Repository navigation
feat(ui): steadier orchestration pages, reviewed reruns and acknowledgements - #1409
Merged
Merged
Conversation
…tions The script compiles the core project, translates each compiled model from Snowflake to DuckDB with Polyglot plus DuckShop-specific rewrites, and writes Snowflake functions DuckDB lacks as project SQL macros. The copy compiles with no errors and builds every model on DuckDB; an E2E test runs the whole path.
A sql freshness query names the Snowflake raw table, which the DuckDB copy does not have, so freshness failed and blocked the models reading that source.
Graphs stay hidden until their first fit has painted, so cards no longer appear at unlaid-out positions and slide into place. The plan page keeps one layout while it loads: one overlay covers the result from the first command check until the plan arrives, and the side panel no longer appears late and narrows the page.
…in the DuckDB copy Formatted TO_TIMESTAMP calls now parse with their format instead of a plain cast, so day-first hub tracking times are no longer NULL. REGEXP_SUBSTR and REGEXP_LIKE now apply Snowflake's i, s and m flags, and e extracts the first group, so case-insensitive matches such as supplier payment terms and refund reasons resolve as they do on Snowflake.
Each first-fit attempt is bounded by a frame, so a fit the flow never runs cannot keep the graph hidden, and an empty graph shows its controls.
The plan page now reads like the terminal plan: work that runs first, one section per change with its query, config or schema diff, then the routine models, functions, seeds and warnings. Unchanged tables and views are no longer counted as skipped. The lineage graph moves to its own tab, and a selection too large to draw shows only the models the plan changes. A job run can reprocess a date range (latest data, yesterday, the last 7 or 30 days, or custom dates), kept in the command as its cursor-window flags. Rebuilding from scratch leaves the run menu and is done by editing the command. Run opens the plan page; with ?flow=panel it opens a side panel on the job with the same range and a compact plan, while the two are compared.
…e panel Pages now switch as soon as a link is clicked and show placeholder rows until their first read answers, instead of holding the click for 200 ms and showing loading text only after 400 ms. Page reads start when a link is hovered, so they are usually ready by the click. List totals and tab counts wait for the list instead of reading zero. The lineage inspector keeps the selected node's facts while they reload and shows the loading state for a newly selected node, instead of showing the previous node and then flipping to Loading. The graph can zoom out to the whole project, its page paints before the graph is built, and a graph laid out before is not laid out again.
…b sidebar out in sections Every job run now chooses its data in one place: the latest data, a date range for jobs with a cursor, or a full refresh, in the plan page and the side panel alike. An option still being typed in a job run's command, such as `--`, holds Run until it is finished instead of being refused. The job sidebar shows recent runs with their outcomes, what starts the job and when, its definition and its settings, in headed sections like the lineage inspector. A full refresh plan lists each materialization with its model count.
…ide lists between node clicks The run timeline moved its window to the server's clock on every answer, which read the timeline again as soon as each answer arrived: jobs, triggers, the newest runs and every trigger's ticks, several times a second for as long as the overview or runs page was open. It now follows the server's clock only when that is a minute away, and moves on with time once a minute. A node's detail reused nothing between clicks and read every check, job and trigger in the project to show one node's. Those lists are now kept until the next change signal, or a minute at most.
Browsers now receive the control API's reads and the embedded UI's files gzip-compressed; a project's check or run listing is a tenth of its size or less on the wire. Sign-in responses, which carry the CSRF token, writes and the location API stay uncompressed, so a secret is never compressed beside text a requester chose.
kvlonge
enabled auto-merge (squash)
November 11, 2026 04:43
…ector's like its facts
Drops the side-panel run flow that was offered for comparison, and the compact plan layout only it used.
…n earlier command
…when a node's facts fail
…most specific job A node built by several member jobs is drawn in the job building the fewest nodes, so a small job inside a broad one is no longer an empty box; each box counts every node its job builds. A collapsed box shows only its header, and lineage to its nodes meets the box.
A job selected with upstream operators (+model) includes the sources it reads. A build never executes a source, so the worker reported each one as skipped and every such run finished as partial with no failure behind it. Runs now record and report only the nodes they execute; a source in a build's selection gets no node row and no outcome.
Anyone who can see a run can acknowledge it. Each acknowledgement and its withdrawal is an append-only security audit event, and the overview moves acknowledged items into a collapsed Acknowledged list with who and when. The list now shows one problem per job, its latest finished run, so a later success clears it and acknowledging never brings an older failure back. A problem's detail names what went wrong (failed and skipped node counts) instead of how the worker stopped.
…one level or kind Rerunning from a run's page or from the runs list opens a review page with the nodes each rerun repeats and the permissions it needs; nothing reruns until Run is pressed, as for every other run. The log's level and kind filters are single choices with All, like the checks page.
… worker A build never runs a selected source, so the worker marked each one skipped and a job selected with upstream sources ended partial with nothing wrong. The earlier server-side exclusion broke older workers and in-flight group runs during a server-first upgrade, so it is reverted: the server again keeps a row per selection key, and the worker completes a build's selected sources as succeeded once the command returns.
- The overview lists one problem per job; only a failure or a full run of the job settles it, so a cancelled or narrowed run never hides a failure. - The rerun review page falls back to rerun all for an unknown mode, and after a partial failure narrows to the runs that didn't start. - Pages no longer shift as they load: the first paint waits briefly for the web fonts, the top bar and runs toolbar hold their width, and the checks table keeps its header while rows load. - Opening the SQL panel on the lineage graph keeps the zoom. - A group run's timeline folds each job's nodes.
…succeed Sources a build completes without running no longer count towards a partial outcome, in the worker and in a group's parent, so a build whose models all failed ends failed as before. The rerun page now leaves out runs whose rerun already started, so Run never repeats one, and falls back to rerun all for an unknown mode on several runs.
kvlonge
added
the
full-verify
Amp for GitHub
kvlonge
changed the title
kvlonge
disabled auto-merge
November 11, 2026 09:40
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
The orchestration UI felt slow and jumpy on a realistic project, a job selected with upstream sources ended
partialwith nothing wrong, and acknowledged problems kept their place on the overview. A DuckDB copy of DuckShop gives a realistic project to demonstrate against.Changes
sqb plan; every job run and every rerun is reviewed there before it starts.Verification
make check-cipasses on the final head; svelte-check 0 errors; vitest 202 passed.