Skip to content

armgit / floci

Public Imported from GitHub
0

Light, fluffy, and always free - The AWS Local Emulator alternative

ci: stop the compat jobs writing per-PR image caches (#2725)

Hector Ventura54dacfc56fa3
diff --git a/.github/workflows/compatibility.yml b/.github/workflows/compatibility.yml
index 90c6fea874..494f633d77 100644
--- a/.github/workflows/compatibility.yml
+++ b/.github/workflows/compatibility.yml
@@ -38,8 +38,12 @@ jobs:
build-args: VERSION=latest
load: true
tags: floci-compat-shim-test
+ # Read-only; warm-compat-caches.yml produces this scope on main. Unlike
+ # floci-native above, these layers are worth caching: the image builds
+ # FROM a published floci/floci tag and its cost is a microdnf install
+ # that changes only when the Dockerfile does. Writing from here would
+ # put a per-PR copy no other PR can read back into the budget.
cache-from: type=gha,scope=compat-shim
- cache-to: type=gha,scope=compat-shim,mode=max
- name: Unflagged aws sqs call routes to AWS_ENDPOINT_URL
run: |
@@ -133,8 +137,14 @@ jobs:
file: docker/Dockerfile.native-package
tags: floci:test-native
outputs: type=docker,dest=/tmp/floci-native-image.tar
- cache-from: type=gha,scope=floci-native
- cache-to: type=gha,scope=floci-native,mode=max
+ # No layer cache here, deliberately. The `COPY native/$TARGETARCH/`
+ # below brings in a binary that changes every commit, and Docker layer
+ # reuse is sequential, so nothing after it can ever be reused. Measured:
+ # the build imported a cache manifest but reported **zero CACHED
+ # layers** — it was write-only. `mode=max` still stored the whole image
+ # including that volatile multi-hundred-MB binary layer, per PR ref,
+ # which was the largest single consumer of the repo's 10 GB budget.
+ # The step costs ~19-28s either way.
- name: Compress native image
run: gzip /tmp/floci-native-image.tar
diff --git a/.github/workflows/warm-compat-caches.yml b/.github/workflows/warm-compat-caches.yml
index e9af842ee7..f93b5aa64b 100644
--- a/.github/workflows/warm-compat-caches.yml
+++ b/.github/workflows/warm-compat-caches.yml
@@ -34,6 +34,9 @@ on:
# (go.sum is gitignored, so filtering on it would match nothing.) Only 8
# commits in 90 days touch this directory, so watching all of it is cheap.
- 'compatibility-tests/sdk-test-go/**'
+ # compat-shim's inputs (see the `shim` job below)
+ - 'docker/Dockerfile.compat'
+ - 'bin/awslocal'
# No *.tf entry: compat-terraform and compat-opentofu resolve nothing at
# build time — `COPY . .` is their final layer, after the yum install and
# bats clones — so .tf changes invalidate no cached work.
@@ -93,3 +96,31 @@ jobs:
# No `load:` — the image itself is discarded; only the cache is wanted.
cache-from: type=gha,scope=${{ matrix.test }}
cache-to: type=gha,scope=${{ matrix.test }},mode=max
+
+ shim:
+ name: warm / compat-shim
+ # Produces the scope that compatibility.yml's compat-aws-cli-shim job reads.
+ # Worth warming, unlike floci-native: this image builds FROM a published
+ # floci/floci tag and its cost is a microdnf install of python3/pip/unzip that
+ # changes only when the Dockerfile does — genuinely stable layers, with nothing
+ # volatile above them.
+ runs-on: ubuntu-24.04-arm
+ timeout-minutes: 20
+
+ steps:
+ - name: Checkout repository
+ uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+ - uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c # v4.2.0
+
+ - name: Build and cache compat shim image
+ uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7.3.0
+ with:
+ context: .
+ file: docker/Dockerfile.compat
+ # Must match the consumer's build-args, or the layers key differently
+ # and the warm is silently useless (compatibility.yml:38).
+ build-args: VERSION=latest
+ # No `load:` — the image is discarded; only the cache is wanted.
+ cache-from: type=gha,scope=compat-shim
+ cache-to: type=gha,scope=compat-shim,mode=max