The compose file shipped its own ollama service, so 'docker compose pull'
during deploy re-downloaded the Ollama image and a deploy that starts the
AI stack would spin up a second Ollama alongside an existing/shared one.
- ollama service moved behind a 'bundled-ollama' compose profile, so it is
excluded from the default pull/up (no duplicate, no re-download)
- ai-service no longer depends_on ollama and is documented to point at a
shared instance via OLLAMA_BASE_URL (e.g. http://<host>:11435)
- deploy.sh no longer names ollama in 'compose up'
To run a self-contained Ollama: docker compose --profile bundled-ollama up
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
CRA's build runs fork-ts-checker in a forked process whose IPC needs
more than Docker's default 64MB /dev/shm; too little segfaults
'npm run build' (RpcIpcMessagePortClosedError / SIGSEGV) with no compile
error. Set shm_size on the frontend image build so production deploys
don't hit this. (CI runners need the same via their container options.)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
cesnimda
cesnimda