- Ya
- 10 Mei 2026, 09:04 UTC
- Mwandishi
- Kamo
- Ahadi ya
- 929eba5
A single replica is the bottleneck for the dictionary sync workflow: each gunicorn worker handles one request then auto-restarts (~5-10s window of connection-refused), and the 2-CPU limit caps inference throughput. With both runtime callers and a 12k-string sync hitting the same instance, requests queue past TranslateService's per-call timeout (now 90s) and most chunks time out — the last full sync run took 3+ hours and produced only 2.9% Chinese in zh.json. Two replicas double effective CPU (4 cores total) and double the gunicorn worker pool, giving requests somewhere to land while one worker is recycling. Both pods schedule on k1m1 (only node with the hostPath /var/lib/libretranslate where argos packages live); k1m1 has ample headroom (~92% CPU free, ~67% memory free).