bluejay/bluejay-infra
1
0
Fork 0
Code Issues Pull Requests 15 Actions Packages Projects Releases Wiki Activity

fc-llm-bridge: repoint Ollama to GX10 NodePort (fix AZ MTU black-hole)

Browse Source 
The PROD-VLAN VIP 10.0.57.201 MTU-black-holes Agent Zero's ~150KB requests
(full prompt + 108 MCP tools) -> connection reset mid-stream -> AZ 'same message
again' loop. Switch FlowerCore__Chat__OllamaBaseUrl to the INFRA-VLAN NodePort
10.0.56.14:30976 (same VLAN as the old cluster, carries 150KB fine). Verified:
150KB POST = 200 via NodePort, times out via VIP. NodePort pinned to 30976 on GX10.
This commit is contained in:
Andrew Stoltz
2026-06-14 15:12:05 -05:00
parent 62ca7dacf6
commit ef6afdd577
1 changed files with 10 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
apps/fc-llm-bridge/fc-llm-bridge.yaml
View File

@@ -164,9 +164,16 @@ spec:
name: fc-llm-bridge-api-keys
key: spare-2
optional: true
# Shared.Chat — Ollama (edge1 Pi 5 + AI HAT+, matches bridge default)
# Shared.Chat — GX10 Ollama via the INFRA-VLAN NodePort (10.0.56.14:30976),
# NOT the PROD-VLAN MetalLB VIP (10.0.57.201:11434). The cross-VLAN path to
# the VIP MTU-black-holes LARGE requests: Agent Zero's full prompt (458-line
# system prompt + 108 MCP tool descriptions ~150KB) times out / resets mid-
# stream there ("Connection reset by peer" in OllamaClient.ChatStreamAsync),
# which made AZ loop on "you have sent the same message again". The NodePort is
# same-VLAN as the old cluster (no inter-VLAN hop) and carries 150KB fine.
# (Small chat/embed requests still work on the VIP; only big agentic prompts broke.)
- name: FlowerCore__Chat__OllamaBaseUrl
value: "http://10.0.57.201:11434"
value: "http://10.0.56.14:30976"
- name: FlowerCore__Chat__HttpTimeout
value: "00:05:00"
# Tier routing override (Wiring A, 2026-06-14): repoint Agent Zero's
@@ -175,7 +182,7 @@ spec:
# Anthropic key is currently 401); Cheap was gemma3:4b which CANNOT
# call tools (400 does not support tools) — fatal for an agentic loop.
# qwen2.5 instruct supports the tool-calling loop; GX10 has the memory.
# OllamaBaseUrl above already points at the GX10 VIP (10.0.57.201).
# OllamaBaseUrl above points at the GX10 NodePort (10.0.56.14:30976).
- name: FlowerCore__Chat__ModelRouter__DefaultRoutes__Balanced__Provider
value: "Ollama"
- name: FlowerCore__Chat__ModelRouter__DefaultRoutes__Balanced__Model