feat(buildkit-bench): the benchmark leaves with its answer

sean commented

2026-09-10 09:00:04 +00:00

Owner

The three throwaway daemons compared storage classes; the answer is zfs-nvme-128k, which the shared daemon has run for days. Deleting the cluster file is enough because flux-system is prune:true, so the Kustomization is pruned and its own finalizer removes all 22 objects including 90 GiB of cache claims -- unlike Stage 5a-3b, where a prune:false parent orphaned Kustomization/minio-operator. Inventory drops to 4 pins in 3 files, the number its own note predicted. Also fixes three rustfs-identities comments that still described tempo.yaml's converge Job and the rustfs-mirror directory, both retired with MinIO; nothing is NAMED mirror any more. https://claude.ai/code/session_01KZoQin34jeyt6nDGqvJA76

sean added 1 commit

2026-09-10 09:00:04 +00:00

feat(buildkit-bench): the benchmark leaves with its answer

binjovi/ci Binjovi failed the frozen plan

Details

16036c3d1f

Three throwaway BuildKit daemons differing only in the StorageClass their
cache used, so the classes could be compared under a real build. The answer
is in and has been in production for days: the shared daemon's
`binjovi-buildkit-cache` runs `zfs-nvme-128k`. The comment in
tests/check-roll-gate-exemptions.sh said "the directory leaves with the
answer", so it does -- with 90 GiB of cache claims and 22 objects.

ONE COMMIT IS ENOUGH HERE, AND THE REASON IS WORTH WRITING DOWN. `flux-system`
owns ./kubernetes/flux/clusters/seanfarm and is `prune: true`, so deleting
the cluster file prunes Kustomization/buildkit-bench, and its own
`prune: true` plus finalizers.fluxcd.io then removes every one of its 22
inventory entries: the namespace, three ConfigMaps, three Services, three
Deployments, three PVCs, five Certificates, two Issuers, a
CiliumClusterwideNetworkPolicy and a CiliumNetworkPolicy.

Stage 5a-3b of the MinIO retirement got this wrong three hours ago. It
deleted `minio/operator/` and `cluster/minio-operator.yaml` in one commit,
but the parent there (`cluster`) is `prune: false`, so Flux ORPHANED
Kustomization/minio-operator instead of pruning it -- and the orphan then
pointed at a path that no longer existed, went Ready: False, and could never
prune the nine objects it owned. Check the parent's prune mode before
deleting a Kustomization.

The upstream inventory drops to 4 pins of moby/buildkit across 3 files,
which is exactly the number its own note predicted: "back to 4 when that
directory goes".

Also corrected here: three comments in rustfs-identities/ that still
described the migration. binjovi-logs.yaml, forgejo.yaml and zot.yaml each
claimed tempo.yaml "still carries its own converge Job" (it retired with
MinIO) and pointed at "the mirror Job in rustfs-mirror/<x>.yaml" (that
directory retired with MinIO too). Nothing in the rustfs tree or the cluster
is NAMED mirror any more -- there was no rename to make, only stale prose
describing deleted files. They now say what is true: no identity file carries
a converge Job, the bulk copy that used these scoped keys is gone, and every
consumer has read RustFS since.

Claude-Session: https://claude.ai/code/session_01KZoQin34jeyt6nDGqvJA76

sean force-pushed sean/retire-buildkit-bench from 16036c3d1f

binjovi/ci Binjovi failed the frozen plan

Details

to 79ac9be3ca

binjovi/ci Binjovi completed the frozen plan

Details

2026-09-10 09:11:38 +00:00

Compare

binjovi-bot merged commit 79ac9be3ca into trunk

2026-09-10 09:17:28 +00:00

Rows
Columns

feat(buildkit-bench): the benchmark leaves with its answer #539