catalogue: Huawei OptiXtrans E6608T — 7x TMB1LDCD terminal shelf (28x 100GE client, 14x 200G line) #1739
No reviewers
Labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
jaylouisw/packetforge!1739
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "catalogue/1116-2155"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Fixes #804
Adds
devices/huawei/optixtrans-e6608t.json— OptiXtrans E6608T (7x TMB1LDCD, DC chassis): 28x 100GE QSFP28 client + 14x CFP2 200G line.Evidence
optixtrans-e6608.json(6x) andoptixtrans-e6616.json(14x).What the request got wrong, corrected here
The T is not an E6608 with a smaller name: it is "OTN equipment supporting only OTN services" with no cross-connect (Huawei's summary: "7 slot @ 2U, Ponder + Optical"). So
capabilitiesisclionly, there is no OTN/PKT/SDH unified switching, and the entry carries nofabricblock — the seven boards are independent transponders that cannot groom between each other.Gates
node scripts/syntax-check.js— 364 files, all OKnode scripts/audit-port-types.js— "No port contradicts its declared type"node test/hardware-invariants.test.js— 26 passed, 0 failed (2332/2332 distinct keys)node scripts/check-control-bytes.js— cleandocs/DEVICE_REGISTRY.mdregenerated (never hand-edited): 2,332 models, 28,500 sockets, 951 port shapesThis branch conflicts with
main, so it was not rebased:A conflict is a decision about behaviour rather than a mechanical fix, so this is left for a human. Rebasing by hand and force-pushing will let the automation pick it up again.
Surgery run 2026-11-16 22:08 — this PR is red for one reason: three hard-coded counts in
test/now see a third TMB1LDCD shelf.Local run of the head
9dde979(detached,node scripts/run-tests.js --concurrency=8): 293 suites, 10,395 assertions, 3 failed,audit-port-typesclean,duplicate-models21/21. The three failures are all the same cause — this PR addsOptiXtrans E6608T (7x TMB1LDCD, DC chassis)with aliasTMB1LDCD, so aB1LDCDquery legitimately returns 3 shelves (E6608, E6616, E6608T) where the suites assert 2. The same two suites onorigin/main(4e14afe) are green (13/13, 19/19), so it is this PR's content, not main.Which lines (all mechanical count/ranking updates, no behaviour question):
test/device-aliases.test.js:95ok(board.length === 2, 'B1LDCD finds both shelves that take the board', String(board.length));→ actual
3; expectation needs to be 3 (or>= 2).test/device-aliases.test.js:99const multi = filterDevices('b1ldcd 14x', specs, {});ok(multi.length >= 1 && /E6616/.test(multi[0].model), ...)→
multi[0].modelis nowOptiXtrans E6608T (7x TMB1LDCD, DC chassis)(its label carries "14x 200G line"), so the ranking assertion no longer holds — the new shelf outranks E6616 forb1ldcd 14x.test/palette-search-keys.test.js:96ok(warm[7].length === 2, 'the B1LDCD alias still finds both shelves', String(warm[7].length));→ actual
3; expectation needs to be 3.Why I have not fixed it:
test/is outside the surgery loop's remit — I may only regenerate generated docs and resolve mechanical merge conflicts, never author source or test changes. This needs the owning loop (or Jay) to land the three updated expectations on this branch; after that the PR is ready and the armed auto-merge will take it. Nothing else in the diff is at fault — the catalogue data,README/docscount regeneration andscripts/build-wiki.jsall regenerate cleanly, and the PR is 2 commits behind4e14afewith no conflict.— surgery watchdog
Surgery run 22:23 — two new facts, still needs Jay's decision, deliberately not rebased.
Main moved twice since the diagnosis (
326ea7fc#1740,473d4c96#1741), so the branch is now 4 behind / 1 ahead. The conflict surface has widened from whatpr-rebasereported at 22:03:git merge-tree origin/main 9dde9798now conflicts in 5 files —README.md,docs/CONNECTORS.md,docs/DEVICE_REGISTRY.md,docs/wiki/Home.md,scripts/build-wiki.js. Three of those are generated (DEVICE_REGISTRY.md,wiki/Home.mdcome fromscripts/build-wiki.js), so they are regenerable — but regenerating them cannot turn this PR green.test/device-aliases.test.js:95and:99andtest/palette-search-keys.test.js:96are byte-identical onorigin/mainto what was reported:board.length === 2,warm[7].length === 2,/E6616/. The blocker is unchanged.Rebasing was skipped on purpose:
catalogue/1116-2155is checked out in the livepf-wt/firmwareworktree, so a force-push from here would desync a running loop for zero benefit — a clean rebase still fails those 3 assertions. The one decision needed is still: who lands the 3-line count bump intest/(293 suites / 10,395 assertions / 3 failed is the expected-total arithmetic).9dde9798828229c866c8Rebased onto
473d4c96→8229c866; the git conflict is gone. What is left is three assertions intest/, which are the catalogue owner's, not mine to edit.Verified after the rebase (
git merge-tree origin/main origin/catalogue/1116-2155): 0 conflicts, 0 behind, 1 ahead. The five previously-conflicting files were all count lines, so they were regenerated rather than merged by hand:node scripts/regen-registry-doc.jsrewrote README,docs/CONNECTORS.md,docs/DEVICE_REGISTRY.md,docs/wiki/Home.mdandscripts/build-wiki.jsto 2,340 models / 28,806 sockets / 958 distinct port shapes (main was at 2,339 / 28,764 / 957; the new shelf is +1 model, +42 sockets, +1 shape).test/doc-catalog-counts.test.jspasses 43/43 on the rebased tree.What still fails, and it is not a conflict — the new shelf really is a third
B1LDCDshelf, so main's hard-coded 2 is now wrong:test/device-aliases.test.js:95board.length === 2, "B1LDCD finds both shelves"test/device-aliases.test.js:99"b1ldcd 14x"→ E6616 firsttest/palette-search-keys.test.js:96warm[7].length === 2The two counts are a mechanical 2→3. The middle one is a question only the catalogue loop can answer: for
b1ldcd 14x, is E6608T now the right first hit (it is a 14x 200G line chassis) with E6616 demoted, or should E6616 keep the top slot? Either answer is a two-line test edit plus the assertion's wording.Needs the catalogue loop to land that edit on top of
8229c866(or on main — the count bump is independent of this PR). Until then this PR stays red on Test (Node 18/20/22) for a reason that is not a merge problem.scripts/anddocs/were regenerated only; no engine or behaviour change was made by me.8229c866c8855c071fd6855c071fd63fc15a0b6b3fc15a0b6baa3586bf24aa3586bf24404ff1889d404ff1889d91de30877291de30877283162ef88a83162ef88a0e790aadd80e790aadd8f614ae87b8f614ae87b89615f066c09615f066c06c6bc4ae72Three stale
=== 2counts in other suites — that is the whole red, no engine change needed.Reproduced on this head (
f614ea87) in the runner's ownnode:20-bookwormimage: 296 suites / 10,479 assertions, 3 failed, all of them the B1LDCD alias count:The new OptiXtrans E6608T really is filled with 7x TMB1LDCD, so a third shelf now takes the board and both suites' hard-coded
=== 2is stale:test/device-aliases.test.js:95—ok(board.length === 2, 'B1LDCD finds both shelves that take the board', ...)test/palette-search-keys.test.js:96—ok(warm[7].length === 2, 'the B1LDCD alias still finds both shelves', ...)test/device-aliases.test.js:98— theb1ldcd 14xcase ranks the new shelf first, where the assertion wants/E6616/; decide whether14xshould prefer the E6608T (14 line-side CFP2 200G) or the E6616, then fix the expectation or the ranking.These are count/ranking edits in cases the catalogue loop owns (
test/is not mine to change), and they are the only thing standing between this PR and a green run. Green on push ⇒ Arm auto-merge takes it.6c6bc4ae727d82a597197d82a59719e9221b9530Rebased onto main — conflicts were counts and generated docs only, resolved by regeneration, never by hand.
Head
7d82a597→e9221b95on7733b4e3.mergeableis now true.The conflicts
README.md,docs/CONNECTORS.md,docs/DEVICE_REGISTRY.md,docs/wiki/Home.md,scripts/build-wiki.js— every hunk was a catalogue count (2,343 vs 2,340) or a vendor/category row. I took main's side of all five and then rannode scripts/regen-registry-doc.js, which rewrote the values itself:No device JSON was touched;
devices/huawei/optixtrans-e6608t.jsonapplied cleanly.Verified on the rebased head
audit-port-typesOK ·check-control-bytesOK ·check-canonical-urlsOK ·check-doc-citationsOKcatalogue-hygiene 25 · faceplate-catalogue 14 · optics-catalogue 27 · duplicate-models 21 · doc-catalog-counts 43 · repo-paths 8 — all green.
CI is re-running on
e9221b95; if it comes back green this arms auto-merge as normal.Follow-up: the rebased head
e9221b95is red too — same three checks, so it was never the conflicts.Test (Node 18),(Node 20),(Node 22)all failure; classify, smoke, docker build and arm-auto-merge green. That is the identical shape this PR had at7d82a597, and the identical shape #1749 has at42852c53— three unrelated trees, three red test legs.In the same tree everything the change actually touches is green: catalogue-hygiene 25, faceplate-catalogue 14, optics-catalogue 27, duplicate-models 21, doc-catalog-counts 43, repo-paths 8, and the four static gates.
I could not read the failing log to say more:
/actions/runs/<id>/jobs/<n>/logsreturns 200 with an empty body on this build and the REST actions endpoints 404, so a red leg here is currently unauditable from the outside. Flagging it rather than guessing — the merge stays blocked on a red run by the standing rule.e9221b95307284dfc0787284dfc07892721e1114end, so a paste closes itself #176892721e1114bd32372887bd3237288742264410984226441098561d5bac48561d5bac48b91328b942Rebased
561d5bac→b91328b9ontoc6ac775b— conflicts gone, 0 behind. The red is three hardcoded totals intest/, and here are the exact lines for the catalogue owner to bump.The rebase (mechanical, generated files regenerated not hand-merged)
Only
docs/CONNECTORS.mdanddocs/DEVICE_REGISTRY.mdconflicted, both pure counts. Took main's side and re-rannode scripts/regen-registry-doc.js, which rewrote all five count surfaces itself:docs/CONNECTORS.md,docs/DEVICE_REGISTRY.md,docs/wiki/Home.md,scripts/build-wiki.js)devices/huawei/optixtrans-e6608t.jsonis byte-identical to your head. Verified locally on the rebased tree:audit-port-typesOK,check-control-bytesOK, anddoc-catalog-counts/catalogue-hygiene/faceplate-catalogue/duplicate-models/repo-pathsall green.The remaining red is 3 lines, all in files that already exist on main
test/device-aliases.test.js:95—ok(board.length === 2, 'B1LDCD finds both shelves that take the board')→ the shelf makes it 3.test/device-aliases.test.js:98—ok(/E6616/.test(multi[0].model), 'a multi-token query can mix an alias with a model word')forb1ldcd 14x:E6608T (7x TMB1LDCD, DC chassis)now ranks first.test/palette-search-keys.test.js:96—ok(warm[7].length === 2, 'the B1LDCD alias still finds both shelves')→ 3.These are catalogue-owner edits (
test/is not mine to change), not a Jay decision — except the middle one, which is the only behaviour question: shouldb1ldcd 14xstill rank the E6616 first now that a 7-board shelf also matches?Suggested durable fix: derive the expected count from the specs (or
>= 2plus "every hit really carries TMB1LDCD") so the next alias-bearing shelf doesn't red the suite. Fourth head on this PR, so I am not queuing another CI cycle myself.b91328b942027a2f215b027a2f215b1531f9da6b1531f9da6bd7adf402edd7adf402ed493554a473493554a47393291abc37Test (Node 20) is red on
93291abc— three stale hardcodes intest/, nothing wrong with the catalogue data. I reproduced locally on this exact head; the device JSON, registry regeneration and byte checks all pass.Three assertions, all in files that already exist on main and that this PR does not touch:
test/device-aliases.test.js—ok(board.length === 2, 'B1LDCD finds both shelves…'). This PR makes it 3 shelves that take a TMB1LDCD (E6608T 7x joins E6616 and E6608T 14x). Hardcoded count.test/device-aliases.test.js— theb1ldcd 14xmulti-token case asserts the first hit matches/E6616/. With the 7x shelf present, the ranking putsOptiXtrans E6608T (7x TMB1LDCD, DC chassis)first, so the "E6616 ranks first" assumption no longer holds even though both halves of the query still match something.test/palette-search-keys.test.js—ok(warm[7].length === 2, 'the B1LDCD alias still finds both shelves'). Same count, same answer: 3.The data is self-consistent —
board.lengthandwarm[7].lengthagreeing at 3 is what a correct catalogue produces. What needs deciding is test-side and belongs to whoever ownstest/: either derive these counts from the alias table instead of hardcoding them, or state explicitly that a new alias-bearing shelf is expected to update them. I don't edittest/, so this stops here.93291abc37745dc0634b745dc0634b5093e841115093e841117386758c887386758c88b8bd487cd6b8bd487cd632fa87851832fa87851843b757899743b7578997ccf73f9202ccf73f92023f6076e7173f6076e7172db9079a442db9079a444531cc33caRed on all three legs, and reproduced here — two assertions, both caused by the new device's model string.
node test/device-aliases.test.json4531cc33: 11 passed, 2 failed, and the failures are exactly the two "Against the real catalogue" cases:Cause:
devices/huawei/optixtrans-e6608t.jsonhas the modelOptiXtrans E6608T (7x TMB1LDCD, DC chassis).TMB1LDCDcontainsB1LDCD, so the board alias now matches three devices, not the two the test asserts (board.length === 2), andfilterDevices('b1ldcd 14x')ranks the E6608T above the E6616.This is a behaviour question, not a count to bump, so I am not resolving it:
TMB1LDCDthe same board asB1LDCD? If not, the model string should not embed it in a way that the alias tier treats as the board name — that is a catalogue change.=== 2and its "E6616 first forb1ldcd 14x" expectation both need widening to three — that is atest/change owned by another loop.Tell me which, or have the owning loop make the call.
docs/DEVICE_REGISTRY.md, the README count and the wiki page in this diff all regenerate cleanly, so the device itself is the only open question. Not merging red.Rebased onto
473d4c96→8229c866; the git conflict is gone. What is left is three assertions intest/, which are the catalogue owner's, not mine to edit.Verified after the rebase (
git merge-tree origin/main origin/catalogue/1116-2155): 0 conflicts, 0 behind, 1 ahead. The five previously-conflicting files were all count lines, so they were regenerated rather than merged by hand:node scripts/regen-registry-doc.jsrewrote README,docs/CONNECTORS.md,docs/DEVICE_REGISTRY.md,docs/wiki/Home.mdandscripts/build-wiki.jsto 2,340 models / 28,806 sockets / 958 distinct port shapes (main was at 2,339 / 28,764 / 957; the new shelf is +1 model, +42 sockets, +1 shape).test/doc-catalog-counts.test.jspasses 43/43 on the rebased tree.What still fails, and it is not a conflict — the new shelf really is a third
B1LDCDshelf, so main's hard-coded 2 is now wrong:test/device-aliases.test.js:95board.length === 2, "B1LDCD finds both shelves"test/device-aliases.test.js:99"b1ldcd 14x"→ E6616 firsttest/palette-search-keys.test.js:96warm[7].length === 2The two counts are a mechanical 2→3. The middle one is a question only the catalogue loop can answer: for
b1ldcd 14x, is E6608T now the right first hit (it is a 14x 200G line chassis) with E6616 demoted, or should E6616 keep the top slot? Either answer is a two-line test edit plus the assertion's wording.Needs the catalogue loop to land that edit on top of
8229c866(or on main — the count bump is independent of this PR). Until then this PR stays red on Test (Node 18/20/22) for a reason that is not a merge problem.scripts/anddocs/were regenerated only; no engine or behaviour change was made by me.0a725462c526b35b4ca8Node 18 green, Node 20 + 22 red at
26b35b4c— and the failing legs have no log left to read. Not touching this one; recording the split so it is not re-pushed blindly.Run 4794, head
26b35b4c(behind main by 0, 2 commits ahead: addsdevices/huawei/optixtrans-e6608t.jsonand updates the count assertions intest/device-aliases.test.js/test/palette-search-keys.test.js):The Node 18 log is intact and ends
Job succeeded(catalogue 2345 models, every topology reference resolved, CI gate passed). Jobs 2 and 3 return:So there is no failing assertion, no step name and no annotation for either red leg.
Why this looks like runner load, not content: the tree is identical across all three legs and the leg that ran first (Node 18) passed. Same signature as #1749, whose only failure is a wall-clock bound (
test/trunk-allowed-remove-from-all.test.js:167,ms < 3000) that measures 1.4-1.6 s on an idle host. Two unrelated PRs going red only on the long legs points at the shared runners being saturated, not at either diff.Recommendation: re-run jobs 2 and 3 as-is. A rebase changes nothing here (already 0 behind) and there is no conflict. If they stay red after a clean re-run, read them in the Forgejo UI rather than the API.
For Jay: the missing
actions_logobjects mean a red leg can currently be undiagnosable. If this recurs, runner log retention/cleanup is worth a look.The content is clean; the red is environmental. Reproduced the two touched suites locally at
26b35b4cand both pass.That is the same count the PR's second commit installs, so the assertions are not the problem and neither is the catalogue entry.
The shape of the run agrees. Every
Testleg onmainfinishes in 42-58 s (measured on3c70523e,1e3adbd1,7d4afb2b), and this PR's Node 18 leg finished in 50 s — but Node 20 failed at 11m8s and Node 22 at 11m1s, two nearly identical durations on legs that do the same work. The job budget istimeout-minutes: 15, so nothing was killed by the workflow; those two legs stalled for ~10 minutes that Node 18 did not.Re-running is the unblock, and only a browser can do it. I re-checked the route rather than assuming:
POST /jaylouisw/packetforge/actions/runs/4794/rerunexists but is CSRF-gated — token auth getsInvalid CSRF token.(400) and no_csrfcookie is ever issued to a token-authenticated request, and/api/v1/.../actions/runsis still 404 on this build. So please hit Re-run on run 4794 (jobs 2 and 3: Node 20 and Node 22) from a logged-in session; I cannot trigger it.Not rebasing: the branch is 1 behind, mergeable, and
mainnow carriesd39b7a87"ci: stop rebasing pull requests that are already red" — a rebase would burn two more runner slots to re-ask a question this run already answered.Rebased onto
473d4c96→8229c866; the git conflict is gone. What is left is three assertions intest/, which are the catalogue owner's, not mine to edit.Verified after the rebase (
git merge-tree origin/main origin/catalogue/1116-2155): 0 conflicts, 0 behind, 1 ahead. The five previously-conflicting files were all count lines, so they were regenerated rather than merged by hand:node scripts/regen-registry-doc.jsrewrote README,docs/CONNECTORS.md,docs/DEVICE_REGISTRY.md,docs/wiki/Home.mdandscripts/build-wiki.jsto 2,340 models / 28,806 sockets / 958 distinct port shapes (main was at 2,339 / 28,764 / 957; the new shelf is +1 model, +42 sockets, +1 shape).test/doc-catalog-counts.test.jspasses 43/43 on the rebased tree.What still fails, and it is not a conflict — the new shelf really is a third
B1LDCDshelf, so main's hard-coded 2 is now wrong:test/device-aliases.test.js:95board.length === 2, "B1LDCD finds both shelves"test/device-aliases.test.js:99"b1ldcd 14x"→ E6616 firsttest/palette-search-keys.test.js:96warm[7].length === 2The two counts are a mechanical 2→3. The middle one is a question only the catalogue loop can answer: for
b1ldcd 14x, is E6608T now the right first hit (it is a 14x 200G line chassis) with E6616 demoted, or should E6616 keep the top slot? Either answer is a two-line test edit plus the assertion's wording.Needs the catalogue loop to land that edit on top of
8229c866(or on main — the count bump is independent of this PR). Until then this PR stays red on Test (Node 18/20/22) for a reason that is not a merge problem.scripts/anddocs/were regenerated only; no engine or behaviour change was made by me.26b35b4ca8afa0f78642