Bohan Zhang

The same problem. It occurs on every dataset.

Change `agent.append(trainCurve, epoch, total_loss[0])` to `agent.append(trainCurve, epoch, total_loss.cpu().numpy()[0])` may solve the problem. It works on python3.5 and pytorch1.1.0.

tracked by #4586

please follow https://github.com/risingwavelabs/risingwave/pull/5647 and change the read process to an async stream.

I briefed the pr and I think this can be merged to branch `ns-sink-meta` instead of main

I think I identified the problem. I add some log here https://github.com/singularity-data/risingwave/blob/8abbb65c4414ac08bef57cae325d93885291568c/src/meta/src/stream/source_manager.rs#L561-L573 and I got the following log ``` {52: ['24-18', '24-7'], 56: ['24-2', '24-5'], 36: ['24-13', '24-8'], 49: ['24-13',...

related #4817

> And DLQ for corrupted bytes that are not even json? json deserialization is quite deterministic. I think it is okay to drop corrupted ones.

As shown in #4626, the behavior of the json parser is as expected. I believe I was confused by json and debezium json before. Json parser can insert NULL in...

> There's https://docs.rs/prost-reflect/latest/prost_reflect/index.html, probably someone is working on this. I am looking into this and I will take the issue