Improvements from Franken 1.2.2 to 1.5.0

[thomas-hiron]

Hello,

I just updated Franken from 1.2.2 to 1.5.0 and I have two questions.

Configuration

First, here is my prod configuration:

{
    frankenphp {
        num_threads 0
        worker {
            file /app/public/index.php
            env APP_RUNTIME Runtime\FrankenPhpSymfony\Runtime
            env FRANKENPHP_LOOP_MAX 500
            num 80
        }
    }

    auto_https disable_redirects
}

* {
    encode zstd br gzip

    route {
        root * /app/public/
        php_server
    }
}

Before the update, available RAM was like this, workers restarting every 500 request:

But now with 1.5.0, here is the new RAM graph:

First question

How is this even possible that worker restart release so little RAM and consume almost no RAM?
I would have expect better RAM handling with less memory leaking but not that much!

Second question

This one may explain the first one if my understanding is correct.

It's about database connections. I had continuously 80 DB connections, one for each worker:

But now I have around 15 connections max:

Does this mean not all workers are handling requests?

Thought

I have seen this reply: #1486 (reply in thread) which says:

The reason you are seeing 1, 2, 3, 4, 5, 6 instead of 1, 1, 1, 2, 2, 2, 3, 3, 3 is that requests are dispatching to threads in order. So if thread 1 is busy, check thread 2, if thread 2 is busy, check thread 3 and so on. This helps workers be more efficient by primarily using the 'hot' threads and minimizes connections to external services.

Has it been introduced after 1.2.2?

Is this the reason for my DB connections drop? As long as I don't have 80 concurrent requests, I won't have 80 DB connections?
That may also explain RAM usage: what is responsible for keeping the app in memory, the worker or the thread its using?
For exemple, if I have 15 concurrent requests at most, my app is bootstrapped only 15 times?

That's a lot of questions in the end, but I have a big sale event in a month and I'd like to understand better how are workers and threads working!

In any case, thanks for the work, this is huge! 👏

[AlliBalliBaba]

Yeah this performance improvement was introduced in #1126 (1.3 I think). We changed it from dispatching requests round-robin-style to prioritizing lower-number worker threads, which reduces the number of connections, memory usage and contention. You should only need 80 connections if you actually handle 80 concurrent requests.

The app will still be bootstrapped 80 times with 80 threads though, connections are usually only established once they are first needed. 1.5 also introduced a new max_threads configuration, which allows starting threads at runtime and actually only bootstraps once needed. So a configuration like this would always have 5 threads, but bootstrap up to 80 threads once they are needed (and stop them if they are not):

frankenphp {
    max_threads 80
    worker {
        file /app/public/index.php
        num 5
    }
}

The spikes you were seeing before in memory usage were probably due to a small memory leak on big POST requests, which was fixed in #1350 and a general known memory leak in older PHP versions. Memory consumption should currently be pretty stable across requests, I'd still recommend restarting threads after a few thousand requests though.

[thomas-hiron]

Got it, thanks! Your answer is really helpful.

I did some checking and I've seen that the app is bootstrapped at startup by each PHP thread. So that's great for me, I need the app ready as soon as possible!

About FRANKENPHP_LOOP_MAX configuration, is it 500 requests by PHP thread or for the worker?
And what happens to the 80 bootstrapped app?

[AlliBalliBaba]

It's 500 per thread. The 80 bootstrapped threads will just be dormant and wait for requests. Threads waiting for requests don't consume CPU, but will consume about 100Kb + the weight of your bootstrapped app in memory.