Event sourced persistence

[ghost]

Why

We want the hydra-node to be efficient in processing events to yield a high throughput on processing transactions.

Work done on as part of #186 has demonstrated that the primary bottleneck to faster transaction processing inside the node was the state persistence logic which simply overwrites the full state with whatever new state has been produced. While we also persist synchronously the ServerOutputs in the API server, this has a low impact on tx processing as we use an append-only file.

What

• Architectural changes are documented in an ADR Introduce ADR for event-sourced persistence #940
• Performance must be better than before
  • Can use hydra-cluster benchmarks to ensure this
• Retain a means to inspect the internal head state in a human-readable format
• No semantic changes of the end-to-end or higher level tests in hydra-node (e.g. Behavior)

How

• Do a spike to validate Introduce ADR for event-sourced persistence #940 and identify definite steps we want to do here
• Create or make sure tests exist
  • A node can be restored from disk
  • (maybe) The persisted state is not growing too fast
• Introduce StateChanged - the actual event in this event sourcing - with a single, dummy StateReplaced HeadState constructor
• Introduce an aggregate :: HeadState -> StateChanged -> HeadState function and change Outcome to return StateChanged instead of NewState
• Enumerate all the StateChanged cases and move state update from update to aggregate cases incrementally Introduce event-sourcing style in the HeadLogic #999
• Change persistence to use events Load and save event sourced state #1000
• Potentially update the ADR to reflect what was actually done + make it accepted

[ch1bo]

Should regroom this if needed with #940