A High-Performance Logger with PHP

Recently, at Leaseweb, another successful Hackathon came to an end. There was a lot of fun, a lot of coding, a lot of coffee and a lot of cool ideas that arose during these 2 days. In this blog post we want to share one of these ideas that made us proud and that was really fun to work on.

1. The motivation

As Leaseweb, we strive to know our customers better so that we can actively empower them. For that, we need to start logging events that have value for the business. So basically we want to implement a simple technical service (let’s call it a ‘Logging Service’) that accepts any kind of event with some rich data (in other words a payload), and then logs it, without interfering with the execution of the client.

Our Logging service would, on each request, return an ”OK” so that the client can continue with its own execution but still keep on logging events asynchronously.

2. Our tech stack and our tech choice for POC

Typically at Leaseweb, we have a pretty standardized technology stack that revolves around PHP+Apache or PHP+Nginx. If a server or an application is built using this kind of stack, we are bound to typical synchronous execution. This client sends a request to our application (in this case our Logging Service) and the client needs to wait until our application sends back a response after it finishes all the tasks. This is not an ideal scenario. We need a service that runs asynchronously, a service that receives the request, says ”OK” so that the client can continue with his execution, and then does its job.

In the market, there are several tools that would enable us to do this such as languages and tools like NodeJs, Golang or even some message queuing services that could be wired into our PHP stack. But as we are PHP enthusiasts at Leaseweb, we want to use our language of choice without adding dependencies. That is how we discovered ReactPHP (no relation to the front-end framework React). ReactPHP is a pure PHP library that allows the developer to do some cool reactive programming in PHP by running the code on an endless event loop :).

3. Implementation of the POC

With some ReactPHP libraries, we can handle HTTP requests in PHP itself. We no longer need a web server that handles the requests and creates PHP processes for us. We can just create a pure PHP process that handles everything for us and if we want to fully use the hardware power from our machine, we can create several PHP processes and put them behind a load balancer.

After choosing our technology we started playing around and implementing our idea. In order to make sure that asynchronous PHP would solve the majority of our concerns, we started benchmarking.

First we needed to script something to benchmark. Therefore we implemented 3 different scenarios:

  1. The all-time-favorite: An endpoint that prints Hello World.
  2. An API endpoint that calls a 3rd party API that takes 2 seconds to reply and proxies its response.
  3. An API-endpoint that accepts a payload with POST and logs it via HTTP to Elasticsearch (This is what we really want).

We implemented these 3 scenarios in two different stacks:

  • Stack A
    Traditional PHP+FPM+Nginx
  • Stack X
    4 PHP processes running on a loop with ReactPHP behind an Nginx load-balancer. The reason why we chose 4 processes was solely because this number looks good :). There are some theories in which suggestions are made regarding the number of processes that should run on a machine when using this approach. We will not go into further detail on this in this article. Note that in both Stack A and Stack X we used the exact same specs for the hardware server. On both, we had a CPU with 8 cores.

Then we ran some stress tests with Locust:

3.1 1st Benchmark – Hello world!

For the first benchmark, we just wanted to see how the implementation of the Scenario 1 would behave in both of our stacks.

Hello World with Stack X
Figure 1: Hello World with Stack X
Hello World with Stack A
Figure 2: Hello World with Stack A

We can see that both of the stacks perform very similar. The reason for this is that the computation needed to print a ”Hello World!” is minimal, therefore both of our stacks can answer a high amount of requests in a reliable way.

The real power of asynchronous code comes when we need to deal with Input/Output (I/O (reading from a DB, API, Filesystem, etc) because these operations are time-consuming (see [zhuk:2026:event-driven-with-php]). I/O is slow and CPU computation is fast. By going with an asynchronous approach, our program can execute other computations while waiting for I/O.

It is time to try this theory with the next benchmark:

3.2 2nd Benchmark – Response from a 3rd-party API

Following what we described in the previous section, the power of asynchronous code comes when we deal with input and output. So we were curious to find out how the APIs would behave if they need to call a 3rd party API that takes 2 seconds to respond and then it forwards its response.

Let’s run the tests:

Response from a 3rd-party API with Stack X
Figure 3: Response from a 3rd-party API with Stack X
Response from a 3rd-party API with Stack A
Figure 4: Response from a 3rd-party API with Stack A

With the benchmarks illustrated in figure 3 and 4, it’s encouraging to see we already have very interesting results. We ran a stress test where we gradually spawn 100 concurrent users that send a request per second, and we can easily see that the more the concurrent users grow, the less responsive Stack A becomes. Stack A is achieving an average response time of 40 seconds, while Stack X maintains an average response time of 2 seconds (which is the time that the 3rd-party API takes to respond).

With Stack A, each request made creates a process that will stay idle until the 3rd-party API responds. The implication is that, at some point in time, we will have hundreds of idle processes waiting for a reply. This will cause an immense overload on our machine’s resources.

Stack X performs exceedingly well. This is because the same processes that wait for the reply from the 3rd-party API will continue doing other work during their execution, for example, handling other HTTP requests and incoming responses from the 3rd-party API. With this, we can achieve much more efficiency in our stack.

After observing these results we wanted to push it a bit harder – we wanted to see whether we could break Stack A entirely. So we decided to run the same stress test for this scenario but this time with 1000 concurrent users.

Response from a 3rd-party API with Stack X with 1000 users
Figure 5: Response from a 3rd-party API with Stack X with 1000 users
Response from a 3rd-party API with Stack A with 1000 users
Figure 6: Response from a 3rd-party API with Stack A with 1000 users

We did it! We can see that at some point Stack A is unable to handle the requests anymore so it stops responding completely after reaching an average response time of 60 seconds. Stack X remains perfectly smooth with an average response time of 2 seconds :).

3.3 3rd Benchmark

It was indeed fun trying to see how the stacks behave with the previous scenarios but we want to see how it behaves in a real-world scenario. Next, we wanted our API to accept a JSON payload via an HTTP post and log it to an Elasticsearch cluster via HTTP to keep it simple (Scenario 3).

How the stacks work in a nutshell:

  • Stack X receives an HTTP Post request with the payload, sends a response to the client saying OK and then logs it to Elasticsearch (asynchronously).
  • Stack A receives an HTTP Post request with the payload, logs it to Elasticsearch and sends a response to the client saying OK.

Let’s bombard it with Locust again and why not with 1000 of concurrent connections right away:

Logging payloads with Stack X
Figure 7: Logging payloads with Stack X
Logging payloads with Stack A
Figure 8: Logging payloads with Stack A

We can see that we can achieve a pretty reliable and high-performance logger. And this only with pure PHP code!!

Because our intention was always to push the limits, we chose this benchmark with 1000 concurrent users being spawned gradually. Stack A at some point stops handling the requests, while the Stack X always keeps a pretty good response time, around 10ms.

4. What can we use it for now?

With this experiment, we pretty much built a central logging service!! Coming back to our main motivation, we can use this to log whatever we want, and we want to start logging meaningful domain events, from any application within our system with a simple non-blocking HTTP Request. For example, if we start logging meaningful events, we can get to know our customers better. If we log all of this into Elasticsearch we can also start making cool graphs from it. For example:

Graph with Business Events
Figure 9: Graph with Business Events


Graph with End-user Actions
Figure 10: Graph with End-user Actions

Since this approach is so highly responsive, we can even start using it to log anything and maybe everything, where our logging exists in a central endpoint. System monitoring, near-real-time-analytics, domain events, trends, etc, etc. And all of this with pure PHP :).

5. Cons of the approach and future work

When using ReactPHP there are some important considerations, which in some scenarios can be seen as cons, that usually they are not applicable to projects that follow an architecture similar to Stack A.

  • ReactPHP uses reactive/event-driven programming which is a paradigm that might have a big learning curve.
  • Long-running PHP processes can lead to memory leak and in case of failures, they could affect all the current connections to the server
  • These processes need to be constantly carefully monitored in order to avoid and predict fatal failures.
  • The usage of blocking functions (functions that block the execution of the code) will massively affect the performance for all the connections to the server.

Also, some extra work on the ”Operational / Infra side” is needed to make sure that we continuously check on the process’ health and if something goes wrong, create new ones automatically. We also need to work on the way we deploy the code. We need to make sure that we restart a process sequentially when deploying new code so that our service can finish serving the requests that it has queued at that time.

6. Conclusion

Pushing the limits of our preferred technology is one of the most fun things to do. PHP is not a usual choice if there is the need for an high-performing application, but thanks to a lot of great work of the community around, solutions like ReactPHP starts to emerge. That opens a new path to discover new programming paradigms, it introduces different mind-sets on how to approach a problem and it challenges the knowledge we have regarding the technology.

Challenging what we already know, is one of the most interesting things that we can do because it takes us out of our comfort zone and helps us to become more mature. It is really fun and it makes us realise that we can never know everything.

We would like to thank and acknowledge everybody in the communities around the tools we used in this fun experiment 🙂

Some useful links:

by Joao Castro and Elrich Faul