Files
Ekklesia/docs/architecture/decisions/0042-bounded-streaming-request-bodies.md
T
ekklesiaandClaude Opus 5 e36485bf99 feat: let a module route stream an image-sized upload, explicitly and bounded
ADR 0017 capped a module route at 1 MiB and said large media belonged to a
future storage design. That design arrived — ADR 0030 gave modules a bucket,
ADR 0031 took Ekklesia out of the read path — and it did not solve the write
path: signed write URLs were deliberately declined, so the only way bytes
reach a bucket is a module's own route, and a photograph is not 1 MiB.

Extend the route contract rather than adding a second HTTP system. A route
declares `bodyMode`: `buffered` (the default, and what every existing route
already meant) keeps the 1 MiB ceiling and materializes the body before the
handler runs; `streamed` raises the ceiling to 32 MiB and counts bytes as the
handler consumes them, holding nothing. The mode is explicit rather than
inferred from the size, because inferring it would mean an author who raised a
limit silently changed how their own handler receives its body and when a 413
can reach their own code.

32 MiB holds one 25 MiB image plus bounded overhead and is also the platform's
absolute ceiling: no route in any mode may declare more, and no unbounded
option exists in the contract. The unit is one file per request — twenty
images are twenty individually bounded requests, and nothing here can express
a 500 MiB aggregate whose failure would waste everything already sent.

Replace the buffering dependency with the platform's own limiter. A declared
Content-Length over the limit is refused synchronously, before authorization
and before a byte is read. Buffered mode now counts what actually arrives
rather than trusting the header, which closes a gap the previous
implementation had. Streamed mode wraps the body in a counting transform;
on overflow the stream errors, the source is cancelled, and the response is
replaced with the same sanitized 413 even if the handler swallowed the error —
a limit a handler can swallow is not a limit. An unread body is cancelled so
an abandoned upload releases its socket, and an abort stays an ordinary failed
request rather than becoming a 413.

Ordering is unchanged and now strictly better: in streamed mode no byte is
read until the handler reads it, and the handler runs only after authorization
succeeds, so an unauthorized caller cannot push an upload through at all.

The live suite uploads a real 25 MiB body over a real socket end to end.

Co-Authored-By: Claude Opus 5 <[email protected]>
2026-08-27 15:57:44 -05:00

8.9 KiB

0042: Explicit, bounded streaming request bodies

Context

ADR 0017 gave every public route an enforced request-body ceiling and capped a module route at 1 MiB, with an explicit note that "large media transfer belongs to a future dedicated storage design, never to widening this ceiling."

The storage design arrived — ADR 0030 gave modules a scoped object store, ADR 0031 took Ekklesia out of the read path with signed delivery URLs — and it did not solve the write path. ADR 0031 deliberately declined to offer signed write URLs, so the only way bytes reach a module's bucket is through that module's own HTTP route. That route is capped at 1 MiB, and an ordinary photograph is not 1 MiB. A church uploading a slide for a lobby screen cannot, today, get the file in.

Two properties of the existing implementation also need naming, because they are what the change has to preserve and what it has to fix:

  • The pinned body-limit middleware fully buffers any body with no declared Content-Length before deciding. At 1 MiB that is a bounded, acceptable cost. At image sizes it would mean the platform holding every concurrent upload in memory.
  • With a declared Content-Length within the limit it counts nothing at all, trusting the header. Node's HTTP framing makes that safe today, but it is trust the platform does not need to extend.

Decision

Extend the existing route contract. There is no second HTTP system, no upload endpoint, no signed write URL, and no new middleware stack — one route declaration, one limiter, one 413.

A route declares how its body reaches it

ModuleRouteDefinitionInput gains one optional field:

http.route({
  method: "POST",
  path: "/uploads",
  permission: "signage.slides.write",
  maxBodyBytes: 33_554_432,
  bodyMode: "streamed",
  handler,
});
  • buffered (the default, and what every route written before this meant) — the platform reads and counts the whole body before the handler runs, so the handler is promised a fully available, bounded body and an oversized request is rejected before any authorization or handler work happens. Ceiling: 1 MiB, unchanged.
  • streamed — the platform counts bytes as the handler consumes them and never materializes the body. Ceiling: 32 MiB.

The mode is explicit rather than inferred from the size. Inferring it — "over 1 MiB means streamed" — would mean an author who raised a limit silently changed how their own handler receives its body, and when a 413 can arrive relative to their own code. That is exactly the class of invisible change this repository refuses.

Omitting the field means buffered, so no existing route changed behaviour, and the canonical parsed form always carries a mode so nothing downstream re-applies the default or disagrees about it.

The ceilings, and why 32 MiB

MODULE_ROUTE_MAX_BODY_BYTES_CEILING stays at 1 MiB for buffered bodies: a body the platform materializes is a body the platform is holding, and 1 MiB is the most a JSON request has any business being.

MODULE_ROUTE_STREAMED_MAX_BODY_BYTES_CEILING is 32 MiB, sized to hold one 25 MiB image plus bounded overhead — a multipart envelope, its boundaries and part headers, and room for a client that pads. It is also MODULE_ROUTE_ABSOLUTE_MAX_BODY_BYTES: no route, in any mode, may declare or accept a larger single request body, and no unbounded option exists anywhere in the contract. A route declaring more than its mode's ceiling fails registration, exactly as an over-ceiling value always has.

The unit is deliberately one file per request. A UI offering twenty images issues twenty individually bounded requests, each of which the platform can reject cheaply and each of which fails alone; a 500 MiB aggregate request would make one bad byte waste everything already sent, and nothing in this contract can express one.

Enforcement, in one limiter

The platform's own limiter replaces the buffering dependency. It is the same code for both modes and differs only in when it counts:

  1. A declared Content-Length above the limit is refused synchronously — before authorization, before the handler, before a single body byte is read. That is the cheap rejection an ordinary too-large upload takes. A malformed or contradictory header is treated as absent rather than as a rejection, because the byte counting bounds the request either way and refusing could only turn a legitimate request away.
  2. buffered: bytes are counted as they arrive and the request is rejected the moment the running total passes the limit — so a lying or absent Content-Length is bounded by what actually arrived rather than by what was claimed, which the previous implementation did not do. At most the route's own declared allowance is ever held.
  3. streamed: the body is wrapped in a counting transform and handed to the handler. The moment the total passes the limit the stream errors, the source is cancelled, and the response is replaced with the same sanitized 413 — even if the handler caught the error and answered anyway. A limit a handler can swallow is not a limit.

A body no handler read is cancelled after the request, so an abandoned upload releases its socket rather than draining. An aborted upload surfaces as an ordinary failed request, never a 413: nothing exceeded a limit.

Ordering is unchanged and remains deliberate: trusted-origin check, then the body contract, then permission evaluation, then the handler. In streamed mode no body byte is read until the handler reads it and the handler runs only after authorization succeeds, so an unauthorized caller can never push an upload through the platform at all — strictly better than the buffered path.

Timeouts stay where they were: the server's own headersTimeout/requestTimeout. A slow upload is a slow request, and the platform already has one answer for that rather than a second per-route one.

What a module author must know

In streamed mode a 413 can arrive after the handler has started. A handler that persists as it reads must be prepared for a partially consumed upload — write to a temporary key and commit only after the read completes. The platform cannot do that for the module, because only the module knows what a partial write means to its own data.

Consequences

  • An ordinary image upload works, through the route contract a module already had, with the platform holding none of it.
  • The buffered path now counts bytes it previously trusted a header for, closing a gap rather than opening one.
  • Every existing route and every existing test is unchanged: the default is the old behaviour, and the 413 contract, ordering, and sanitization are identical.
  • apps/api owns ~60 lines of stream plumbing it previously delegated. That is the price of a limiter that neither buffers a large body nor can be swallowed by a handler, and it is enforced by tests rather than by reading the dependency's source.
  • Media larger than 32 MiB — a video — still has no path in. That is deliberate and named below.

Alternatives

  • Raise the single ceiling to 32 MiB. Rejected: every existing route would silently gain the ability to have 32 MiB buffered in memory on its behalf, and await request.json() in a module handler would stop being a bounded operation.
  • Infer streaming from maxBodyBytes. Rejected: it makes a size change silently change delivery semantics and error timing for the author's own code.
  • A separate upload endpoint or second HTTP system. Rejected: a second route table, a second authorization path, a second CORS and origin policy, and a second failure vocabulary, to solve something the existing route contract can express with one field.
  • Signed write URLs to object storage. Deliberately still deferred (ADR 0031). A signed write URL is a bearer capability to create an object, which needs quota, content-type, and lifecycle answers this pass does not have — and a module that wants to validate, transform, or authorize per byte cannot use one anyway.
  • Let the handler enforce its own limit. Rejected: a limit enforced by the party being protected against is not a limit, and every module would reimplement it slightly differently.
  • Trust Content-Length in streamed mode. Rejected: the running byte count costs nothing and removes the need to reason about framing at all.

Deferred

Signed write URLs, resumable or chunked multi-request uploads, per-module or per-installation upload quotas, content-type or magic-byte validation, virus scanning, image transformation, progress reporting, and any body larger than the absolute ceiling — a video pipeline is its own design, not a larger number here.