Implement group commit into journal

[marvin-j97]

No description provided.

[jcdickinson]

Thinking about contributing to this, what's the plan? I assume something along the lines of:

1. Create a transaction ID
2. Mark journal entries with the transaction ID
3. Create transaction completed journal event with the transaction ID

That should allow multiple active transactions to interleave. The main question is binary compatibility in the journal, is that a concern?

[marvin-j97]

@jcdickinson The main idea is to amortize the fsync costs in multithreaded workloads. Right now it's effectively singlethreaded:

t1:insert
t1:sync
t2:insert
t2:sync
t3:insert
t3:sync

The rough idea is the same as what RocksDB is doing:

As most other systems relying on logs, RocksDB supports group commit to improve WAL writing throughput, as well as write amplification. RocksDB's group commit is implemented in a naive way: when different threads are writing to the same DB at the same time, all outstanding writes that qualify to be combined will be combined together and write to WAL once, with one fsync. In this way, more writes can be completed by the same number of I/Os.

https://github.com/facebook/rocksdb/wiki/WAL-Performance

So it becomes:

t1:insert
t2:insert
t3:insert
t1:sync

One problem is that the insert() method is not aware of sync=true/false, however it should be implementable for Batch or WriteTransaction, when their durability is set to Sync*, e.g. let tx = keyspace.write_tx().durability(Some(SyncData)).

It should be easy to benchmark this by writing with 1, 2, 4, 8 threads with SyncData or SyncAll on a fairly slow device (e.g. HDD).

If this cannot be implemented without a breaking change, this would be a good candidate for a v3 change. TBD