#efficient-software #ai
# [[Epistemic status]]
#shower-thought #to-digest
# Related
> [!TODO] Related
# TODO
> [!TODO] TODO
> explain how I deploy & monitor fast without human errors - all automated on the cloud
> github + gcp + ovhcloud k8s + ovhcloud container registry etc.
> helm + wait for nodes to boot - don't replace if failure, notify every k8s event to the team automatically
> `make release` -> done
# Continuous delivery with Kubernetes
## Building [[Embedding is the dark matter of intelligence|embeddings]] & indexes at boot
The problem is that you don't want to shut down your old version until the new version is fully ready to take the load. Also you don't want the new version to take over if the boot fail.
```yaml
# Wait until embeddings and index built
startupProbe:
initialDelaySeconds: 60
periodSeconds: 60
# Embeddings + index might take up to 240 minutes depending on compute resources
# (1000 embeddings on 16GB RAM 4 CPU => ~20 min)
timeoutSeconds: 18000
successThreshold: 1
failureThreshold: 600
exec:
command:
- cat
- /home/docker/indexes/knn.index
```
# External links