e7625cd322
Ostatni krok fazy 4 KB (plan §2 Decyzja 2, §5): kb-query przestaje być martwe
przez ~16 h/dobę, gdy SOLARIA (GPU) śpi — zapytania embeduje wtedy lokalna
Ollama CPU na PIHA (wolniej: ~790 ms+ vs ~207 ms na GPU, ale działa).
Nowy serwis services/ollama-piha (GitOps, owner_node: piha):
- ollama/ollama:latest (arm64 natywnie), OLLAMA_KEEP_ALIVE=0 — model zwalnia
RAM natychmiast po każdym wywołaniu (spike, nie rezydent; PIHA dzieli 8 GB z HA)
- bind wyłącznie 127.0.0.1 + LAN_BIND_IP (192.168.31.5), nigdy 0.0.0.0/Tailscale
- named volume ollama_piha_models (NVMe data-root) zamiast bind-mounta — obraz
biega jako root w kontenerze i bind łamałby wzorzec uid PIHA (oskar=1004,
kontenery uid 1000, setgid pi)
- override hosts/piha/runtime/ollama-piha: mem_limit 2560m (wartość startowa
z planu, do potwierdzenia kalibracją na żywo), świadomie bez mem_reservation
- pull bge-m3 to jawny, ręczny krok deployu (README) — obraz nie ma modeli
kb-query — maszyna stanów fallbacku (app/embed_router.py):
- health-check SOLARII (GET /api/tags, timeout 1.5 s) z cache 30 s — zero
sondowania per request; po powrocie SOLARII ruch wraca na GPU w ≤30 s
- primary up → embed na SOLARII z twardym timeoutem 3 s; błąd W TRAKCIE
zapytania = jednorazowe przełączenie (krok 3b planu): status down na 30 s
i TO SAMO zapytanie leci na fallback — user nie widzi błędu SOLARII
- primary down → embed prosto na ollama-piha (bez twardego timeoutu: CPU +
zimny load modelu to legalnie pojedyncze sekundy)
- 503 tylko gdy oba backendy padłe (lub fallback nieskonfigurowany)
- inwariant modelu, druga połowa: każdy backend weryfikowany raz, leniwie przy
pierwszym użyciu, że /api/tags zawiera EMBED_MODEL (bge-m3 — ta sama wartość
co startowy check przeciw document_chunk.model/document_summary.embedding_model);
niezgodność = ERROR log + 500, nigdy ciche liczenie dystansów między
różnymi przestrzeniami embeddingów; leniwie, bo śpiąca SOLARIA nie może
blokować startu serwisu
- odpowiedź /search: nowe pole embed_backend ("solaria"|"piha") + sol_status
wg realnego świata routera (UI już renderuje down jako "offline (fallback
embed)"); log INFO backend=... elapsed_ms=... per zapytanie
- /healthz: sol_status przez cache routera (spójny widok z routingiem) +
fallback_status (żywa, tania sonda /api/tags)
Konfiguracja spójnie przez env (compose + env.example + service.yaml + README):
EMBED_PRIMARY_URL (zastępuje OLLAMA_URL), EMBED_FALLBACK_URL (pusty = brak
fallbacku, zachowanie sprzed kroku 2), EMBED_{PRIMARY,FALLBACK}_NAME,
EMBED_HEALTH_TTL_S/EMBED_HEALTH_TIMEOUT_S/EMBED_PRIMARY_TIMEOUT_S.
Testy: 39 pass (14 nowych w test_embed_router.py: cache TTL, failover w trakcie
zapytania, powrót po TTL, oba padłe, mismatch modelu na primary i fallbacku,
tag "bge-m3:latest" vs "bge-m3"); docker build + smoke (importy + uvicorn do
guardu KB_DSN) OK; compose config OK dla obu stacków.
Deploy (Oskar, na PIHA z mastera po merge):
cd ~/homelab-codex-ws && git pull
# 1. ollama-piha
cp services/ollama-piha/env.example services/ollama-piha/.env
docker compose -f services/ollama-piha/docker-compose.yml \
-f hosts/piha/runtime/ollama-piha/docker-compose.override.yml \
--env-file services/ollama-piha/.env up -d
docker exec ollama-piha ollama pull bge-m3 # ręczny krok, obowiązkowy
services/ollama-piha/healthcheck.sh
# 2. kb-query (dopisać fallback do istniejącego .env)
echo 'EMBED_FALLBACK_URL=http://192.168.31.5:11434' >> services/kb-query/.env
docker compose -f services/kb-query/docker-compose.yml \
-f hosts/piha/runtime/kb-query/docker-compose.override.yml up -d --build
services/kb-query/healthcheck.sh
# (deploy-node.sh też podniesie oba serwisy z hosts/piha/services.yaml,
# ale pull bge-m3 i .env pozostają ręczne)
Weryfikacja: testy A/B/C w services/kb-query/README.md (backend=solaria przy
SOLARII online; backend=piha przy symulacji offline; powrót na GPU w ≤30 s).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
77 lines
3.4 KiB
Markdown
77 lines
3.4 KiB
Markdown
# ollama-piha
|
||
|
||
Local CPU Ollama on **PIHA**, serving exactly one purpose: the **fallback embed
|
||
backend** for `kb-query` while SOLARIA (the GPU node, ~16 h/day powered off)
|
||
sleeps. Model: `bge-m3` — the **same** model as SOLARIA's Ollama, because query
|
||
embeddings must live in the same vector space as the pgvector index
|
||
(`document_chunk.embedding VECTOR(1024)`); a different/smaller model is not an
|
||
option (module 5 phase 4 plan §2 decision 2).
|
||
|
||
Expected latency: bge-m3 embeds in ~207 ms on SOLARIA's GPU vs ~790 ms on x86
|
||
CPU; on the Pi 5 expect single seconds per query (plus model load, since the
|
||
model is never resident — see below). Slower but alive beats fast but dead.
|
||
|
||
## Design constraints
|
||
|
||
- **`OLLAMA_KEEP_ALIVE=0`** (pinned in compose): PIHA is the RAM-bound 8 GB box
|
||
shared with Home Assistant. The model is unloaded immediately after every
|
||
call — a transient ~1.5–2 GB spike per embed, ~100 MB idle daemon, never a
|
||
resident cost.
|
||
- **`mem_limit: 2560m`** (host override, `hosts/piha/runtime/ollama-piha/`):
|
||
hard cgroup ceiling, plan §2 D2 starting value. The cgroup OOM killer
|
||
restarts this container instead of the host OOM killer picking a victim
|
||
(which could be Home Assistant). Confirm/trim after live calibration.
|
||
- **Bind**: `127.0.0.1` + `LAN_BIND_IP` (192.168.31.5) only — kb-query calls it
|
||
over the host LAN interface (same pattern as kb-query → kb-postgres:5433).
|
||
Never `0.0.0.0`, never a Tailscale bind, no public ingress.
|
||
- **Storage**: Docker named volume `ollama_piha_models` (NVMe data-root), not a
|
||
bind mount — the ollama image runs as in-container root and would break
|
||
PIHA's uid pattern (host oskar=1004, containers uid 1000, setgid group pi)
|
||
if it wrote to a shared bind directory.
|
||
|
||
## Deploy (PIHA, master, after merge)
|
||
|
||
```bash
|
||
cd ~/homelab-codex-ws && git pull
|
||
cp services/ollama-piha/env.example services/ollama-piha/.env # LAN_BIND_IP
|
||
docker compose -f services/ollama-piha/docker-compose.yml \
|
||
-f hosts/piha/runtime/ollama-piha/docker-compose.override.yml \
|
||
--env-file services/ollama-piha/.env up -d
|
||
```
|
||
|
||
**Then pull the model — this does NOT happen automatically:**
|
||
|
||
```bash
|
||
docker exec ollama-piha ollama pull bge-m3
|
||
```
|
||
|
||
Verify:
|
||
|
||
```bash
|
||
services/ollama-piha/healthcheck.sh # checks container + API + bge-m3 present
|
||
time curl -s http://127.0.0.1:11434/api/embeddings \
|
||
-d '{"model":"bge-m3","prompt":"test kalibracyjny"}' | head -c 80
|
||
```
|
||
|
||
(`deploy-node.sh` on PIHA also picks this service up from
|
||
`hosts/piha/services.yaml` once `.env` exists — the `ollama pull bge-m3` step
|
||
stays manual either way.)
|
||
|
||
## Calibration (plan §5 step 4 — gate, not formality)
|
||
|
||
Before trusting the fallback under load, on live PIHA at a normal (not
|
||
night-quiet) hour: run a few embeds as above while watching
|
||
`docker stats ollama-piha`, note peak RAM and wall time. Verdict per plan §5
|
||
step 5: keep as default fallback / tune `mem_limit` / fall back to explicit
|
||
503 degradation.
|
||
|
||
## Relation to kb-query
|
||
|
||
kb-query's router (`services/kb-query/app/embed_router.py`) health-checks
|
||
SOLARIA with a ~30 s cache and only sends embeds here while SOLARIA is down.
|
||
kb-query verifies at first use that this backend actually serves `bge-m3`
|
||
(`/api/tags`) and refuses to embed against a mismatched model. Configuration:
|
||
`EMBED_FALLBACK_URL=http://192.168.31.5:11434` in `services/kb-query/.env`.
|
||
See `services/kb-query/README.md` for the fallback verification plan (tests
|
||
A/B/C).
|