ADR 0188: a provider declares what it derives for each consumer, and the mesh tells both ends

Issue 124: a value the mesh's own rule produced reached neither end as a
statement. The object store's provisioner derived each consumer's bucket in
its own code; all three consumers transcribed the rule into their own
definitions, one of them wrong, and each of the three also named the machine
it happens to run on.

A served value may now name the consumer the mesh is serving. Design 27
amended; issue 124 resolved.
This commit is contained in:
2026-10-02 21:24:31 +02:00
parent 983fd412c6
commit 0b9fc90885
4 changed files with 180 additions and 6 deletions
@@ -0,0 +1,130 @@
---
topic: the mesh
status: accepted
date: 2026-10-02
deciders: jochen
reconstructed: false
extends: 02-DECISIONS/0048-a-provider-creates-the-credential-the-mesh-minted.md
---
# 188. A provider declares what it derives for each consumer, and the mesh tells both ends
## Context
An arrangement between a consumer and a provider is delivered entirely by the mesh. Where the
provider is, which port it answers on, what name the consumer must present, where its password
is — each arrives as a fact the consumer reads from its binding, or as `${bound:…}` filled into a
file before the declaration leaves the control plane. The provider invents none of it and hands
none of it back ([ADR 0048](0048-a-provider-creates-the-credential-the-mesh-minted.md)).
One kind of value escapes that. Where the **provider names the resource** — a bucket, a database,
a vhost — the name is derived from the consumer, per consumer, and the mesh has no way to carry
it. `serves` is a literal block in the provider's definition: the same values for every consumer.
A provisioner's contract takes a provision and returns nothing. So a value the mesh's own rule
produced reaches neither end as a statement; it is recomputed at one end and transcribed at the
other.
The object store is the instance ([issue 124](../04-ISSUES/124-a-consumer-cannot-be-told-what-its-provider-derived/00-report.md)).
Its provisioner normalises the login the mesh minted into a bucket name and creates, checks and
removes exactly that; the rule lives in twenty lines of the module's own TypeScript. Its three
consumers each write the answer into their own definition by hand. Two transcribed it correctly;
one named a predecessor's bucket, and would have authenticated successfully and been refused on
every object, which reads like a credential fault and is not one.
Even corrected, the transcriptions are wrong in a second way. Each is `mesh-<node>-<slug>`, so
each **names the machine the module happens to run on today** — a definition stating a fact about
one installation, which [ADR 0155](0155-a-definition-names-no-installation-and-how-that-is-checked.md)
forbids and whose check does not catch because the name is not a domain. Move any of the three to
another machine and its configuration points at a bucket its key cannot open.
The shape is not the object store's. A database provisioner that prefixed names, a queue provider
that scoped vhosts, any provider that derives a resource from who is asking: each forces the
consumer to reproduce somebody else's rule and keep it in agreement by hand.
## Decision
**1. A served value may name the consumer the mesh is serving.** A `serves` block, which is
literal today, may interpolate the mesh's own statement of who the consumer is:
- `${consumer:as}` — the identity the mesh minted for this consumer, exactly as the login it is
told to present ([ADR 0049](0049-a-consumers-identity-fits-the-tightest-backend.md));
- `${consumer:as:dns}` — the same identity written as a DNS label.
Nothing else. **The mesh learns no protocol here; it spells its own name in an alphabet it already
knows.** The identity is the mesh's, minted by the mesh, already capped at twenty characters
because of what an S3 access key accepts; `dns` is that same name with its separator written `-`
instead of `_`, which is the whole of the difference between the mesh's identifier alphabet and
the one buckets, vhosts and hostnames use. A provider that needs a prefix or a suffix writes it
around the placeholder, because a served value is a string.
The rejected alternative is **the provider returning values from provisioning** — the natural
channel, since the provider is what derived them. It is rejected for three reasons, in order of
weight. It inverts the delivery the mesh is built on: a grant would carry data the provider wrote
rather than only data the mesh minted, and [ADR 0048](0048-a-provider-creates-the-credential-the-mesh-minted.md)
removed exactly that second path once already. It makes a consumer's declaration incomplete until
its provider's reconcile loop has run, so a consumer could not be composed before a provider
answered — a bootstrap order the mesh does not have and does not want. And it puts the rule where
nothing can check it: a value that arrives from a running process cannot be refused at resolution,
only discovered wrong later, which is the failure this record exists to end.
**2. The mesh resolves it once, per consumer, and tells both ends from the one resolution.** At the
moment a consumer's declaration is composed, the mesh knows exactly who the consumer is. There, and
only there, the placeholders are filled. The result reaches:
- the **consumer**, as the served facts in its binding file and as `${bound:<provision>:<key>}` in
any file it writes — unchanged mechanisms, carrying one more key;
- the **provider**, as `serves` on that consumer's entry in its contributions file, so the
provisioner is *told* the name rather than recomputing it.
**The provider stops deriving in code and starts declaring.** One statement, filled once, delivered
to both ends: the two cannot disagree, because there is no second computation to disagree with.
**3. A served value stays settled before it is per-consumer.** Settings still compose into `serves`
([ADR 0174](0174-a-node-varies-a-module-through-settings-and-kept-regions-never-an-edit.md)), and the consumer
placeholders are filled after that, so an operator may set a prefix and the mesh still derives the
rest. A `${consumer:…}` naming a fact or an alphabet the mesh does not have is refused when the
definition is parsed, with what it may say.
**4. A consumer may no longer name the resource its provider derives.** With the value delivered,
a literal in a consumer's definition is not merely redundant — it is the one thing that can
disagree with what the provider will actually create. The three object-store consumers lose their
hand-written bucket names in this change.
## Consequences
- One more thing a definition may say, and one less thing a module may be wrong about. The
vocabulary grows by a placeholder; the catalogue loses three literals that named this
installation's control node.
- A provider's naming rule becomes readable in its definition instead of in its source. `minio`'s
`bucketFor` goes; the manifest says `"bucket": "${consumer:as:dns}"` and the provisioner uses
what it is given.
- A provider that already serves consumers keeps serving them: the derived value equals what the
code derived, so no bucket, database or login changes name. This is a change of **who says it**,
not of **what is said**.
- The mesh now holds a rule in another system's alphabet — one rule, `dns`, stated once. A second
alphabet is a decision, not an addition: the cost of each is that the mesh must be right about
somebody else's naming, and that cost is only worth paying where the mesh already mints the name.
## How this is checked
- A served value naming an unknown fact or alphabet is refused at parse, with the list of what it
may say — tested on both halves of the message.
- Resolving a consumer whose provider derives a value puts that value in the consumer's binding
file, in its `${bound:…}` substitutions, and in the provider's contributions entry for that
consumer — one test asserting the three agree, because agreeing is the whole point.
- Two consumers of one provider on one machine get two different derived values, and neither gets
the other's.
- A catalogue-wide test refuses a consumer definition that writes a literal where its provider
derives: the provider's `serves` names the key, so the catalogue can say which definitions
transcribe one.
- `dns` is checked against the identity the mesh actually mints, not against an invented string:
the test derives an identity with `ConsumerIdentity` and asserts the label it becomes.
## References
- [issue 124 — a consumer cannot be told a value its provider derived for it](../04-ISSUES/124-a-consumer-cannot-be-told-what-its-provider-derived/00-report.md)
- [ADR 0048 — a provider creates the credential the mesh minted, and seals nothing](0048-a-provider-creates-the-credential-the-mesh-minted.md)
- [ADR 0049 — a consumer's identity fits the tightest backend](0049-a-consumers-identity-fits-the-tightest-backend.md)
- [ADR 0174 — a node varies a module through settings and kept regions, never through an edit](0174-a-node-varies-a-module-through-settings-and-kept-regions-never-an-edit.md)
- [ADR 0155 — a definition names no installation, and how that is checked](0155-a-definition-names-no-installation-and-how-that-is-checked.md)
- [design 27 — a module requires, the mesh resolves](../03-DESIGN/01-to-be/27-a-module-requires-the-mesh-resolves.md)
+1
View File
@@ -188,6 +188,7 @@ python3 00-META/checks/index.py fail if stale
- **0185** — [A control plane behind its seat's row serves what it can](0185-a-control-plane-behind-its-seats-row-serves-what-it-can.md)
- **0186** — [A ban list never holds a neighbour, and the mesh's own bans are its own wherever they hang](0186-a-ban-list-never-holds-a-neighbour.md)
- **0187** — [A dead tracker is not the machine's failure](0187-a-dead-tracker-is-not-the-machines-failure.md)
- **0188** — [A provider declares what it derives for each consumer, and the mesh tells both ends](0188-a-provider-declares-what-it-derives-for-each-consumer.md)
### Its tiers, from the bottom up
@@ -2,7 +2,7 @@
layer: to-be
status: in-progress
code: [mesh-controller internal/catalogue]
updated: 2026-09-30
updated: 2026-10-02
decisions:
- 02-DECISIONS/0112-a-module-definition-names-no-node-mesh-or-path.md
- 02-DECISIONS/0155-a-definition-names-no-installation-and-how-that-is-checked.md
@@ -14,6 +14,7 @@ decisions:
- 02-DECISIONS/0084-which-provider-serves-a-consumer.md
- 02-DECISIONS/0046-a-module-configuration-is-its-assignments-not-its-manifest.md
- 02-DECISIONS/0038-the-mesh-assigns-the-port.md
- 02-DECISIONS/0188-a-provider-declares-what-it-derives-for-each-consumer.md
---
# 27 — A module requires, the mesh resolves
@@ -206,6 +207,22 @@ name when nothing sets it. That is the contract half of this design's operator p
the placeholder allows: the definition says which values reach which requirement, and nothing else
does. *How it is checked:* the unit tests named in issue 173, and the plan comparison that closed it.
*A provider says once what it derives for each consumer (2026-10-02,
[ADR 0188](../../02-DECISIONS/0188-a-provider-declares-what-it-derives-for-each-consumer.md),
[issue 124](../../04-ISSUES/124-a-consumer-cannot-be-told-what-its-provider-derived/00-report.md)):*
where a provider **names the resource** it gives each consumer — a bucket, a database, a vhost — the
name is derived per consumer, and a literal `serves` block could not carry it. A served value may
now name the consumer the mesh is serving: `${consumer:as}`, the identity the mesh minted, and
`${consumer:as:dns}`, that same identity written as a DNS label. Nothing else — **the mesh learns no
protocol here; it spells its own name in an alphabet it already knows.** Settings are laid on first,
so an operator may still set a prefix and the mesh derives the rest. The mesh fills it at the one
moment it knows who the consumer is, and the one filled value reaches both ends: the consumer, as
its binding's served facts and as `${bound:<provision>:<key>}` in any file it writes; the provider,
as `derived` on that consumer's entry in its contributions file, so its provisioner is told the name
rather than recomputing it. A consumer that writes the derived value into its own definition instead
of asking for it is refused, naming the placeholder to use. *How it is checked:* the unit tests in
ADR 0188's "how this is checked", each run against the unchanged controller first.
## How a definition reads what was resolved
**One form, naming a requirement and a field of its contract.** A definition that needs the database's
@@ -214,7 +231,9 @@ name in a configuration file writes the same thing: the requirement's name and t
controller fills it at resolution.
This one form replaces the placeholders that exist today, one per mechanism: bound values, secrets,
ports and machine facts.
ports and machine facts. It subsumes the consumer placeholder too — a value a provider derives is
read by the consumer exactly as any other field of the contract is, and `${consumer:…}` is only
how the *provider* states the rule.
**The seat placeholder stays, for the controller alone.** The controller composes its own
declaration and reaches the store and broker it made before any module existed, so it cannot be
@@ -1,9 +1,9 @@
---
status: located
status: resolved
opened: 2026-09-26
located-in: [mesh-controller internal/catalogue/declaration.go, mesh-sdk src/provisioner, mesh-catalog modules/minio]
fixed-by:
amended-design:
located-in: [mesh-controller internal/catalogue, mesh-sdk src/provisioner, mesh-catalog modules/minio]
fixed-by: 02-DECISIONS/0188-a-provider-declares-what-it-derives-for-each-consumer.md
amended-design: 03-DESIGN/01-to-be/27-a-module-requires-the-mesh-resolves.md
---
# 124 — A consumer cannot be told a value its provider derived for it, so it transcribes one
@@ -63,3 +63,27 @@ compares it to what the provider will actually create. The one wrong instance wa
- What would have caught the wrong instance? A test that resolves a consumer's grant and compares the
bucket in its own configuration against the one the provider would create is a check that could
exist today, for any interface, without the mechanism above.
## Answered, 2026-10-02 — [ADR 0188](../../02-DECISIONS/0188-a-provider-declares-what-it-derives-for-each-consumer.md)
The channel is the provider's own `serves` block, which may now name the consumer the mesh is
serving: `${consumer:as}` and `${consumer:as:dns}`. The mesh fills it once, where it knows who the
consumer is, and delivers the one filled value to both ends — the consumer's binding and its
`${bound:…}` substitutions, and the provider's contributions entry, so a provisioner is told the
name rather than deriving it. Each open question above, answered:
- **Should a provider return values from provisioning?** No. It would make a grant carry data the
provider wrote, make a consumer's declaration wait on its provider's reconcile loop, and put the
rule where nothing can refuse it. The reasoning is in the record.
- **Or should `serves` say a value is derived?** Yes, and the mesh performs the derivation — but it
learns no protocol doing it. The only fact is the identity the mesh itself minted, in one of two
alphabets it already knows.
- **Should a consumer that names the resource be refused?** Yes. A consumer's file that already
contains the value the mesh is about to derive for it is refused at resolution, naming the
placeholder to write instead. That is the check this report asked for, and it is exact rather than
heuristic: a derived value carries the identity minted for this consumer on this machine, which
nothing else would spell out.
minio's `bucketFor` is gone; its manifest serves `"bucket": "${consumer:as:dns}"`. The three
consumers' hand-written bucket names are gone with it — each of them also named the machine the
module happens to run on, which is the second thing wrong with a transcription.