Sunsetting the Techempower Framework Benchmarks

I've contributed a few optimisations to some implementations in these benchmarks, but as I read the code of many other implementations (and some frameworks) I lost most of the trust I had in these benchmarks.

I knew that once a benchmark is famous, people start optimising for it or even gaming it, but I didn't realise how much it made the benchmarks meaningless. Some frameworks were just not production ready, or had shortcuts made just for a benchmark case. Some implementations were supposed to use a framework, but the code was skewed in an unrealistic way. And sometimes the algorithm was different (IIRC, some implementation converted the "multiple sql updates" requirements into a single complex update using CASE).

I would ignore the results for most cases, especially the emerging software, but at least the benchmarks suggested orders of magnitudes in a few cases. I.e. the speed of JSON serialization in different languages, or that PHP Laravel was more or less twice slower than PHP Symfony which could be twice slower than Rails.

I always liked these benchmarks, I've been following them since the earliest rounds.

One thing to note is how much things have improved over that time. Numbers that used to top the benchmarks would now be seen as "slow" compared to the top performers.

The other useful thing about these benchmarks is being able to easily justify the use of out of the box ASP.NET Core.

For many languages, the best performers are custom frameworks and presumably have trade-offs versus better known frameworks.

For C# the best performing framework (at least for "fortunes") is aspnet-core.

That side-steps a lot of conversations that might otherwise drag us into "Should we use framework X or Y" and waste time evaluating things.

Are the benchmarks gamed? Yes of course, the code might not even be recognisable as Asp.NET Core to me, but that doesn't really matter if I can use it as an authoritative source to fend off the "rewrite in go" crowd, and it doesn't matter that it is gamed, because the real-world load is many orders of magnitude less than these benchmarks demonstrate is possible.

I really liked these benchmarks, and would check in with them from time to time.

No benchmark is perfect, but these ones cover such a wide variety of different languages and frameworks, it's a good resource for getting a rough idea of the kind of performance that a given stack is capable of.

I don't know much about TechEmpower the company, it seems to be a small consultancy, maintaining this project probably takes non insignificant resources from them.

The end of the project seems kind of unceremonious, but they don't owe anything to anyone.

Hopefully an active fork emerges.

Sad to see this. I had so much fun implementing a http server (called httpbeast) from scratch to get as far up these benchmarks as possible.

I do agree with others here that it was possible to game them, but it still gave a good indication of the performance bracket a language was in (and you could check if interpreted languages were cheating via FFI pretty easily).

Feels like the end of an era.

This text lacks information about why it is being sunset.

well done to the techempower team for the work done.

though the benchmarks were not exactly 100% accurate - they gave good estimates on how different frameworks / perform in handling web tasks.

they also helped people move to simpler / lighter web frameworks that are more performant and kind helped usher in the typical 'Sinatra/express' handlers for most web frameworks e.g .net core

they also showed the performance hit of ORMs vs RAW. so yeah well done.

Engineering has kind of moved on in a weird way from web frameworks. Now AI just writes document.getElementById('longVariableName') javascript and straight SQL without complaining at all. The abstraction isn't as important as it used to be because AI doesn't mind typing.

Would you know any alternative?

[dead]

[dead]

[dead]