Hello, world!

Hello, world! In this blog, we’d like to share our ideas and experiences related to testing and debugging distributed systems.

Testing and debugging software is difficult. Especially, testing and debugging distributed systems is known to be very difficult. Why so difficult? We believe the difficulty comes from lacking good tools for the distributed systems specific bugs.

Testing, debugging, and verification techniques for removing bugs in software have a long history. Even programming is difficult task since the beginning of its history, significant effort of researchers and engineers is succeeding at establishing techniques for fighting against major important bugs. For example, modern programming languages tend to have their own GC mechanisms and they are very effective for removing memory leak bugs. In addition, tools like valgrind can help programmers to detect these bugs even software is written in C or C++. Many other techniques were also established for other types of bugs, so developing software seems to be becoming easier than ancient days.

However, times are changing. In these days, it is clear that importance of a new category of software, distributed systems, is rising. Although the idea of distributed systems is very old, its modern implementations, e.g. Apache Hadoop, are introducing significant benefit to today’s world. Distributed systems are essentially different from non-distributed systems. They combine multiple computers for highly availabile, highly durabile, and scalable performant systems. As a result, they can enable new sort of services called cloud computing and big data. Therefore everyone loves these systems.

Sadly, recent studies shows that bugs in the distributed systems are hard to detect and tend to introduce critical failures e.g. permanent data loss. Of course distributed systems share many types of bugs (e.g. memory leak, race conditions) with non-distributed systems, so existing debugging techniques are also effective for removing these bugs in the systems. However, distributed systems have their own types of bugs and the critical failures tend to be introduced by such bugs.

What are the distributed systems specific bugs? Studies of this area is in very early stage so there’s no mature categorization, but I can list some classes which can be seen in many systems:

  • distributed race conditions caused by interleaving of messages over network
  • incorrect handling of hardware e.g. disk failure, node failure, network partition
  • performance degrading, especially a case of losing scalability

Of course all of the above three classes are critical, but first and second ones are especially emergent because they are related to correctness of systems. Incorrect systems are not valuable even their performance is good.

Though programming methodologies are evolving, these bugs remain hard to be debugged because of lacking good tools. In succeeding posts, we will describe the difficulties of the debugging, possible candidates of solutions, and a tool we are working on.