Jd.com 2: how to prevent oversold in the deduction process of commodity inventory?

Click on “ End of life ”, Focus on , Official account

Daily technical dry goods , First time delivery ！

In the process of commodity purchase , Inventory deduction process , The general operation is as follows ：

• select According to the goods id Query the inventory of goods .

• According to the number of orders , Calculate whether the inventory is sufficient , If the inventory is insufficient, an exception of insufficient inventory will be thrown , If there is enough stock , Then subtract the deducted inventory to get the latest inventory residual value .

• set Set the latest inventory remaining value .

The pseudo code of the above process is as follows ：

//  According to the goods id Get the remaining inventory of goods 
select stock_remaing from stock_table where id=${goodsId};
 
//  Operating inventory 
//  Compare inventory 
if(stock_remaing <quantity){
   //  Throw the exception of insufficient inventory 
}
else{
//  Inventory value after deduction 
int new_stock=stock_remaing - quantity;
}
 
//  According to the goods id Set the calculated inventory 
update stock_table set stock_remaing =${new_stock} id=${goodsId};

1、 Concurrent modification of database storage oversold

If the isolation level of database transactions is not serialization （serializable）, According to the characteristics of transactions , During concurrent modification , Write overwrite may occur .

hypothesis , Remaining inventory of goods stock_remaing by 100, Customer A Place an order 20, Customer B Place an order 30, When deducting inventory concurrently , There may be oversold . If the customer A And the customer B At the same time, the remaining inventory is 100, The value submitted after the transaction will overwrite the value submitted by the previous customer , It is possible that the remaining inventory is 80 perhaps 70. The process is as follows ：

2、 Lock update Repository

In order to control transactions , Prevent write overwrite , You'll think of using select for update The way , Lock the inventory of this commodity , Then do the rest .

The process is as follows ：

above , Use pessimistic locking , In distributed services , If concurrency is high , Stock deduction is a serial operation , Efficiency is very low .

3、 Update with optimistic lock

At the time of the update , Use （CAS+ Version number update ）+ Retry conditions （ Number of retries or retry time limit ） Update inventory in an optimistic way . here , If , Customer A And the customer B At the same time, the remaining inventory is read 100, At the time of the update , One operation will fail . The process is as follows ：

This method can greatly improve the concurrency , It can also ensure the consistency of data ; It is controlled by the conditions of retry times and retry time , It can prevent the database pressure caused by too many retries .

4、 Can it be executed by direct decrement ？

When deducting inventory , Some people propose not to implement select, Calculation ,set Three stage operation , The way of direct deduction , And make a judgment when the deduction is less than zero . The pseudocode is as follows ：

update stock_table set remaing_stock=remaing_stock-${quantity}
where id = goods id
and remaing_stock>${quantity};

In a distributed service call , Because the network is abnormal , Get server exception , Maybe when the microservice is invoked , There is a service retry . for example , Gateway timeout for scenario , Service retry mechanism . here , This method does not satisfy idempotency , And there are many deductions . for example , When the same user deducts inventory , Service retry , In extreme cases , The user has performed inventory deduction for multiple times , Then there is oversold .

5、 have access to redis Do you want to deduct inventory ？

Because I haven't studied redis Source code , For this way, I refer to Daniel's reply , The answer is that you can use redis Transaction deduction balance of , But in CAS Mechanism ratio mysql No advantage , High performance is due to its memory storage , The side effect is the risk of data loss .

PS： Prevent missing this article , You can like it , It's easy to browse and search