monolog
monolog copied to clipboard
Seldaek
Issue with Logging to stdout in Laravel 10.21 using Monolog v3 EXCEPTION#Cannot log request: fwrite(): Write of 378 bytes failed with errno=32 Broken pipe
Monolog version 3
Hello,
I'm currently working with Laravel 10.21 and have encountered an issue when attempting to write logs to stdout using Monolog version 3. My configuration is as follows:
'stdout' => [
'driver' => 'monolog',
'handler' => StreamHandler::class,
'formatter' => '\Monolog\Formatter\JsonFormatter',
'with' => [
'stream' => 'php://stdout',
],
],
However, this setup results in an exception:
Exception: /var/www/html/vendor/monolog/monolog/src/Monolog/Handler/StreamHandler.php::162 -> EXCEPTION#Cannot log request: fwrite(): Write of 378 bytes failed with errno=32 Broken pipe
I noticed a similar issue previously reported as #1634, but it seems the problem persists. Any guidance or suggestions to resolve this would be greatly appreciated.
Thank you for your support.
Best regards,
It is because of this:
https://github.com/Seldaek/monolog/blob/main/src/Monolog/Handler/StreamHandler.php#L118
is_resource is not enough to handle writing to an already closed socket
[EDITED - the first solution was not the proper one]
The library should handle recreation of $this->stream on fwrite() error here: https://github.com/Seldaek/monolog/blob/main/src/Monolog/Handler/StreamHandler.php#L149
As the only way to tell if the pipe is broken is basically when trying to write to it.
Also assuming fwrite will always succeed is a bit overoptimistic: https://github.com/Seldaek/monolog/blob/main/src/Monolog/Handler/StreamHandler.php#L160C24-L160C24
Sample pseudocode for this (as fwrite does not have to write everything in single call we could also check number of bytes written):
protected function streamWrite($stream, LogRecord $record): bool
{
$result = fwrite($stream, (string) $record->formatted) === false
if ($result === false)
return false
if ($result === length($record))
return true;
return false;
}
Although locally the fwrite should be usually able to write all the bytes to the stdout (although I would not be doing the same on network sockets).
See also comments on #1634 - I'm still waiting for a way to narrow this down.
@Seldaek thanks.
This is really the same class of problem like programming over the network (although the pipe is a local object, it can be closed by the OS at any time really) - closing the pipe on the "remote" side (so the side external to the library execution context) is not detectable until you try to write to the pipe. The explanation of this can be read here - there is not much we can do about it really as it is all coded in the kernel:
https://stackoverflow.com/questions/8369506/why-does-sigpipe-exist
Yeah fair enough, maybe we need to check before writing or handle that failure at least and reopen the stream if possible. I am still curious why this happens now and it didn't before though.
Well, possibly due to file descriptors leak, I think this is the obvious scenario that can happen locally.
Is it possible something with the server images changed, outside of any PHP changes, that could have caused this?
Very unlikely unless PHP 8 implements is_resource or write() differently. Writing to an already closed socket will always behave like that.
Other option is lower sysctl limits for file handles (so the bug has always been there but had less chance of occuring).
https://github.com/Seldaek/monolog/issues/1862
There is also this interesting claim - which would explain the issue.
We have the same issue in production, occured after upgrade to PHP 8.2 and Monolog 3
EDIT: so far it seems that file descriptors build up until reaching the limit of the kernel, preventing any new fwrite. Rebooting our servers fix the issue for a while. It seems like the stream isn't closed properly, or at least not in all cases. I've noticed that it should be done during the __destruct() phase of the handler, but is there some edge cases where it isn't called? (php-fpm process crash or something else)
Hello! Encountering the same issue on many of my projects as I log to stdout with json formatter.
Just to grasp this better, as it seems nobody feels like digging into it (+1s are nice buuut they don't solve problems..). Does it happen only in long running worker processes, or also when using fpm just processing web requests?
For me, it only happened in fpm web requests. Our jobs never had this issue.
Ok thanks, this is odd though..
[@Rydgel] so far it seems that file descriptors build up until reaching the limit of the kernel, preventing any new fwrite. Rebooting our servers fix the issue for a while. It seems like the stream isn't closed properly, or at least not in all cases. I've noticed that it should be done during the
__destruct()phase of the handler, but is there some edge cases where it isn't called? (php-fpm process crash or something else)
The way PHP works, resources should be closed and file handles released at the end of a request, even if you never call fclose() explicitly. It just garbage collects everything, so if it started happening with PHP 8.2 I am wondering if it is some bug in PHP itself.
Per https://github.com/Seldaek/monolog/issues/1634#issuecomment-1864715807 maybe PHP 8.1 is the cause?
I still don't really see any clear pattern here, except for everyone using Laravel.. So maybe it's something Laravel does?
Closing the handle every time seems to cause a 15-20% slowdown in log writes vs just opening it once, if you do log a lot. If you log only one line obviously it is the same, so I am not sure if this is acceptable or not. Perhaps if we ensure at least batch writes happen in one stream open the impact would be small enough.
If you want to try the fix, you can require monolog/monolog dev-fclose2 as 2.99 (2.x) or dev-fclose3 as 3.99 (3.x). I'd be happy to get confirmation that this fixes the problem before merging as it is strictly speaking a performance-degrading patch, I don't want to do it if not worth it.
See #1882 (2.x) and #1883 (3.x) for the patches.
Hey @Seldaek my team experiences it on custom PHP application. I can perhaps confirm that the issue started with 8.x branch however I would not necessarily assume that the resources are always "garbage-collected" - we use PHP-FPM and the worker processes will live through some requests.
@Seldaek and perhaps others - can you try out this quick patch I composed ? Sadly I dont have time to test it.
https://github.com/Seldaek/monolog/pull/1884
As already mentioned, only happens in fpm context, not for long running processes/cronjobs.
~I will check if we can run the patch to see if it resolves the issue.~
Going to try out the patch now.
@Seldaek you write that "except for everyone using Laravel".. We don't use laravel. We have symfony project, php version 8.1.27, monolog version 2.9.1. And this issue happening in fpm requests.
Ok then Laravel is excluded too. I really don't know what is causing it except for having too many parallel requests/workers doing stuff perhaps and reaching system file handle limits. Anyway hopefully my patch does help.
Ok then Laravel is excluded too. I really don't know what is causing it except for having too many parallel requests/workers doing stuff perhaps and reaching system file handle limits. Anyway hopefully my patch does help.
Our application is running in docker. It seems that there is no limit on the number of files in docker (systemctl show docker | grep LimitNOFILE= output LimitNOFILE=infinity). Also pm.max_children setted to 15, so unrealistic that we can hit hard limit of open files. If only someone could make a demo stand with load testing on a clean application to try to reproduce this issue. And may be edit pm.max_requests to 1 to make sure that the case is in php-fpm. If no one will take it, then I can do it, but later
composer require monolog/monolog:"dev-fclose3"
gives me, any quick idea how to resolve it?
Problem 1
- Root composer.json requires monolog/monolog dev-fclose3, found monolog/monolog[dev-fclose3] but these were not loaded, likely because it conflicts with another require.
Problem 2
- launchdarkly/launchdarkly-php is locked to version 4.3.0 and an update of this package was not requested.
- launchdarkly/launchdarkly-php 4.3.0 requires monolog/monolog ^1.6|^2.0|^3.0 -> found monolog/monolog[dev-main, 1.6.0, ..., 1.x-dev, 2.0.0-beta1, ..., 2.x-dev, 3.0.0-RC1, ..., 3.x-dev (alias of dev-main)] but it conflicts with your root composer.json require (dev-fclose3).
Problem 3
- illuminate/log v10.48.8 requires monolog/monolog ^3.0 -> found monolog/monolog[dev-main, 3.0.0-RC1, ..., 3.x-dev (alias of dev-main)] but it conflicts with your root composer.json require (dev-fclose3).
- laravel/lumen-framework v10.0.3 requires illuminate/log ^10.0 -> satisfiable by illuminate/log[v10.48.8].
- laravel/lumen-framework is locked to version v10.0.3 and an update of this package was not requested.
I guess composer.json should have an entry in:
"extra": {
"branch-alias": {
"dev-main": "3.x-dev"
}
},
@TheLevti that's why I said to put "dev-fclose3 as 3.99" as require so that composer aliases it and it passes the ^3.0 requires other packages may have on it.
Worked, running it now on our test system. This might take some time to validate. So I guess in a few days I will let you know if the issue still occurs.
Would be nice if someone else with the same issue can do the same.
Great, thanks for trying, I'll merge and release in a couple weeks maybe once we have some real world feedback :)
Instead of closing the file after each log, could we do that on shutdown? Like registering a shutdown function or so to close the resources. Or if its during a framework lifecycle, part of a request/response cleanup step.
🤔
Ok,let's see where the patch leads us. We are going to deploy this change further into production next week and monitor it there. Unfortunately it's hard to reproduce reliably, that's why I think it would be important that more people deploy it in their application and report back their result.