Replies: 4 comments 6 replies
-
|
What session driver / database driver are you using? How many CPU cores does your server have? Are you setting
It's still interesting though that you are seeing such a difference for big queries. I wonder what the bottleneck here could be, since at that point you're mainly benchmarking your database and PHP/eloquent. I'll try and see if I can reproduce this. |
Beta Was this translation helpful? Give feedback.
-
|
So I managed to reproduce your benchmarks and they reveal something interesting. Not necessarily about FrankenPHP, but more about MySQL (which is the bottleneck here). When reading 1000 Eloquent models per request, if I limit FrankenPHP to just 3 workers (3 CPU cores), it will handle the same amount or even more requests, which indicates that MySQL is at its limits. Paradoxically, MySQL will then be able to handle more requests if the requesting server is slower, aka if there are less connections. When using the This is kinda surprising since I expected MySQL to only be slower on writes, not on reads. But apparently PostgresSQL outperforms MySQL on big reads significantly? (or the I can give you the full breakdown + comparisons + recommendations if you share the Docker image you are using (I'm not sure how to fully set up Apache with Laravel). |
Beta Was this translation helpful? Give feedback.
-
|
I'd also ask if you are using pdo-mysql or mysqli driver. pdo-mysql, in it's default configuration is pretty terrible (all mysql driver development is on mysqli) and insecure. |
Beta Was this translation helpful? Give feedback.
-
|
Laravel uses On the other hand, worker mode is especially well suited to work with PostgresSQL since it keeps the connection open between requests. PgSQL will allocate a certain amount of memory for each connected client, which is slow on first connection but makes it more efficient afterwards. |
Beta Was this translation helpful? Give feedback.
-
PDO tries to provide a generic interface for all databases, so many drivers (such as mysql) cannot take full advantage of the database itself. Mysqli is a specific interface for MySQL that takes full advantage of the database and its features. There really should be a specific one for Postgres and other popular databases as well. A generic interface like PDO is great for some applications, but when you need the full power of the database, it is not enough. This is why frameworks like WordPress explicitly use MySQLi and not PDO. Edit to add: the PDO implementation is VERY insecure for Mysql unless you specifically change some settings. The default settings are absolutely bad on so many levels... it's a known issue (somewhere) but won't be changed for compatibility reasons. Postgres default settings are not bad though. |
Beta Was this translation helpful? Give feedback.
-
|
So, while looking deeper into this, I discovered something interesting... This is just from reading the C code though, so it would be good to look over an actual service and see how many connections are actually being utilized by the workers. (there is actually a To get around this, you can disable |
Beta Was this translation helpful? Give feedback.
-
|
Hmm I'll look into the amount of connections. It's weird though that this doesn't impact small queries |
Beta Was this translation helpful? Give feedback.
-
|
Hmm that doesn't seem to be it. Or at least I'm seeing 40 connection from 40 workers (and 2 from PHPMyAdmin) |
Beta Was this translation helpful? Give feedback.
-
|
Seems this was a red herring. The actual issue is the resource leak fixed in #1350 |
Beta Was this translation helpful? Give feedback.
-
|
@Macek007 do you see the same behavior when setting |
Beta Was this translation helpful? Give feedback.
-
What happened?
I’ve been experimenting with Laravel Octane + FrankenPHP. Initially, the benchmarks were quite promising, but after real-world usage, I discovered a few bottlenecks. Here’s my experience.
I have two very similar Dockerfiles: one is based on the php:8.3.13-apache-bookworm image, and the other on dunglas/frankenphp:php8.13.3-bookworm. Both use the same
php.iniconfiguration. The FrankenPHP container is started with:I ran benchmarks on an API for an ERP system that involves numerous database queries, some of which return large datasets.
Response on Cached Requests
The first test was on requests that use cache (using spatie/laravel-responsecache). All tests were conducted with the command:
The results were similar when testing in the browser.
PHP:
FrankenPHP:
The results were excellent, showing around a 150% improvement with FrankenPHP.
Response on Requests Returning Simple Data
For this test, I used requests that return simple data, such as enums or lookup tables. Typically, these involve only a single query.
PHP:
FrankenPHP:
Again, the results were impressive, with FrankenPHP showing about a 122% improvement.
Requests with Large Datasets
Next, I tested requests that return a dataset of 1,000 rows using Eloquent models with 15+ relations. I confirmed there were no N+1 query issues, and the database queries themselves took around a few hundred milliseconds (up to 500ms). However, the remaining time was spent in PHP runtime. After some xdebug tracing, I found that a significant amount of time was spent on model hydration.
PHP:
FrankenPHP:
In this scenario, FrankenPHP was about 20% slower.
I’ve spent many hours trying to eliminate various factors without much success. I tried adjusting
php.inisettings, aggressive OPCache and JIT conf, testing on both a local machine (MacBook + Docker) and a server environment (Ryzen + Debian + Kubernetes).Does anyone have any ideas? Has anyone experienced something similar?
Build Type
Docker (Debian Bookworm)
Worker Mode
Yes
Operating System
GNU/Linux
CPU Architecture
x86_64
PHP configuration
Relevant log output
No response
Beta Was this translation helpful? Give feedback.
All reactions