The portal publishes what can be asked for and the database records
what each cluster was given. Those two were connected exactly once, at
allocation, and never looked at together again - so the questions that
matter after an application is published had no way to be asked:
which live clusters are running this
has the tag moved since they were allocated
which tags can the registry not delete
All three are answerable from data already being written. cluster_app
has had an index on name since it was created and nothing ever queried
it. This is the thing that asks.
prunable is the retention rule made operable. Two things protect a tag
and they guard different mistakes: the catalogue, because a new cluster
may still ask for it, and a live cluster's record, because that cluster
may be rebuilt - and a rebuild that cannot find its image does not fail
loudly, it builds a cluster running different software while every
record still says the same version. Not a retention window: a
constraint, in one query. Run against the real registry it protected
desktop:xfce, which no catalogue entry mentions and a live cluster
holds.
The registry client is here rather than in the build script although
that script asks the same question, because the other half of the
answer comes from the database - and assembled by hand from two
printouts it would eventually be assembled wrong. A repository that
cannot be listed makes the whole answer incomplete and says so, because
a partial list of "safe to delete" is worse than none.
The first version of `apps list` swallowed the error from its own usage
query and printed "-" in the CLUSTERS column, which reads as "nothing
holds this" - the answer an operator would act on when deciding a tag
is safe to delete. It printed that the first time it was run, against a
database one migration behind. The column now prints "?" when the
question could not be asked, because that and "0" are different facts.
docs/23 is the plan the five items came from, and what each turned out
to involve.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>