Transactions in NoSQL?

Question

I'm looking into NoSQL for scaling alternatives to a database. What do I do if I want transaction-based things that are sensitive to these kind of things?

Answer 1

Generally speaking, NoSQL solutions have lighter weight transactional semantics than relational databases, but still have facilities for atomic operations at some level.

Generally, the ones which do master-master replication provide less in the way of consistency, and more availability. So one should choose the right tool for the right problem.

Many offer transactions at the single document (or row etc.) level. For example with MongoDB there is atomicity at the single document - but documents can be fairly rich so this usually works pretty well -- more info here.

Answer 2

This is the closest answer I found which would apply to any NoSQL database. It's on a 2007 blog post from Adam Wiggins of Heroku.com:

The old example of using a database transaction to wrap the transfer of money from one bank account to another is total bull. The correct solution is to store a list of ledger events (transfers between accounts) and show the current balance as a sum of the ledger. If you’re programming in a functional language (or thinking that way), this is obvious.

From: http://adam.heroku.com/past/2007/12/17/a_world_without_sql/ (His website is great for ideas on scalability.)

I interpreted the above paragraph as:

1. Create a database for member accounts.
2. Create a messaging queue. Nickname it "ledger".
3. Add in background workers to fulfill each request in the queue.

More info. on queues/background workers: http://adam.heroku.com/past/2009/4/14/building_a_queuebacked_feed_reader_part_1/

The client (aka member or customer) follows these steps to take out money:

1. Submit a request to take out money.
2. Request is sent to server.
3. Server places it in a queue. The message is: "Take out $5,000."
4. Client is shown: "Please wait as request is being fulfilled..."
5. Client machines polls server every 2 seconds asking, "Has the request been fulfilled?"
6. On server, background workers are fulfilling previous requests from other members in first-in/first-out fashion. Eventually, they get to your client's request to take out money.
7. Once request has been fulfilled, client is given a message with their new balance.

You can use Heroku.com to create a small mock-up quickly if you are comfortable with Node.js or Ruby/Rack.

The general idea seems pretty easy and much better than using transactions baked into the database that make it super-hard to scale.

Disclaimer: I haven't implemented this in any way yet. I read about these things for curiosity even though I have no practical need for them. Yes, @gbn is right that a RDBMS with transactions would probably be sufficient for the needs of Timmy and me. Nevertheless, it would be fun to see how far you can take NoSQL databases with open-source tools and a how-to website called, "A Tornado of Razorblades".

Answer 3

NoSQL covers a diverse set of tools and services, including key-value-, document, graph and wide-column stores. They usually try improving scalability of the data store, usually by distributing data processing. Transactions require ACID properties of how DBs perform user operations. ACID restricts how scalability can be improved: most of the NoSQL tools relax consistency criteria of the operatioins to get fault-tolerance and availability for scaling, which makes implementing ACID transactions very hard.

A commonly cited theoretical reasoning of distributed data stores is the CAP theorem: consistency, availability and partition tolerance cannot be achieved at the same time. SQL, NoSQL and NewSQL tools can be classified according to what they give up; a good figure might be found here.

A new, weaker set of requirements replacing ACID is BASE ("basically avalilable, soft state, eventual consistency"). However, eventually consistent tools ("eventually all accesses to an item will return the last updated value") are hardly acceptable in transactional applications like banking. Here a good idea would be to use in-memory, column-oriented and distributed SQL/ACID databases, for example VoltDB; I suggest looking at these "NewSQL" solutions.