenceInputStream where stores a list of NativeInputStream.

[afiller]

Hi xfeep,
thank you for your reply and - by the way - for your great project. The library is so amazing and I love using it, it fills such an important gap by realising dynamic NGINX redirects.

I just tested it out and it's in both cases (contains body/misses body) a SequenceInputStream:

2023-08-17 09:59:00[warn][21190][pool-1-thread-1]>>>java.io.SequenceInputStream
2023-08-17 09:59:00[warn][21190][pool-1-thread-1]<<<<

2023-08-17 09:59:02[warn][21186][pool-1-thread-1]>>>java.io.SequenceInputStream
2023-08-17 09:59:02[warn][21186][pool-1-thread-1]<<<<{"event": {"event_timestamp_ms": 1692259142008, "product_id":....

Best regards,

Andreas

[xfeep]

Hi, Andreas,
I'm glad to see this library is useful to you.
Please show the two values of bodyInputStream.available(). The first one is got at ExternalRESTCallHandler.invoke(request). The second one is got at InvocationHandler.run().
BTW I find thead pool mode is used (viz. jvm_workers > 0) and requst.hijack(..) / request,handler().hijack(..) is used together with thead pool mode. In most general cases requst.hijack(..) is used with plain mode to avoid blocking nginx.

[afiller]

Hi xfeep,

we never had problems with thread pool mode and NGINX and we are using this configuration quite long already. But I will keep an eye on it after your input.

The result is interesting:

In NOT successful cases the .available() is 401 and later 0.

In successful cases (so the content can also be read in run()) it is 1241 in both methods, so before starting the thread and inside the thread.

Summarized: When the stream is already at 1241 it is also possible to read it later. If it's not (401) it's empty later (0).

Hope it helps a bit in investigating. I keep my setup as it currently is for further testing.

[afiller]

PS:

We are not using "jvm_workers" (so the default of 0 should be used). The thread name comes from the thread pool executor I use in my class NginxJavaRingHandler for handling the parallel request:

Constructor:
ExecutorService executorService = Executors.newCachedThreadPool();

Usage:
// Run thread in thread pool executor
executorService.execute(new InvocationHandler(nginxRequest, channel));

...so this is simply internally in my class to be able to handle the hijacked requests in parallel.

[xfeep]

PS: java.io.SequenceInputStream.available() only returns the first stream's available(). If there're serveral chunks in the body the value which available() returns is not always the total length of the body. So both .available() is 401 and .available() is 1241 are OK.

[afiller]

Thank you for your explanation and clarification. I adjusted my implementation now that it reads the body before going starting into the thread. The body will be handed over in the constructor of the same. It's less threaded in the end, but working stable. Thanks again!

[xfeep]

@afiller Hi Andreas,

Please try a workaround way for better performance.

// get a user mananged body reference
Inputstream  umbody = (Inputstream)MiniConstants.BODY_FETCHER.fetch(nginxRequest.nativeRequest(), null);
// access umbody in another thread and then close it.
...

I've also created a related issue for the future enhancement about the hijack API.