3d98068ead92a95ea4b714f31f01119100a3c099 braney Wed Sep 16 10:26:18 2026 -0700 docent: preflight prints how the target is configured, refs #37892 make test TARGET=hgwdev-you points the whole directory at another server, and a red script there can be that machine's configuration rather than a bug. A different curatedHubPrefix genuinely changes what a quickLift hop produces, and a db.trackDb with a private table in front of the shared one changes what exact: true counts. Both failures look exactly like a bug. The README already warned about it, in prose, which arrives after someone has spent an hour. make preflight now prints the settings that decide it, so a redirected run carries its own explanation in the log: target https://hgwdev-braney.gi.ucsc.edu/cgi-bin /usr/local/apache/cgi-bin-braney/hg.conf central.db hgcentraltest db.trackDb trackDb_braney,trackDb curatedHubPrefix braney browser.quickLift on browser.quickLiftAlignments on browser.recTrackSets on central.db is on the list because preflight already checks named sessions and that setting says which hgcentral it looked in. The value printed is the EFFECTIVE one. The reader follows include and delete the way hg/lib/hgConfig.c does, with a later assignment winning over an earlier one, so a sandbox conf that sets nothing still shows what it inherits from the shared conf it includes. Reading cgi-bin-braney/hg.conf on its own says browser.quickLiftAlignments is unset; the CGI there sees it on, from the include on line 2. A ts park is the mirror case: 37547's frozen conf sets db.trackDb twice and the second one, 114 lines later, wins. Only the six keys in HG_CONF_KEYS are ever printed. The includes lead to hg.conf.private, which holds database passwords, so the reader parses whatever it is pointed at and prints nothing off that list. There is no way to ask a browser over http what its hg.conf says, so which file to read is a lookup by convention: genome-test and hgwdev read /usr/local/apache/cgi-bin/hg.conf, an hgwdev-<name> sandbox or demo reads cgi-bin-<name>, and a ts park on 127.0.0.1 is found by port in its ports.tsv. Anything off this machine -- hgwbeta, the RR -- says the conf cannot be read from here, which is true and better than a guess. Even with that in the log, a config difference and a code difference can still look alike. The reliable way to tell them apart is to swap only the binary: drop a control build's CGIs into the same sandbox, leave its hg.conf alone, and re-run. If the failures follow the binary they are the code. That experiment is what turned nine plausible failures on the #37547 branch into nine proven ones, and both READMEs now say so. Verified against all three shapes: genome-test, hgwdev-braney, hgwdev-demo9, and the #37547 park on 127.0.0.1:48087. Both test directories preflight clean. diff --git src/hg/utils/docent/tests/README.txt src/hg/utils/docent/tests/README.txt index d4a5b3ffaa5..67494249425 100644 --- src/hg/utils/docent/tests/README.txt +++ src/hg/utils/docent/tests/README.txt @@ -1,171 +1,198 @@ Docent tests ------------ Run by hand, not by the kent tree's `make test`: make test # every *.docent.yaml here make test T=composite # just one make parity # one script FAST and slow, and twice over make derive # the derivation alone, against expected/ (no browser) make derive-accept # rewrite those baselines, then read `git diff expected/` make test TARGET=hgwdev-demo9 # the same scripts, against another server TARGET overrides the `target:` each script carries, for every target above, and takes the same values it does: a shorthand (rr, genome-test, hgwdev, hgwbeta), a bare hgwdev-<name> sandbox or demo, or a full .../cgi-bin URL. It is how you try a suite against a branch build -- a sandbox, a ticket park from `ts`, a demo browser -- without editing the scripts. `make preflight TARGET=...` checks that server rather than the one the scripts name, so the fixture check and the run agree. Read a redirected run's failures with the server in mind. A script asserts what its OWN server draws, so a red one somewhere else can be the other machine's trackDb rather than a bug: a demo sandbox that carries only one assembly fails every script on the others, and a sandbox trackDb with a track the RR has not released changes what `exact: true` counts. Redirecting is for trying a suite elsewhere, not for moving it: the committed scripts stay pointed at the server they were written against, which is the one the nightly reads. +`make preflight` prints how the target is configured, so the log says which server was +driven AND how it differs from the one the scripts name: + + target https://hgwdev-braney.gi.ucsc.edu/cgi-bin + /usr/local/apache/cgi-bin-braney/hg.conf + central.db hgcentraltest + db.trackDb trackDb_braney,trackDb + curatedHubPrefix braney + browser.quickLift on + browser.quickLiftAlignments on + browser.recTrackSets on + +Those are read off the hg.conf the server reads, following its includes the way +hg/lib/hgConfig.c does, so the value printed is the EFFECTIVE one -- a sandbox conf that +sets nothing still shows what it inherits from the shared conf it includes. Only a fixed +list of settings is printed, because hg.conf includes hg.conf.private. + +It works for a server on this machine: genome-test, hgwdev, an hgwdev-<name> sandbox or +demo, or a ticket park from `ts` on 127.0.0.1 (looked up by port in its registry). For +hgwbeta or the RR it says the conf cannot be read from here, which is true and is better +than a guess. + +Even with that in the log, a config difference and a code difference can still look +alike. The reliable way to tell them apart is to swap only the BINARY: drop a control +build's CGIs into the same sandbox, leave its hg.conf alone, and re-run. If the failures +follow the binary they are the code. + Most tests drive a real browser against a real server, so they need the network and the shared Playwright install (/hive/groups/browser/uiTest/pw; see ../README.md). That is why none of this is part of the tree-wide test target: a broken network would fail the build. A test is an ordinary Docent script that asserts with `expect:`. It passes by exiting 0. `expect:` is the only verb that CHECKS anything, so a test with no `expect:` step in it tests nothing: `track:` accepts a name no assembly has and still exits 0. Other verbs do fail a run, so do not read the line above as "nothing else can stop it". A verb throws when it cannot do what it was told -- `mouseover:` cannot find the item, `loadSession:` cannot find the file, `drag:`, `convert:` and `go:` likewise -- and docent.js turns any step's throw into `step N (verb) failed` and exit 1. None of them looks at whether the page came out right, which is the part only `expect:` does. A script named *.xfail.docent.yaml is expected to FAIL, and the run fails if it passes. That is how a trap gets pinned rather than merely written down. `make derive` is the cheap half: DOCENT_DERIVE=1 resolves the `track:` steps against trackDb and prints the cart variables without opening a browser, in about a second. It is where Docent's own decisions live, and the baselines in expected/ are what catch a change to visVars() or tdbHideTargets() that a rendered page would hide. The trackDb listing is cached in $TMPDIR for a day (docent.js, TDB_TTL), and a cold fetch prints one provenance line that a warm run does not. That line would make the first `make derive` of the day differ from a baseline captured warm, for a reason that has nothing to do with trackDb changing, so the makefile strips it from both the run and the baseline. Everything else trackDb says about itself is kept, including the two lines that report a hub genome or an unreachable hubApi. What is covered --------------- selftest session: -> expect: -> loadSession:, on hg38 at SHH. Saves the cart, changes the view, restores it from the local file, checks rows both times. heavysession the same three steps as selftest, on a Recommended Track Set: 34 rows in, saved, moved away, loaded back, `exact: true` on both halves. selftest round-trips two rows, which barely reaches outIfNotPresent() in hgSession -- the function that writes a trackDb default for every track that is deliberately NOT in the cart, and the one a broken save-and-reload path shows up in. A path that dropped four rows left selftest green. firstrequest a track turned on has to be drawn by the request that turned it on, with no `go:`, `open:` or `convert:` in between. The bug it exists for lags by exactly one request, so any script that navigates before asserting reads a correct image and passes. It names wgEncodeRegMarkH3k4me1 for the reason in its header: a top-level track or a default-visible child would pass on the broken build too. composite clinvar with clinvarCnv hidden: the two-request split (#37953). One request would leave clinvarCnv_sel=1 and the CNV row drawn. views hideKids on the VIEW that holds the subtrack, with the sibling views hidden by name. Also covers the `_sel` checkbox, since the subtrack is `parent wgEncodeRegDnaseSignal off`, and pins the superTrack side effect below. views.xfail the same thing aimed at the COMPOSITE instead, which loses the row. Expected to fail. supertrack varsInPubs hideKids + one member: `exact: true`, because a test that only checked the member was present would pass with all six drawn. urllen {cCREs: hideKids} must not become the 1701-variable, 42,020-character GET that Apache answered with 414. Checks `noText: "Too Long"`, since a 414 renders as a perfectly good page; the derive baseline pins it at 3. customtrack addCustomTrack: with inline BED, tabs and newlines surviving the trip. scale a 3x run draws the same rows as a 1x one. ordered `ordered: true` on rows:, and the fact that a row which was not drawn is reported by rows: alone rather than failing the order check as well. ordered.xfail the same two rows named the wrong way round. Expected to fail -- a flag that cannot fail is not a check, it is a second copy of the set test. pagechecks the `expect:` checks that read the PAGE rather than the track image -- `url:`/`noUrl:` on the address, `has:`/`noHas:` on a CSS selector -- plus the positional form of `click:` (`{track:, frac:}`), which follows the item box nearest a point. All four exist for bugs that rows:, height: and text: cannot see: a search term's zero-width space stripped out of a URL (#36387), a center label attached to the wrong row (#37785), and an item that cannot be named at all because its track is `type bigBed 3` (#36335). pagechecks the same four aimed the wrong way at once. Expected to fail. The message .xfail names every check that failed, so one run says which of the four broke. colorchecks `color:`, the one check that reads the track IMAGE: is:/not: on the color a row is mostly drawn in, `part: label` for the center label instead of the items, `at:` for one item rather than the whole row, and the list form. It exists for #36212, where a track that sets both `itemRgb on` and `color` draws its items in the wrong one -- same rows, same height, same names, same tooltips, so nothing but the pixels can tell. colorchecks the same six aimed wrong, all in ONE expect: step so the message has to .xfail name all six. Expected to fail. The comment lists them in order; read the log rather than trusting the exit code. expectfail an assertion that is plainly false. Expected to fail -- if it ever passes, .xfail `expect:` has stopped throwing and every other test here means nothing. make parity FAST vs slow, and a rerun, on composite. FAST drops the dwells and the recording and must not change what the page ends up showing; the rerun catches state left behind in the cart. Two things these tests found ---------------------------- Worth knowing before writing more: * Turning on anything under a superTrack sends `<superTrack>=show`, and every OTHER member then comes up at its own trackDb visibility -- so `hide: all` is undone for them. views asserts wgEncodeRegMarkH3k27ac comes back, rather than working around it. Whether Docent should be cleverer here is an open question, not a settled one. * `hideKids` on a VIEW has to enumerate its leaves (a view holds no sub-containers to stop at), so views sends 188 variables in a 6,986-character request. That is under Apache's 8,190 limit with less room than is comfortable. Its `noText: "Too Long"` is what turns a future overflow into a clear failure instead of a strange one. Still to write -------------- mouseover: by item: on stacked items, and the timing case where a neighbour's tooltip is still up on arrival pinShot: several tooltips in one figure, cursors drawn convert: quickLift onto a GenArk haplotype, hideDefaults re-checked -- note a session taken after it cannot be checked in, see #38046 drag: each of then: zoom / highlight / cancel addHub:, addPublicHub: the two hub attach paths (a stable hub URL is the hard part) montage: panel order, lettering, a named shot that was never taken goShow: the suggestion menu, including a `pick:` that matches nothing loadSession: the three remote forms -- only the local-file form is covered the YAML lint `{item:name}` with no space warns and drops the argument. This needs a test that reads stderr, which the harness does not do yet. hgCollection no script here or in regress/ reaches that CGI at all, and it shares visibility logic with hgTracks by COPY rather than by call: hg/hgCollection/hgCollection.c carries its own isParentVisible(), a verbatim copy of the one in hg/lib/trackHub.c. The copy in trackHub.c was caught by nine scripts in regress/ on the #37547 branch; the copy in hgCollection.c was found by grep afterwards, and would have dropped a container's children out of a saved collection in the same silent way. A test needs a `collection:` verb: the page puts tracks into a collection by dragging between two jsTrees, and `drag:` is the genomic drag-select on the track image, not that. Its buttons are #newCollection, #doNewCollection, #saveCollections and #discardChanges, which is enough to open and save one but not to put a track in it. A test that needs a stable server-side fixture (a hub, a custom track) should carry it in the script rather than assume something on disk. colorchecks is the one exception, and the reason is worth knowing before someone else hits it. `color:` has to address a ROW by name, and a custom track cannot be addressed by name at all: hgTracks assigns its row id (`ct_<name>_<number>`), which is why customtrack asserts on label text instead of on `rows:`. So an inline custom track -- the self-contained way to get a known color onto the page -- is the one fixture this check cannot use. It reads ~/public_html/docentFixtures/itemRgbHub/ instead, which tests/regress/rm36212.xfail needs anyway, and which `make preflight` checks is still there.