Transactions without Transactions

If you're interested in other upcoming videos from Strange Loop, the full release schedule is here and all slides are here. If you want to be notified about Strange Loop announcements in the future, sign up for the mailing list.

wrt the Ticket Sales example.. If there are only 2 seats left and 2 separate processes try and reserve both at the same time, then you may have a situation where by each process is only able to move one of the tickets into the IN_CART status. Both processes effectively initiate a compensatory action which moves both tickets back to an AVAILABLE status.. You end up missing out on selling tickets for which you had 2 buyers for.. did I miss something ?

I totaly agree with the concept of using RDBMS transactions with caution and avoid using them to model business transactions to avoid coupling as was mentioned in the presentation.

However, I'm still not convinced that you need to model all the transactions at the application. In my view, I think you still need atomicity at the scope of an operation on single database. For example if you have 2 entities a parent and a child and you want the child to be deleted if the parent is deleted then you need transactions to do that otherwise if you model it in the application then you will increase the complexity of the application tremendously.

If software managed transactions are being introduced to in languages (like clojure) and frameworks to ease the development of concurrent operation then for sure transactions have their use in persistent data stores which are concurrent by nature.

The problem is that one has to implement all the detection and compensation code. That is not simple.

The ticket sales example isn't any more realistic than the bank transfer. Nobody would implement the whole sale in a single transaction.

The realistic parts are where multiple tickets need to transit state in unison. That is trivial with transactions (and very short lived), but quite a bit of code without transactions.

Also in the ticket example the target object (ticket) is essentially locked for an extended period. What if the target object needs to be updated by other transactions?

What if a part of the transaction is an increment (or similar)? On a failure how do you know if the increment was applied or not?

So summarising the slides:
- in distributed systems you can't rely on the transaction capabilities of a single database.
- but the real world problem you're dealing with still needs transactional semantics (e.g. conservation of money)
- so build the logic of a distributed transaction manager in your application code base.
- Use transactional asynchronous messaging to communicate among the nodes in your network.
- manage the transaction as a multi-step state machine, with identified compensatory actions should any transition in the state machine fail.

Sounds a lot like the problem that spawned distributed transaction monitors 20 years ago. The tech world really does go round in circles sometimes. Wonder if Tuxedo, Top End and their ilk will rise phoenix-like?

I did not have patience to seat through entire presentation. Because of simplistic view and examples provided.
Well, what happens if compensating transaction fails? What happens in highly concurrent environment?

Is Transactional Messaging currently supported by MongoDB?

The idea is the leasing - you put a transient lock to the data while you are working on it. In normal case you would move the data out of the transient state and into a stable state when you are done. Worst comes worst the data store will do the clean up of that piece of data when the lease expires.

You end up missing out on selling tickets for which you had 2 buyers for.. did I miss something ?