Original C++ version from Qingjun Wei - https://github.com/weiqj/pc12306

This is a Java port. I'm interested in finding out how well the architecture works on Java platforms.

"12306" refers to Chinese Railroad website - http://www.12306.cn

There are a lot of trains; each train has a lot of seats; each train runs on a fixed path with multiple stops.

A customer wants to ride from station A to station B; this may require multiple connecting trains and multiple tickets; each ticket is for a fixed seat on a single train.

The question here is how to atomically reserve these tickets. The solution we are exploring here processes all reservation requests sequentially on a single thread. This is a more pleasant programming model when dealing with contention on multiple resources. However, is a single computer powerful enough to handle the tremendous request load on Chinese New Year?

We aren't going to write a complete system; we only want to study performance matrix on the key component of the system. For that purpose, we assume that, the system has constant number of resources -- there are 5000 trains; each train has 3000 seats; each train has 11 stops (i.e. 10 segments).

TrainTicketMap handles query and reservation on a single train. Given a train, and departure-arrival stations, it answers whether a seat is available; and whether it can be reserved (on a followup request). All information are kept in in-memory data structure.

An online ticket purchase is a complex action; fortunately many parts of the action can be executed in parallel. The most difficult part to parallelize is seat reservation. Our solution is to extract that part (in bare minimum) and execute it sequentially. The correctness of the solution is much simpler to prove; and the performance of the solution is very stable and quantifiable.

To handle concurrent interactions between reservation processor and other parts of the system, we use LMAX Disruptor (https://lmax-exchange.github.io/disruptor/), which is designed exactly for this type of application.

See class TestLmax. There is one event publisher, which queues reservation requests. All requests are processed sequentially by consumer C1, which only invokes TrainTicketMap. C1 is really fast. We also throw in C2 and C3 which depend on C1; they are no-op now, but they can be used for transaction logging, response writing, etc.

P1 -->  <request queue>  -->  C1 -----> C2
                                `-----> C3

P1, C1, C2, C3 all run on their own threads; there will be costs of data handover between them, but not much.

TestLmax can easily achieve 1M req/s on a commodity PC. This indicates that the architecture is feasible on Java platforms to handle real world workloads.

If we are building such a high throughput system, every CPU cycle counts; shouldn't we use something more metal instead, like C++?

It depends on economics. Java is cheaper in a lot of senses. Given the budget, time-to-market, and other constraints, you may not be able to afford to find a competent C++ guy that can do the job. But Java programmers that can write something like TestLmax are dime a dozen:)

If Java is adequate to meet the performance requirement of an application, we can turn the question around and ask, Why C++?