Files
atelier/domains/messaging/first-principles.md
T
Jon Chery f61dffbb5a docs(P02): complete messaging domain phase — v0.4
---ci---
project: atelier
phase: 2
milestone: v0.4
status: complete
phase_role: execution
phase_tag: v0.3.2
requirements:
  covered: [ATELIER-97, ATELIER-98, ATELIER-99, ATELIER-100, ATELIER-101]
  partial: []
---/ci---
2026-08-05 16:00:31 +00:00

307 lines
18 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Messaging — First Principles
## 1. The Principles
### P1. Messages are Contracts
A message has an explicit, versioned schema. Producer and consumer
agree on shape before exchange; the schema is the boundary, not a
guess. A schemaless message — a free-form JSON blob the consumer
parses by hope — is a defect: the consumer breaks silently on the
next shape change, and the producer has no contract to evolve
against. This derives from `C1 Correctness` (the exchange must
carry what the parties agreed to) and `C2 Clarity` (the schema
makes the boundary obvious to both sides). This is the
cross-process expression of the contract discipline that
`domains/api/rest` owns for synchronous request/response: the
message schema is to async exchange what the API contract is to
sync exchange. It is distinct from
`domains/concurrency/patterns` Pattern 1 (Message Passing), which
owns the *in-process* channel primitive — here the contract spans
separate systems and survives network failure (D-062). See
`domains/messaging/queues.md` for the queue-flavored application
and `domains/messaging/streams.md` for the durable-log-flavored
application.
### P2. Ordering is a Property, Not an Assumption
Ordering guarantees — per-partition strict, global, or none — are
explicit and documented. "It's FIFO" is a claim that must be backed
by the broker's partitioning contract, not an assumption the
consumer makes and the broker may not honor. A standard queue
delivers in arrival order per receive-node but offers no global
ordering across shards; a FIFO queue delivers strict per-message-
group order but at a latency cost; a partitioned stream delivers
strict per-partition order but only within a partition. Each is a
distinct, declared property. This derives from `C1 Correctness`
(order is a correctness property — a consumer that assumes order
the broker does not provide is wrong) and `C2 Clarity` (the
ordering guarantee is documented, not discovered in production).
This is distinct from in-process ordering, which
`domains/concurrency/patterns` Pattern 1 owns for channels within
one program: messaging ordering survives network failure, broker
restart, and consumer crash-and-retry — a stronger failure model
than thread-local channels (D-062). The `messaging-shared-
subscription` chaos anti-pattern breaches this rule: two consumers
sharing one subscription break per-consumer ordering because the
broker dispatches each message to an arbitrary consumer. See
`domains/messaging/streams.md` for the partition-order contract
and `domains/messaging/delivery-semantics.md` for the interaction
of ordering with the three delivery semantics.
### P3. Consumers are Idempotent
Delivery is at-least-once by default across the network; a
consumer deduplicates via idempotency keys or deterministic
processing. "Exactly-once" is idempotency plus at-least-once, not a
broker guarantee — Jepsen analyses of Kafka, RabbitMQ, and NATS
establish that exactly-once claims require independent
verification, and the durable engineering practice is to make
consumers idempotent under redelivery. A non-idempotent consumer
under at-least-once delivery doubles the effect on every retry; a
non-idempotent consumer under a claimed exactly-once broker is a
bug waiting for the broker's exactly-once invariant to break. This
derives from `C1 Correctness`: correctness under redelivery is the
contract, not a nice-to-have. This parallels
`domains/edge/P5 Edge Operations are Idempotent` (the
cross-partition device-and-cache-flavored analog) and is the
cross-process instance of the retry-safety discipline that
`domains/concurrency/patterns` Pattern 6 (Timeout on Every Block)
implies for in-process retry. It is distinct from in-process
retry because the redelivery comes from the broker across a
network, not from an in-process loop (D-062). See
`domains/messaging/delivery-semantics.md` for the idempotency-key
dedup-store pattern.
### P4. Delivery Semantics are Explicit
At-least-once / at-most-once / exactly-once is a declared choice
per channel, not an emergent behavior. The tradeoff — latency cost,
implementation complexity, operational cost — is made consciously
and documented. At-most-once is fire-and-forget (low latency, lossy);
at-least-once is acked with possible duplication (the default,
requires idempotent consumers per P3); exactly-once is at-least-once
plus idempotency or a transactional two-phase commit (highest cost,
narrowest fit). An unstated semantic is a defect: the consumer
guesses, and the guess is wrong under the first failure. This
derives from `C1 Correctness` (the chosen semantic must hold) and
`C2 Clarity` (the tradeoff is visible to the reader and the
operator). This is the cross-process analog of the explicit-failure-
mode discipline that `domains/errors/patterns` owns for synchronous
code — messaging makes the delivery-mode choice as explicit as an
error-handling choice. See `domains/messaging/queues.md` for the
three-semantics comparison table and `domains/messaging/delivery-
semantics.md` for the correctness properties of each.
### P5. Dead-Letter Handling is Defined
Poison messages — unparseable, repeatedly failing, or exhausting
the retry budget — are routed to a dead-letter queue, not retried
forever or silently dropped. The DLQ is observable and drainable: an
operator can inspect it, replay from it, or discard with audit. An
unbounded retry loop is a livelock: the consumer never makes
progress past the poison message. A silent drop is a correctness
defect: the message vanished with no record. This derives from `C1
Correctness` (poison messages must not livelock the consumer or
silently disappear) and `C5 Reversibility` (the DLQ is the
reversibility mechanism — a dead-lettered message can be reprocessed
after the bug is fixed). This is the cross-process analog of the
bounded-error discipline that `domains/errors/patterns` owns for
synchronous code: a poison message is an error-as-data instance
that must be observable and recoverable, not swallowed. It is
distinct from in-process error handling because the failure spans a
network and a consumer restart (D-062). See
`domains/messaging/delivery-semantics.md` for the dead-letter
strategy comparison table and the DLQ routing rule pattern.
### P6. Backpressure is Bounded
A slow consumer cannot unbounded-buffer the broker or the
producer. Backpressure is explicit: consumer lag is visible,
max-unacked is bounded, the retry budget is capped. A consumer
that falls behind without a visible signal is a silent backlog —
the operator cannot fix what they cannot see, and the broker's
memory grows without bound until it fails. This derives from `C1
Correctness` (a backlog that grows until OOM is a correctness
failure) and `C8 Economy` (the broker's memory is bounded by
design, not by luck). This is distinct from
`domains/concurrency/P9 Bounded Queues` and
`domains/concurrency/patterns` Pattern 5 (Bounded Queue with
Backpressure), which own the *in-process* analog: concurrency's
bounded queue fails by OOM or thread crash; messaging's bounded
backpressure fails by network partition, broker restart, or
consumer crash-and-retry (D-062). The Reactive Streams
specification (`request(n)`, `onNext` bounded) is the in-process
instance; messaging's broker-backed backpressure is the
cross-process instance above it. See `domains/messaging/queues.md`
for prefetch and max-unacked and `domains/observability/metrics`
for consumer-lag as an alert.
### P7. Partitioning is Intentional
The partition key determines ordering, parallelism, and hotspots.
Key choice is a design decision with documented rationale, not a
default. A key that hashes unevenly creates a hot partition that
limits throughput; a key that does not match the ordering need
breaks per-key semantics; a key that is too coarse (one partition
for the whole topic) serializes all the traffic. The partition
count is a capacity bound: too few partitions cap parallelism, too
many partition overhead the broker. This derives from `C4 Locality`
(ordering and parallelism are co-located with the partition) and
`C6 Composability` (the partition is the unit of parallelism and
scaling — consumer groups compose from per-partition workers).
This is the cross-process analog of the locality discipline that
`domains/performance/` owns for generic data-near-compute
optimization: performance's locality is algorithmic (data near
compute); messaging's locality is partitional (order and
parallelism near the partition). See `domains/messaging/streams.md`
for the partitioned-log model and consumer-group rebalance
strategies.
### P8. Replay and Retention are Configured
Retention windows and replay-from-offset are explicit. A message
is not ephemeral by default; the broker is a durable log, not a
pipe. A topic with no retention is a fire-and-forget stream — a
consumer that falls behind loses data permanently; a topic with
infinite retention is an unbounded log — the broker grows until
disk exhaustion. Both are defects: the retention window is a
declared bound, and replay-from-offset is the mechanism that makes
the log durable (re-consumable) rather than ephemeral. This derives
from `C5 Reversibility` (a retained message is reversible — it can
be re-consumed; an ephemeral message is not) and `C7 Observability`
(the durable log is itself an observable record of what happened —
the offset is the position from which to replay). This is the
foundation for `domains/messaging/streams.md` and the rule that
distinguishes a stream from a queue (a queue deletes on ack; a
stream retains for replay). See `domains/messaging/pubsub.md` for
the pub/sub-vs-stream durability boundary.
### P9. Schemas Evolve Compatibly
Schema changes are backward- and forward-compatible by
construction. Breaking changes are versioned migrations, not
silent shape edits. A producer that ships a new field the old
consumer ignores is backward-compatible; a consumer that handles a
missing field the new producer omits is forward-compatible. A
silent schema change — the producer renames a field and the
consumer parses `undefined` — is a P1 violation (the contract was
broken) compounded here as an evolution defect. This derives from
`C5 Reversibility` (a schema change is reversible by versioning —
the old shape is still readable) and `C6 Composability` (producers
and consumers of different versions compose because the schema
evolves compatibly). This parallels `domains/data/migrations`
(schema migration for databases) and `domains/api/versioning`
(API contract evolution): messaging's schema evolution is the
async instance of the same compatibility discipline. See
`domains/messaging/streams.md` for the stream-schema-evolution
angle.
### P10. Messaging is Observable
Consumer lag, DLQ depth, throughput, and consumer-group health are
first-class signals. Silent backlog is a bug, not a feature: a
consumer that falls behind with no lag metric is invisible until
the downstream effect surfaces — by which time the backlog may be
hours or days. A DLQ that grows without an alert is a silent
correctness defect: poison messages are accumulating and no one
knows. This derives from `C7 Observability` (the broker's behavior
is visible to the operator) and `C1 Correctness` (backlog
detection is a correctness bound — unbounded lag is a failure).
This is distinct from `domains/observability/metrics`, which owns
*generic* structured metrics; messaging owns the *broker-specific*
signals — lag, DLQ depth, partition imbalance, consumer-group
rebalance events. See `domains/observability/metrics` for the
generic SLI/SLO discipline and `domains/observability/tracing` for
cross-partition traces.
## 2. Core Principle Trace
Each messaging P-rule derives from one or more core C-rules
(C1C8). The matrix extension lands in P4 of the v0.4 plan; the
traces below are authoritative. Messaging is a broad-derivation
domain touching 7 of 8 core principles (C1, C2, C4, C5, C6, C7,
C8); C3 (Simplicity) is not a primary derivation — messaging is
inherently a tradeoff domain where simplicity yields to the
correctness of delivery guarantees (a simpler-than-necessary
delivery model does not handle the failure cases, per C3's
"simpler than necessary is also a violation").
| P-rule | Core | Why |
|--------|------|-----|
| P1 Messages are Contracts | C1, C2 | Correctness of the exchange; clarity of the schema boundary |
| P2 Ordering is a Property, Not an Assumption | C1, C2 | Correctness of order; clarity of the guarantee |
| P3 Consumers are Idempotent | C1 | Correctness under redelivery |
| P4 Delivery Semantics are Explicit | C1, C2 | Correctness of the chosen semantic; clarity of the tradeoff |
| P5 Dead-Letter Handling is Defined | C1, C5 | Correctness of poison-message routing; reversibility of reprocessing |
| P6 Backpressure is Bounded | C1, C8 | Correctness of bounded backlog; economy of broker memory |
| P7 Partitioning is Intentional | C4, C6 | Locality of order; composability of parallelism |
| P8 Replay and Retention are Configured | C5, C7 | Reversibility of replay; observability of the durable log |
| P9 Schemas Evolve Compatibly | C5, C6 | Reversibility of schema changes; composability of versions |
| P10 Messaging is Observable | C7, C1 | Observability of lag/DLQ; correctness of backlog detection |
## 3. What Violates These Principles
| Violation | Principle Breached |
|-----------|-------------------|
| Schemaless message (no versioned contract; consumer parses by guess) | P1 Messages are Contracts |
| "It's FIFO" with no documented partition contract | P2 Ordering is a Property, Not an Assumption |
| Non-idempotent consumer under at-least-once delivery | P3 Consumers are Idempotent |
| Unstated delivery semantic (at-least-once vs exactly-once guessed) | P4 Delivery Semantics are Explicit |
| No dead-letter queue (poison message retried forever or silently dropped) | P5 Dead-Letter Handling is Defined |
| Unbounded retry budget (no cap; slow consumer stalls the partition) | P6 Backpressure is Bounded |
| Default partition key (no rationale; hotspot or wrong-order) | P7 Partitioning is Intentional |
| Ephemeral broker (no retention; no replay) | P8 Replay and Retention are Configured |
| Silent schema change (producer breaks consumers with no version bump) | P9 Schemas Evolve Compatibly |
| Silent backlog (no lag metric; consumer falls behind invisibly) | P10 Messaging is Observable |
| Shared subscription (two consumers share one subscription; per-consumer ordering breaks) | P2 Ordering is a Property, Not an Assumption (P3 compounding) |
| Blocking consumer (slow downstream call with no timeout; broker redelivers to the stuck consumer) | P6 Backpressure is Bounded |
## 4. Relationship to Other Domains
Messaging systems are the engineering discipline of
**cross-process, cross-system asynchronous communication via
brokers**. Producer and consumer are separate systems; the broker
is the intermediary that brokers delivery, ordering, retention,
and failure semantics. The distinguishing constraints are a
cross-process failure model (network, not crash), explicit
delivery semantics, decoupled producer/consumer lifecycle, and
replay-and-retention as a durable-log property. Messaging overlaps
`domains/concurrency/` by *subject* (messages, queues,
backpressure) but not by *failure model*: per D-062, messaging
owns the cross-process/network-failure-model angle; concurrency
owns the in-process/crash-failure-model angle. The discriminator
is the failure model: concurrency's queue fails by OOM or thread
crash; messaging's queue fails by network partition, broker
restart, or consumer crash-and-retry. Messaging extends
concurrency's bounded-queue/backpressure model to the network-
partition regime. Cross-links are one-directional outward (per
D-026 extended); no back-link edits to v0.1/v0.2/v0.3 content.
- `domains/concurrency/patterns` ← P6 (the broker-backed bounded
queue is the cross-process analog of the in-process bounded
buffer — concurrency Pattern 5 owns in-process; messaging owns
the network-failure-model instance above it, per D-062)
- `domains/concurrency/patterns` ← P3 (idempotent retry is the
cross-process analog of in-process retry-safety — the failure
model differs: broker redelivery across a network vs in-process
loop)
- `domains/observability/metrics` ← P10 (consumer lag and DLQ
depth as alerts; observability owns the generic SLI/SLO
discipline, messaging owns the broker-specific signals)
- `domains/observability/tracing` ← P10 (cross-partition traces
for stream processing; observability owns the generic tracing
discipline, messaging owns the cross-partition propagation)
- `domains/data/schema-design` ← P1, P9 (message schema design
and evolution; data owns the generic schema discipline,
messaging owns the cross-process message-shape instance)
- `domains/errors/patterns` ← P5 (errors as data for message
failures; a poison message is an error-as-data instance that must
be observable and recoverable, not swallowed)
- `domains/edge/iot` ← P4 (the edge↔messaging cross-link
resolves bidirectionally here: edge/iot.md links outward to
messaging/queues for MQTT QoS parallels to delivery semantics;
this first-principles doc acknowledges the back-link — the
edge/iot.md → messaging/queues link from P1 now resolves because
messaging/queues.md exists, completing the bidirectionality per
IDEATE-40)
> Note: the edge/iot.md → messaging/queues cross-link (MQTT QoS
> parallels for delivery semantics) was authored in P1 with a
> dangling reference; this P2 authorship of messaging/queues.md
> resolves it. The bidirectionality is verified in P5
> (ATELIER-114 per IDEATE-40). The cross-link is one-directional
> outward from edge/iot.md; this first-principles doc
> acknowledges the resolution without editing edge/iot.md (per
> D-026 extended — no back-link edits to v0.1/v0.2/v0.3 or to
> P1-authored edge content).