A few years ago, an Internet finance brother told me that with the rapid development of mobile Internet, mobile payment has high concurrency and high surge performance requirements, and the performance of the classic IOE architecture is insufficient, to force the financial technology architecture to evolve into a distributed architecture, multiple nodes need to be scaled out.

At that time, I am doubted. I have seen the performance data of some city commercial banks and joint-stock banks with trillions of assets, and the IOPS of the whole core system (including mobile payment) is only 2~30000, dahang has seen 100,000 yuan, and its IO performance is very surplus. There is no performance shortage as he said.

In the past two years, I have visited many customers, roughly asked them, and learned some real performance data for your reference. Welcome to make bricks.

Include the core transaction peak performance of mobile payment is about 2 to 30000 TPS for the TOP big line and 10,000 for the bottom big line., the most powerful line of joint-stock personal finance is 1.x million TPS, the head joint-stock banks generally range from 2000 to 5000TPS, large-scale provincial city commercial banks rural credit business peak value of 1000TPS, more city commercial banks only a few hundred TPS, the transaction performance of small and medium-sized banks is not as high as imagined..

However, the current computing power is actually very strong. A set of test data is provided for reference. The two-way Qipeng server has 48 cores and 1 master and 2 slave servers. The transaction scenario performance can be measured to 18000 TPS. In fact, if you think about it carefully, only WeChat payment and Alipay have a monthly or even daily activity of nearly 1 billion levels. Indeed, everyone basically uses these two payment eldest brothers to trade every day, the transaction system has hundreds of thousands of TPS, especially for Double 11, the transaction peak may be higher.

However, there are more than 200 banks in our country (6 big banks, 12 joint-stock banks, more than 130 city commercial banks, 30 provincial agricultural credit cooperatives, and nearly 20 agricultural commercial banks that independently build core systems, there are also internet private banks, etc.), more than 200 mobile payment flows are diverted, and the transaction performance is not so high. Now the Double 11 has been extended to one month, and the economic situation is not optimistic, people's consumption desire and demand are degraded, and transaction performance continues to decline. Is performance requirement still the core requirement of distributed architecture transformation?

On the contrary, many customers have reported that the distributed architecture has high latency and many cross-node transaction devices, which naturally consumes a lot of time. Of course, the complexity is also very high, and the performance positioning is very difficult, most of the time, we can't find out who the problem is, and we have to roll out the flowers.

I have communicated with many customers and found that, in fact, the core transaction volume is really not high, most of the core database capacity is concentrated in hundreds of GB ~ 10TB, and even many customers think that more than 90% of the system centralized architecture is sufficient, only systems that are less than 10% or even less require distributed transformation, or more data shards are enough. Similar to a team, if there are more people, they will be divided into more groups. Currently, some users can see hundreds or even thousands of database parts. Is the performance not enough or the data capacity not enough?

Database experts have always said that the performance of single shard is awesome. I have heard of a G database 800000 TpmC, so why do I need so many shards? So many shards bring about a sharp increase in equipment, a large number of transaction node chains, significantly increased latency, extremely high O & M complexity, and simple architecture design. How can it be more and more complex and can it hold?

It is wrong to go to the left and right of everything. If database experts plan so many shards, will they do it for their own benefit: If there are more shards, the database software license will be more, so that they can be sold more. In addition, it is often seen that the I/O capability of the local disk of the server is not good. Database experts say that the CPU needs to be expanded, that is, the server and database license of one master and multiple slave needs to be expanded, storage I/O capability and computing power are two completely different things. Why not analyze the core reason why I/O capability and time delay are not working?

Recently, it has also been seen that the central financial Working Committee has also clearly pointed out that it is very, very strict: unrealistic disorderly innovation and fake innovation will bring huge risks and losses to financial institutions. Business and technology are the same, which is worth pondering. Eventually, they will return to the essence of business and seek truth from facts.

Source: Bi xunshuo