upstream connect error or disconnect/reset before headers, T+40m and still going

Started T+0 03:10 local. p50 1.1s, p95 4.6s, p99 22s, then nothing at all. Body comes back like this, no json, content-type is text/html: upstream connect error or disconnect/reset before headers. reset reason: connection termination Impact: about one call in nine on the primary. failover fires, second vendor picks it up, output shape different enough that our parser bins roughly half of what it hands back. same payload to both, we do not tune per vendor. Mitigation: none yet. dropped the timeout to 2s which made it worse, put it back. end of shift.