Raise CPU limit off the 2-core ceiling

Ya

12 Agosti 2026, 04:05 UTC

Mwandishi

Kamo

Ahadi ya

27f2267

Both replicas sat pinned at 1997m/1999m against a 2-core limit for ~47h. That starved /languages enough that the liveness and readiness probes timed out (443 and 921 failures) and dropped one replica out of the Service endpoints entirely, concentrating all traffic on the survivor and turning ordinary translations into 90s timeouts. k1m1 has 48 cores at ~16% utilisation, and argos inference is CPU-bound across several gunicorn workers, so the headroom is genuinely usable.