Ben Johnson

Nope, it should not

We've also had many users over at https://github.com/timberio/vector bump into this. We'll see if we can contribute an improvement.

@dbhobbs we are experiencing the same issue, did you find a work around? I can also confirm this is specific to `21.2`, `21.1` does not have this issues. @paulswartz that...

> Another would be to strip down the inner sink, just using it as a service and pulling up things like timeouts and retries to the level of the outer...

Yeah, there are a few reasons I marked it as `should`: 1. Logstash does not retry partial failures. 2. The status code returned by Elasticsearch is 201. 3. The most...

> At the least, I think we should be logging and reporting a metric for failed inserts into ES. Agree, and @bruceg implemented logging for partial failures in https://github.com/timberio/vector/pull/2185.

Yeah, and clearly it would be nice to have a test for this if there's a way.

Yep, that was exactly my thought. `c5.*` instances offer exactly 93.75% of the memory that `c4.*` instances have available. I'm thinking we allow an even 10% difference.

Thanks for reporting, will investigate.

Thanks @eyefodder, a work around for now is to add a 2 second sleep before the process exits, but we'll get this resolved so that isn't necessary.