discussion: start_block parameter in decorator [SBK-360]

[mikeshultz]

Problem

Currently the start_block param of the on_() decorator only has an affect on the PollingRunner.

silverback/silverback/application.py

Line 128 in 8b313b1

start_block: Optional[int] = None,

I think this may be a point of confusion since it behaves differently depending on the runner. It sure surprised me seeing blocks from 1m blocks ago during testing. I think it's worth considering either documenting, removing, or somehow supporting in the WebsocketRunner.

Possible solutions

I think there's a few options here.

• Document the current behavior, and maybe show a warning when using websockets
• Remove it completely. This would be easy. We wouldn't need to consider historical events. For isntance, the ApePay use case does event backlog processing during the Silverback startup event, so maybe historical processing is unnecessary in Silverback.
• Using it as a filter in WebsocketRunner. Don't process events from before start_block. This could be used as a kind of "deploy code before available on chain" type thing.
• Backfill events and blocks by polling the missed block range even when using websockets. This would be RPC-expensive but would be more akin to a "traditional" listener that can pickup where it left off. The complexity is an ongoing issue with this implementation, however.

[fubuloubu]

In general, guidance should be that it only supports streaming new data events.

The historical handling is useful for restarts, but I think when we finally utilize the task queue correctly it will just start collecting unhandled items in the queue that a restarted worker would start processing on a restart, and we wouldn't need to backtrack anything

[mikeshultz]

In general, guidance should be that it only supports streaming new data events.

The historical handling is useful for restarts, but I think when we finally utilize the task queue correctly it will just start collecting unhandled items in the queue that a restarted worker would start processing on a restart, and we wouldn't need to backtrack anything

Unhandled jobs aren't really the same thing as missed blocks/events though. start_block causes the runner to queue up jobs for historical blocks/events during startup when polling.