In the term of hiding the temp table, I think we can provide a wrapper class with the database driver. And I'm not sure if we are looking for something like MVCC ( Multi-Version Concurrency Control ) ?
2018-04-02 22:54 GMT+08:00 Willem Jiang <willem.ji...@gmail.com>: > Chris just gave a talk[1] recently about the countermeasure for data > anomalies. > It made us to think if we can find a better way to resolve the isolation > issues. > > It looks like the TCC pattern can prevent the uncommitted states to be saw > by adding the try phases. > And this is what current saga solution is missed. With help of temp table > to store the status of the try operation, we don't need the coordinator > handle much unnormal situation. It could be better if we don't let the > user know about the temp table. But the challenge is how can we hide the > temp table to the user. > > [1] > http://microservices.io/microservices/general/2018/03/ > 22/microxchg-sagas.html > > > > Willem Jiang > > Blog: http://willemjiang.blogspot.com (English) > http://jnn.iteye.com (Chinese) > Twitter: willemjiang > Weibo: 姜宁willem > > On Mon, Apr 2, 2018 at 5:57 PM, Eric Lee <eric.lee....@gmail.com> wrote: > > > Hi, all > > > > As we have discussed the ACID guarantees Saga provides in the previous > > thread[1], it turns out Saga does not provide isolation guarantee. To > > improve user experience, the business logic using Saga may need to > reorder > > to make sure the user-sensitive sub-transaction is the last one to be > > executed. In sceanrios require full ACID support, the implementaion of > Saga > > may need to be compatible with the TCC[2] pattern with an extra try > phase. > > > > Take a transfer application as an example, it contains transfer in and > > transfer out service of two different databases. From the customer's > view, > > the transfer in and transfer out operation is an atomic operation which > > requires both to be executed or nothing executed. However, in the middle > of > > the overall transaction, e.g. the sub transaction of transfer out is done > > and the sub transaction of transfer in is not done yet, if a customer > > checkouts out his/her balance, it will become weird as the balance is not > > equal. The isolation is corrupted at this moment in Saga. > > In TCC, the isolation could be solved using either the reservation or > > compensation which depends on your bussiness logic. > > Reservation: In try phase, use a temporal table to store the credit and > > transaction context. In commit phase, reduce the balance in the account > and > > remove the temporal table. If anything goes wrong, it can execute the > > cancel method to remove the temporal table. In this way, if the global > > transaction fails, it will take no effect on the actual table. Besides, > > when a customer visits his/her balance, we can simply return the value in > > the actual table which is the original value before this transaction > > executed. > > Compensation: In try phase, use a temporal table to record the > compensated > > value and reduce the balance in the account. In commit phase, remove the > > temporal table. If anything goes wrong, it can execute the cancel method > > to recover the balance according to the temporal table and remove the > > temporal table afterward. In this way, when a customer visits his/her > > balance, we can do simple calculation of the value in actual table and > > temporal table to return the origianl value before the transaction > > executed. > > > > Within transaction ids in the table row, each create/update/delete > > operation is idempotent and it simplies a lot of work to make sure > > sub-transactions are idempotent. > > > > > > Any other ideas or suggestions on the isolation support in Saga are > > welcome. Thanks. > > > > > > [1] https://lists.apache.org/list.html?dev@servicecomb.apach > > e.org:lte=1M:a%20question%20about%20acid%20 > > [2] http://design.inf.usi.ch/sites/default/files/biblio/rest-tcc.pdf > > > > > > Best Regards! > > Eric Lee > > >
