M6.4 — GetDesignStream client fan-out (3D + info clients)
| Status | 3D client IMPLEMENTED + REVIEW-FIXED 2026-11-01 — PR intel-agency/odbdesign-3d-client-prototype#67 (branch feat/getdesign-stream, 2299e6b + review-fix 6dd8592; suite 1248/1248 post-fixes, no new warnings — ~35 pre-existing nullable warnings live in untouched files; live stream≡monolithic equivalence gate passed against the merged-tip server AND independently reproduced by the review, which also re-ran the suite). Three-lens review of 2299e6b: 13 reported → 9 unique, all fixed in 6dd8592 — connectivity.proto byte-sync (the 28th proto, stale since the #62-era sync), empty-stream header-seen guard (surfaced as DataLoss, not Internal — Internal is in the transient-retry set), OCE-shape conversion for mid-stream generation cancel, bounded stream-only transient retry (R-4 intact), null-Step present-but-empty parity, StreamTimeoutSeconds docs, DESIGN_STREAM_SMOKE_DESIGN smoke parametrization (Panel-g7162 is not in local TEST_DATA — zero-connectivity live coverage deferred; unit tests cover empty-batch semantics). Awaiting CI/review on 6dd8592. Server #609 merged (99f03f3), deployed + live-accepted 2026-11-02 (nam20485-1287, digest 29a4501d). Info client DONE 2026-11-02 — #27 merged (8679ac8), promoted nam20485→development→main (#28/#29, c98a179); proto-sync only. A stream fallback remains owed to the info client — the §4 deferral premise was wrong and is corrected below (§2 recon note + §4). |
| Parent | getdesign-streaming-oom-plan.md (M6.4 definition, §M6.5 gates, chunk contract) · component-connectivity-client-followups.md (F1–F7 client backlog; F6 = proto upkeep) |
| Split | 3D client → the OdbDesign session that authored #609 (this doc §3). Info client → separate session (this doc §4 + handoff prompt). FixtureCapture --stream stays in this repo and rides the M6.6 PR. |
1. Driver
The #609 dispatch guard rejects over-budget monolithic GetDesign with RESOURCE_EXHAUSTED
(“use the GetDesignStream RPC” — OdbDesignServiceImpl.cpp:272-278). The 3D client renders from the
full Design.fileModel fetched via monolithic GetDesign, so without adopting the stream it cannot
load the two big designs (200-40628, 350-41017) from any post-#609 server. No regression — those
designs OOM today’s server — the update converts the server fix into a usable client feature.
2. Recon (verified 2026-11-01)
3D client — intel-agency/odbdesign-3d-client-prototype
- Clean tree on
development; PR #65 merged; no open PRs (at recon time — #67 opened later the same day; the local checkout now sits onfeat/getdesign-stream). Branch stylefeat/*; remote is the intel-agency org (CI there randomly cancels runs — rerun before re-diagnosing). GetDesignAsyncatsrc/OdbDesign3DClient.Core/Services/Implementations/OdbDesignGrpcClient.cs:122: client-side design cache (hit / in-flight join / generation-canceled eviction).RpcException … ResourceExhaustedhandler at:224currently logs + rethrows — that is the fallback insertion seam.- Streaming precedent exists (
GetLayerFeaturesStream,StreamTimeoutSecondsconfig); real-server smoke pattern exists (tests/…/Rendering/SolidBoardRealServerSmokeTests.cs). - Proto drift: all 27 message protos differ cosmetically only (
option cc_enable_arenas, comments) except two semantic gaps shared with the info client (below). No C# code referencesComponentRecordsByName(grep-clean) → wholesaleprotoc/sync is safe, no code fallout beyond regenerate. service.protomissing:GetDesignStream+DesignChunkcontract,RequestLoadDesign(+ messages).
Info client — nam20485/Odbdesign-info-client-india79-b
(Corrected 2026-11-01, post-execution. The original first bullet said “REST-first; gRPC used only for Connectivity” — wrong even at recon time: design reads went gRPC with #23 on 2026-10-15. That error is what produced the bad §4 deferral.)
- Design reads are gRPC (#23, 2026-10-15):
DesignService.GetDesignArchiveAsync→IOdbDesignGrpcClient.GetDesignAsync(src/OdbDesignInfoClient.Services/DesignService.cs:413) — the method’s own doc-comment calls the gRPCGetDesignresponse “the single content source”. REST covers the design list (GetDesignNamesAsync,:114) and auth. Codegen wired inOdbDesign.ProductModel.csproj. Default branchmain. - Branch flow resolved (was “needs untangling”): #25 (F6) and #26 (F4/F5) both merged to
nam20485; #27 cut from8aa2b95; promotednam20485→development(#28,72f700d) →main(#29,c98a179, 2026-11-02T00:42Z).mainis current; residue is stale localfeature/*branches only. ComponentRecordsByNameverified (was “unverified”): zero hand-written C# consumers — generatedobj/code only — so #27’s proto sync was regenerate-only, no code fallout.
Shared semantic proto gaps (vs server OdbDesignLib/protoc/)
componentsfile.proto— client still declaresmap<string, ComponentRecord> componentRecordsByName = 10;; server deleted + reserved it (never populated by the parser; conformance rule 1: clients index refDes → record locally).edadatafile.proto— server addedoptional uint32 id = 7(NET;ID=) andid = 13(PKG) from the connectivity epic; clients lack them.
3. 3D client scope (ONE PR: feat/getdesign-stream from development) — EXECUTED as #67
- Proto sync (wholesale): copy
OdbDesignLib/protoc/*+OdbDesignServer/protoc/grpc/service.protofrom the post-mergenam20485tip (pre-merge: thenam/getdesign-streamingtip — proto content is frozen, PR approved 272/272). Kills the cosmetic noise so future diffs are meaningful; brings the two semantic fixes +GetDesignStream. - Fallback in
GetDesignAsync: on anyResourceExhaustedfrom the monolithic call, fall back once toGetDesignStreamand reassemble, feeding the existing cache path. Implemented as any-RE rather than matching the guard’s detail only: both RE flavors (dispatch-guard refusal and “larger than max”) are size failures, and bounded chunks remedy both. Pre-streaming servers (Unimplementedon the stream) surface the original refusal. Server decides monolithic-vs-stream; client never guesses size. (Two existing contract tests that pinned the pre-fallback fail-fast behavior were updated: still exactly one monolithic attempt — R-4 — plus one chunk-sized fallback attempt.) - Reassembler per the contract comments in
service.proto:146+(authority):DesignHeaderscalars once;StepSkeletoninsert intofileModel.stepsByName[step_name];FeatureBatchappend into…layersByName[layer_name].featureFile.featureRecords; Design-level repeated batches append in order;ConnectivityBatchMergeFromintoDesign.connectivity;NormalizedMapBatchmaps merge.- Resolve map entries by iteration, never
find()/at()— the #599 address-layout-dependent protobuf-Map bug; the proto comments call this out for client implementers. - Sanitization-colliding step/layer names arrive with the monolith’s last-wins semantics (server guarantee) — reassembler appends/overwrites, never dedupes.
- Tests: unit reassembly from crafted chunk sequences (incl. empty/cancelled mid-stream); equivalence — stream-reassembled Design deep-equals monolithic
GetDesignfor the same flavour onsample_design+ a zero-connectivity design (Panel-g7162); fallback unit test with a mocked channel throwingResourceExhausted; real-server smoke followingSolidBoardRealServerSmokeTests. - Local server for tests: build merged-tip OdbDesign (
linux-dynamic-releasepreset,VCPKG_ROOTexported), run with--designs-dir ../OdbDesignTestData/TEST_DATA(relative path —CommandLineArgstreats leading/as a flag).
Gates: full sln build + suite green (1154+ baseline) + new tests green.
4. Info client scope (separate session, ONE PR) — EXECUTED 2026-11-02 as #27
Shipped: proto sync from server 99f03f3 (message protos + service.proto) + regenerate. Original item 3 collapsed to nothing — the ComponentRecordsByName grep came back clean (§2), so no code fix or targeted test was needed. Branch cut from 8aa2b95 (both #25/#26 already merged — the original “cut from the tip carrying 0e46581” instruction described the pre-#26 picture). Suite 119/119, all CI legs green, merged 8679ac8, promoted to main via #28/#29 (c98a179).
Deferral rationale CORRECTED 2026-11-01. The original “Out of scope (deliberate)” paragraph deferred a GetDesignStream reader on two premises, both stale:
- “the info client has no gRPC design-read consumer” — false since #23 (2026-10-15):
DesignService.GetDesignArchiveAsync→IOdbDesignGrpcClient.GetDesignAsync(DesignService.cs:413) is the design-read path. - “reassembly lands with F3 (REST→gRPC transport switch)” — F3 had already shipped (#23); #27 was its proto refresh, and no reassembly landed with it.
The deferral’s outcome stands as shipped, but it is remaining work, not dead-code avoidance: on any post-#609 server, GetDesignAsync on the two big designs (200-40628, 350-41017) fails with a clean RESOURCE_EXHAUSTED, and the info client currently swallows it — FetchDesignArchiveAsync’s catch-all logs a warning and returns null (DesignService.cs:429-434), so the design renders empty with no user-visible error and no retry. When big-design support is wanted, mirror 3D #67: any-ResourceExhausted → single GetDesignStream fallback → reassembler per the contract comments in service.proto:146+ (iterate maps, never find()/at() — the #599 landmine). While in there, stop nulling genuine transport failures — surface them.
5. Sequencing — updated 2026-11-01
Proto content is frozen → client branches can be cut immediatelydone: info #27 merged + promoted; 3D #67 open (green, MERGEABLE).#609 merge → docker publish → rollout → live acceptanceDONE 2026-11-02: merged99f03f3→ publishednam20485-1287→ rolled out (pod on digest29a4501d…, 00:21Z) → live acceptance passed (big-design stream OK at 3.31 GiB HWM; over-budget monolithic = cleanRESOURCE_EXHAUSTED).- Remaining: review + merge 3D #67 → M6.6 (
nam/m66-residency-budget, cut from99f03f3) with FixtureCapture--streamriding it → info-client stream fallback (§4) whenever big-design support is wanted there.