Running with #49979, I'm still seeing the stall that the change was hoping to fix.

Here's a dump of the goroutines: https://gist.github.com/nvanbenschoten/3cffff71d9847e09e366bfaeed0f15b4.

Notably, we still have a collection of importWorkers all stuck in (*DatumRowConverter).SendBatch. However, we have no one pulling from this channel. After reading through this code a bit, I believe this channel is supposed to be consumed by this goroutine:

That goroutine is not in the stacktrace. runImport is pretty suspicious. It links together two ctxgroup instances here:

However, it only seems to care about producerGroup hitting an error before group (consumerGroup), not the other way around. If ingestKvs throws an error, I don't see anything that would stop this from deadlocking.