Interesting. All the requests up until the failure are around 100ms and we aren't seeing anywhere near hitting resource limits or any postgres slowdowns