> The biggest remaining fault, in the context of web apps (IMO) is the lack of transactions

While not ACID, "tension" documents are a good way to add a level of transactional support.

For example, if you want to update multiple documents, instead of modifying the original documents directly, you create a new document with instructions on how to update the original documents. Then, you roll through the tension documents and apply their changes. Any failures can be dealt with on subsequent passes, only removing the document from the scan once it has been successfully applied.

It is an eventually consistent pattern, but that is a tradeoff you have already accepted by choosing MongoDB in the first place, so it ends up working well for a lot of transaction use cases.

Its actually possible to be fully consistent if you include the "tension" documents when fetching the main docs.

For example, in the canonical transaction example of Alice giving Bob $5 you could insert a document in the transactions collection that says Alice gave Bob $5. When you go to fetch Alice's pending balance you fetch her current document, then you fetch all pending transactions than she is involved in. Same for Bob and all other users. Then each night, you pull that days transactions and apply them to each user and update an "as of" field to make sure no transaction can ever be applied twice.

Now you could argue that this isn't fully consistent because it is possible for Alice to simultaneously give $5 to both Bob and Charlie even if she only has $7 in her account. However she will then have a pending balance of $-3 so it immediately reflects her current situation. This is similar to the real world example of overdrawing you checking account. If you wanted to, you could void one or both of those transactions if you never want to allow a user to go negative.

The system you describe, recording events and replaying events to get the actual state of the system, usually assumes that you can only write one event at a time (writes need to be super cheap). If you can write multiple events in parallel, you won't have isolation and you may read tension documents that will be voided (in your example).

You could use optimistic concurrency control to ensure that only one tension document is written at a time while ensuring that your constraints are respected:

  current_id = atomic increment global state_id
  create tension document (but with active bit set to false)
  check constraints
  atomic:
    if global state_id = current_id
    set active bit to true
  if active bit is false:
    delete tension document and report error

This would provide isolation (you don't see tension documents that are not yet validated) and consistency (your balance is respected).

The atomic operation in RDBMS is usually implemented in one very simple and fast SQL UPDATE query. I believe mongo must provide something similar too.

Although my bank would allow me to overdraw from my checking account, I have a higher margin/tolerance than my brother, so they must perform some constraints check ;-)

This is roughly what MVCC does under the covers. Of course, you don't have to implement it, the database just does it for you, and it's not eventually consistent.

That's a really clever idea. I'll have to give that a shot - I wonder how hard it'd be to wrap up that pattern for use in the popular ODMs?