Skip to content

Bound what a backend contributes

A backend contributes everything it can see. That is right for a config file you own, and wrong when the credential is broader than the application — a Consul prefix or an SSM path the token can read entirely, where this service should only be reading its own subtree.

config.Filtered wraps any backend and bounds its key space:

config.NewStore(ctx,
    config.WithFiles(fsys, "config.yaml"),
    config.WithBackend(config.Filtered(vault,
        config.Allow("db.**", "cache.**"),
        config.Deny("db.password"))),
)

It is a decorator over the Backend interface, so the same rules apply to a secrets manager, a parameter store, a file, or a nested store. There is nothing per-backend to learn.

What the patterns mean

Patterns match the dotted path a key resolves at, not the backend's native addressing — you do not need to know whether a store nests by /, by underscore, or not at all.

Pattern Matches Does not match
db.host db.host db, db.host.name
db.* db.host db.primary.host, db
db.** db.host, db.primary.host db, cache.host
** everything —

The segment boundary is the whole design. * stops at one; ** crosses any number. Deliberately not regular expressions: a regexp invites patterns that cross boundaries in ways that are hard to reason about against a tree, and a mistake in one silently widens a filter rather than narrowing it — the wrong direction when what it is holding back is a secret.

Allow, deny, and which wins

  • Neither — everything passes.
  • Allow only — a whitelist.
  • Deny only — a blacklist.
  • Both — allow, then deny. Deny wins.

Deny winning is the safe reading: a caller who has written both has expressed a narrowing intention twice, and resolving the overlap the other way would let an Allow re-expose something explicitly denied.

A denied key is invisible, not an error

The layer simply does not carry it. Reads fall through to whatever lower layer defines it, and if nothing does, the key is absent exactly as though the backend never held it.

Writes follow from that without a special case: routing skips a backend that defines nothing for a key, so a write to a denied key lands in the next writable layer down — which is what "this backend does not own this key" should mean.

This is a visibility bound, not a security control

Filtering happens after the backend has fetched and parsed. A denied key is still read from the remote system, still crosses the network, and still sits in memory briefly.

If your goal is that the process never receives a value, narrow the credential or the backend's own configuration. Filtered bounds what your configuration exposes, not what your client fetches.

Filtering a secrets backend

A denied key on a Sensitive backend keeps its sensitive marking even though it is hidden. So this still fails:

store.Apply(ctx, config.Set("db.password", "new"))
// → ErrSensitiveLeak

That asymmetry is deliberate, and it is the one place Filtered behaves differently depending on what it wraps.

The leak guard works by asking which source defines a key. A filtered-out key defines nothing — so without this, denying db.password on Vault would make a write of it route into the plain file beneath and not be refused. The filter would have quietly re-opened the hole ErrSensitiveLeak exists to close, and the failure would be a secret on disk with no error at all.

Two consequences worth knowing:

  • A denied secret becomes unwritable anywhere, rather than routing down. That is the intended outcome: "do not expose this secret" should not mean "so write it somewhere worse".
  • Only paths the backend actually holds are guarded. The withheld set is recorded from real data rather than derived from the rules, so denying a key the backend does not have refuses nothing.

Capabilities pass through

A filtered writable backend is still writable; a filtered watchable one still watchable. The wrapper claims exactly what the backend underneath it claims — no more, so the store never takes a path the backend cannot honour, and no less, so a capability you configured is not silently dropped.

ID() and Capabilities() forward unchanged, which matters more than it looks:

  • The write path finds a backend by matching a layer's Source.Name against ID(), so a decorator that renamed itself would contribute layers no write could ever reach.
  • The store reads Capabilities().Sensitive directly, so a decorator returning empty capabilities would strip the flag from a secrets backend entirely.

Both are asserted by tests that fail if the forwarding is removed.