Would be great to get a quantitative estimation of the performance for redis / postgresQL, perhaps some wrapper functions that make caching easier using either backend.
I've noticed that sometimes the reward calculation takes a while -- with many simultaneous games, we sometimes see the entry not present in the reward buffer, but it gets added by a different node by the time its complete. so this can throw errors
Other things we could cache depending on performance:
- node values and prior logits (these would have to get cleared every time the model weights are updated)
Would be great to get a quantitative estimation of the performance for redis / postgresQL, perhaps some wrapper functions that make caching easier using either backend.
I've noticed that sometimes the reward calculation takes a while -- with many simultaneous games, we sometimes see the entry not present in the reward buffer, but it gets added by a different node by the time its complete. so this can throw errors
Other things we could cache depending on performance: