Make the time-to-deploy 2x quicker: 2mins vs 4mins #522
Conversation
This replaces our Dagger on Fly.io setup with Dagger on Namespace.so Before this change, throughout August 2024: - deploys took `4m 36s` 95% of the time - `1` in `14` deploys failed Let's see how the above changes after a few weeks of running. This is the first part of the required change. The second part will follow after this gets merged. We need to: - Cleanup secrets & variables from GitHub Actions - De-provision the Dagger Engine on Fly.io - Remove the deploy config from fly.io dir - Update the INFRASTRUCTURE diagram We still keep K8s to see if we can improve on Namespace.so, the new Speed King 🏎️ Signed-off-by: Gerhard Lazu <gerhard@changelog.com>
a90b8bf to
21f1ed1
Compare
|
On the second run, I was expecting Build, test & publish... to use the cached Docker volume This is how the same run behaved on K8s, where the cache was definitely used: https://github.com/thechangelog/changelog.com/actions/runs/10664539072/job/29556002949 I was expecting that particular step to take around Why did the cached volume not work as expected on Namespace.so? This message might explain it, even if it's somewhat hidden: If cache misses are indeed "very rare", I would expect the next run to have a cache hit. Let's see... |
|
That was it! Second run worked much better: It will be worth keeping an eye on these cache misses over the next few weeks (notice the time difference between the first 2 uncached runs vs the third cached run): |
This was left around in case this change would not work out in practice: - #522 It did, and this is no longer needed. Signed-off-by: Gerhard Lazu <gerhard@changelog.com>
First run, empty cache (first time penalty)
Second run, primed cache
This replaces our Dagger on Fly.io setup with Dagger on Namespace.so
We still keep K8s to see if we can improve on Namespace.so, the new Speed King 🏎️
This is the first part of the required change. The second part will follow after this gets merged. We need to:
Before this change, throughout August 2024:
4m 36s95% of the time1in14deploys failed, meaning we had a93%success rate