In the LevelDB source code, there’s a function for getting a Comparator that seemed a bit strange when I first saw it. It looks like it constructs a singleton, but it’s slightly more complex. The complete code is as follows:

1
2
3
4
5

// util/comparator.cc
const Comparator* BytewiseComparator() {
  static NoDestructor<BytewiseComparatorImpl> singleton;
  return singleton.get();
}

Here, NoDestructor is a template class that, judging by its name, is used to prevent object destruction. Why prevent object destruction, and how is it achieved? This article will delve into these questions.

Anthropic has quietly released the Claude 3.5 model, which not only improves model performance but also supports Artifacts. After initial experience, I feel this is what the future of AI should look like, even more impressive than the previous GPT4-o.

Before Artifacts, if you wanted to use ChatGPT or other LLMs to help implement program functionality, you needed to first provide the functional requirements, then copy the AI’s code into your own environment to run. If it didn’t meet expectations, you had to ask again, modify the code, and run it again. This process needed to be repeated until satisfaction or giving up (some complex code is still not well written by AI currently).

Recently, I encountered a very strange service restart issue in our business, and the troubleshooting process was quite complex. This article will review the process. The problem involves multiple aspects such as C++ thread pools, integer overflow, exception catching, and blocking, which is quite interesting.

Next, I will organize the content of this article according to the problem investigation process, starting with an introduction to the problem background, then listing the initial troubleshooting ideas and methods for locating abnormal requests. Then, through code analysis and some simple use cases to reproduce the problem, we will unveil the mystery of the service restart.