Update Discipline

~14 min · ops, updates

Level 0Downloader

0 XP0/41 lessons0/11 achievements

0/120 XP to next level120 XP to go0% complete

Things that drift

Ollama daemon — major versions add features (OpenAI compat, MLX backend on Apple) and occasionally change defaults.
Model tags — qwen2.5:7b may resolve to a different actual GGUF a month from now.
Capabilities — a model that didn't support tools might gain them via a new release.
System prompt templates — minor template changes are why a Modelfile that worked yesterday produces garbage today.

Update strategy

Update Ollama itself on a known cadence (monthly) — read release notes for the breaking changes.
Pin model tags for production. Use ollama pull qwen2.5:7b-instruct-q5_K_M not qwen2.5. The fully-qualified tag is much less likely to silently move.
Re-run your benchmark suite after a daemon update. Catch performance regressions before they ship.
Keep a "last-known-good" model on disk. If a new pull breaks your prompts, fall back to the previous tag.

Code

Update + verify·bash

# Update Ollama itself (Homebrew)
brew upgrade ollama

# Update each installed model to its latest tag
for m in $(ollama list | tail -n +2 | awk '{print $1}'); do
  echo "=== updating $m ==="
  ollama pull "$m"
done

# Re-run your benchmark — don't assume nothing changed
python3 ~/scripts/ollama_bench.py

# If something broke, downgrade Ollama and pin the previous model tag

Pin tags in production code·python

# DON'T — moving tag, can change under you
adapter = OllamaAdapter(model="qwen2.5:7b")

# DO — fully-qualified, pinned
adapter = OllamaAdapter(model="qwen2.5:7b-instruct-q5_K_M")

External links

Exercise

Run ollama list, identify which models you pulled by short tag (e.g. qwen2.5:7b) vs fully-qualified (qwen2.5:7b-instruct-q4_K_M). Re-pull the short-tag ones with explicit fully-qualified tags. Confirm ollama show reports a stable tag.