FaunaDB has a per-database replicated transaction log. Once a transaction has been globally committed to the log, it is applied to each local partition that covers part of the transaction. By this point, the transaction's order with respect to others in a database and results are determined. While writes require global coordination to commit, reads across partitions are coordinated via a snapshot time per query, which guarantees correctness.

In short, writes require a global round-trip through the transaction pipeline; reads are local and low latency.

This is a very good answer. So if I understand you correctly (please correct if I do not), atomicity is handled on a per connection basis (writes cannot be distrubted). And there may be high latency in distributing a transaction, but read consistency is guaranteed by timestamp (equivalent to versioning).

Is this correct?

EL

One thing that makes this easier is that FaunaDB does not support session transactions, rather you must express your transaction logic as a single Fauna query, which is executed atomically. Transactions can still involve arbitrary keys, however.

And yes, for reads, by default the coordinating node chooses a timestamp and uses that to query all involved data partitions. Each partition will respond with the requested data as of at that timestamp, or will delay responding until it has caught up.

One nice thing about this approach is that any chosen timestamp is enough to provide a consistent snapshot view of the dataset at that time. This ends up being useful for bulk or incremental reads, where a longer running process needs a stable view of the dataset.

Without session transaction, how does the application performs transactional read modify write?

Okay, Last question because I'm in Asia and it's 7:30 and I am tired. (But yea, I'm a Midwestern computer scientist who just happens to be in Asia ;-)

Question #2 -- Two, or Three, or More transactions occur simultaneously that want to change data D. each of these transactions send out transaction logs that contradict each other. What happens?

EL

snapshot time as in wall clock time? What is the time source?

It's variable, and the client can provide it. By default, we use the greater of wall clock time or the highest timestamp transaction timestamp the node serving the query has seen.

In that case, your database loses data and is definitely not strongly consistent. Why do you claim it is?

To be clear, read-write transactions are applied in a consistent order derived from the transaction log. Read-only transactions can rely on the fact that read-write transactions are totally ordered, and execute without global coordination while still providing a guarantee of serializability. Wall clock time is used as a suggestion only, in order for the database's logical timestamps to track as close to real time as possible. We will have more information about how Fauna meets its consistency guarantees in a future blog post.

you can't have a read transaction that's serializable that doesn't go through global coordination unless all of the data is guaranteed to only ever be local. Looking forward to the blog post.

looking forward to Jepsen...

Looks like FaunaDB uses Raft[1], so I'd expect that data is sharded into multiple consensus groups, like Spanner or Megastore. That would mean consistency on a single shard/consensus group is basically just dependent on reading from and writing to the Raft leader.

1: https://news.ycombinator.com/item?id=13645876