Next

Sentry Issue: ENGINE-HF

ErrorException: Allowed memory size of 134217728 bytes exhausted (tried to allocate 116526568 bytes)
  File "/lib/elgglib.php", line 947, in fatalErrorShutdownHandler
    \Sentry\captureLastError();
  File "[internal]", line 0

added scoped labels

Example request: https://www.minds.com/fs/v1/thumbnail/1026508015016828928/xlarge

Link to media page: https://www.minds.com/media/1026508015016828928

Offending section of code in /api/v1/thumbnail.php:

            try {
                $finfo = new \finfo(FILEINFO_MIME);
                $contentType = $finfo->buffer($contents) ?: 'image/jpeg';
            } catch (\Exception $e) {
                error_log($e);
                $contentType = 'image/jpeg';
            }

Issue appears related to finfo under PHP 7.3.

Running a referenced test script show the clear performance degradation between our 7.1 and 7.3 builds when calling finfo.

While I'm not able to reproduce the OOM condition in our sandbox, I am able to demonstrate the performance degradation.

PHP 7.1

/var/www/Minds/engine # ITERS=100 php test.php
Took: 0.094574928283691
/var/www/Minds/engine # ITERS=1000 php test.php
Took: 0.91444492340088

PHP 7.3

/var/www/Minds/engine # ITERS=100 php test.php
Took: 0.42086601257324
/var/www/Minds/engine # ITERS=1000 php test.php
Took: 4.215017080307

Likely this is a good place to start looking for the OOM issue.

The referenced performance relates to finfo::file() whereas we are using finfo::buffer().

Using finfo::buffer() does not cause a slow down.

Continuing by instrumenting some mock code to show memory usage before and after finfo::buffer()

There is also the possibility that something prior to this call is using too much memory and this final call tips it over the limit where it would not have done previously.

The use of $finfo->buffer() to get the mime type has an impact on the memory overhead of the script. The difference between PHP 7.1 and PHP 7.3 is minimal.

I suspect the real issue is that the entire image file is read into memory from S3 and then passed around before being written to the HTTP response.

After a couple of requests these values are stabilised:

finfo usage: 3585240
after usage: 3782552
total usage: 3782552

finfo is the memory delta just before and after the finfo calls. total is the delta from the start of request to the end after is the memory usage at the end of the script

finfo call accounts for most of the total memory usage of the script though is probably just when the memory is assigned for the file.

With and S3 backend, we already set the mime type to the metadata using the same finfo call when the file was written so we shouldn't need to repeat this process again.

Looking to see if we can extend the S3 calls to fetch the metadata...

Sentry issue: ENGINE-HF