Repository navigation
UI: Fix clearing a task instance doing nothing after Confirm - #74725
Merged
Merged
Conversation
Confirming a task instance clear opens a second dialog that re-runs the dry run and then sends the clear. Since the zag dialog upgrade that came with Chakra 3.37, closing the clear dialog underneath dismisses that second dialog twice over: the layer stack dismisses layers above a removed one, and the clear dialog hands focus back to its trigger. The dry run finished but the clear request was never sent, and the queued-task warning never appeared. The second dialog now ignores both of those dismissals. Outside clicks and Escape still close it. closes: apache#74712
kaxil
requested review from
bbovenzi,
choo121600,
guan404ming,
pierrejeambrun,
ryanahamilton,
shubhamraj-git and
vatsrahul1001
as code owners
October 14, 2026 11:56
1 of 2 tasks
Member
|
Do we know when this was introduced? |
Member
Author
ashb
approved these changes
Oct 14, 2026
Member
Author
|
The specific change inside that bump is chakra-ui/zag@9a9381d ( I confirmed it by pinning just |
bbovenzi
added a commit
to astronomer/airflow
that referenced
this pull request
Oct 14, 2026
The UI no longer calls getExecution, so the route keeps the shape it had on main rather than gaining order_by, in_loop and region nesting that nothing reaches. Its translations go with it. Ordering by iteration is withdrawn. A row's own region index is its map slot once a mapped task runs inside a loop, so it does not order by what the column shows, and the enclosing pass is an ancestry walk rather than a column. Sorting on region_index while the cursor reported the pass would page wrongly, which is worse than the raise it replaced. Clearing is fixed on main by apache#74725: the confirmation dialog was dismissed twice over by the layer stack, not by its mutation observer being torn down. This drops the change that guessed otherwise, which also reopened the main dialog mid-clear on the gated path. A link that moves the coordinate now clears the pass it leaves behind, so the breadcrumb and the Gantt cannot name a pass the page is not on. The loop filter is sent only with a run, since the endpoint rejects it without one. The grid structure is not requested without a Dag, and the task instance stream is only materialised for a Dag with loops. Also removes two imports of a module this branch deleted, which were failing static checks. Co-Authored-By: Claude <noreply@anthropic.com>
Contributor
Backport successfully created: v3-4-testNote: As of Merging PRs targeted for Airflow 3.X In matter of doubt please ask in #release-management Slack channel.
|
ashb
pushed a commit
that referenced
this pull request
Oct 14, 2026
…irm (#74725) (#74738) Confirming a task instance clear opens a second dialog that re-runs the dry run and then sends the clear. Since the zag dialog upgrade that came with Chakra 3.37, closing the clear dialog underneath dismisses that second dialog twice over: the layer stack dismisses layers above a removed one, and the clear dialog hands focus back to its trigger. The dry run finished but the clear request was never sent, and the queued-task warning never appeared. The second dialog now ignores both of those dismissals. Outside clicks and Escape still close it. (cherry picked from commit ea805e3) closes: #74712 Co-authored-by: Kaxil Naik <kaxilnaik@gmail.com>
bbovenzi
added a commit
that referenced
this pull request
Oct 14, 2026
* Filter task instances by loop and iteration Regions are how the server locates an execution; they are not something a user should have to know to narrow a table. The loop filter is expressed the way the Dag was written -- the loop's task group id and the pass number -- and the server resolves that to the regions behind it. Loop and iteration are one control rather than two. An iteration index means nothing until you say which loop counts it, so the two could previously be set independently into combinations that matched nothing: a pass without a loop, or a pass the loop never reached. Both now sit behind a single pill whose options come from the loop's own summary, so only passes that exist are offered, and removing the pill clears both params. This follows the Run state pill, which already spreads one value over several URL params. The same options render as a strip above the table, because a loop's shape -- how many passes it took and how they went -- is what you came to the page for, and a menu you have to open is a poor place for it. Each chip is also the control that narrows to it, and clicking the active one clears it. Past ten passes the chips collapse into the rich dropdown, the threshold task tries already settled on. A looped group no longer has a tab of its own. Its page is its task instances, the same as any other group, with the strip and filter carrying the loop vocabulary. The strip is scoped to the loop the group belongs to rather than every loop in the Dag; group ids are dotted paths, so a group nested inside a loop reports the loop enclosing it. Grid links now carry the newest region coordinate. A task inside a loop has one row per iteration, so an unscoped link was ambiguous and the API rejected it with "a loop producer requires an explicit scope". isRegional answers whether a task instance carries a region; it is not the same question as whether a loop produced it, and a mapped expansion is regional without being in a loop. Cross-run scope and the mapped-tasks tab ask isLoopTaskInstance instead, backed by the in_loop the schema already reports. Affected-task selection goes back to being keyed by coordinate. Keying it by row id meant an exclusion the user ticked was lost when a retry came back under a new id -- the case its own test describes. The filter registry no longer fetches a Dag structure for every filter bar in the app; the page that knows which Dag is in view supplies the loops. Drops the translation keys the rewrite orphaned. Co-Authored-By: Claude <noreply@anthropic.com> * Select a loop iteration from the breadcrumb A pass is part of where you are rather than a filter you applied -- the same task id means a different run of it in every pass -- so it becomes its own breadcrumb level, built on the switcher the Dag and run levels already use. A task group can be read whole and keeps an "all" entry; a task instance is one row and must name a pass. Picking one moves the region coordinate as well as the filter, so the page shows that pass rather than merely narrowing to it, and a pass carried over from a different loop is ignored instead of being read as a position in this one. The task instances list loses its iteration strip. The loop pill already offers every loop in scope with its passes, so the strip was a second way to say the same thing. The loop filter now survives navigation. Coordinates pin one execution and must not follow you to another, but a pass is where you are in a loop, so it stays as you move between the loop's tasks; a page outside that loop ignores it. Task instances can be ordered by iteration, and carry a Loop column so a list spanning loops says which one each row is from. The gantt outlines the pinned pass, over the whole group or the single row depending on which is being read, and clicking a bar selects the pass it was drawn for. Graph links carry the newest region coordinate for the same reason grid links do: a task inside a loop has one row per pass, so an unscoped link is ambiguous and the API rejects it. Checkbox options move to the left of a dialog footer, away from Cancel and Confirm. This replaces a margin hack in the clear dialog and gives the other dialogs the same shape. Also fixes clearing a task instance, which had been failing silently: the dialog closed itself as soon as the request was sent, and unmounting took its mutation observer with it, so neither the error toast nor the query invalidation ever ran. Closing is left to the mutation's success path. Co-Authored-By: Claude <noreply@anthropic.com> * Address review: drop the dead execution changes and the broken sort The UI no longer calls getExecution, so the route keeps the shape it had on main rather than gaining order_by, in_loop and region nesting that nothing reaches. Its translations go with it. Ordering by iteration is withdrawn. A row's own region index is its map slot once a mapped task runs inside a loop, so it does not order by what the column shows, and the enclosing pass is an ancestry walk rather than a column. Sorting on region_index while the cursor reported the pass would page wrongly, which is worse than the raise it replaced. Clearing is fixed on main by #74725: the confirmation dialog was dismissed twice over by the layer stack, not by its mutation observer being torn down. This drops the change that guessed otherwise, which also reopened the main dialog mid-clear on the gated path. A link that moves the coordinate now clears the pass it leaves behind, so the breadcrumb and the Gantt cannot name a pass the page is not on. The loop filter is sent only with a run, since the endpoint rejects it without one. The grid structure is not requested without a Dag, and the task instance stream is only materialised for a Dag with loops. Also removes two imports of a module this branch deleted, which were failing static checks. Co-Authored-By: Claude <noreply@anthropic.com> * Address review: structure-based loop scope, polling, and an exact region filter Loop membership is read from the Dag structure rather than the task id. A loop declared with prefix_group_id=False does not name its members after itself, so matching on the prefix lost them the breadcrumb, the filter and the Gantt highlight. region_id goes back to naming one region exactly, and stops rejecting the cross-run form. The logical scope of a pass -- the loop's own region plus the expansions nested in it -- is what loop_id and iteration are for, so the two filters keep one meaning each. The coverage for that pair comes back alongside the region case, including the two guards. The loop options follow the run the way the loop summary does, so a running loop's passes and their states no longer sit stale behind a five minute staleTime. Also fixes the task instance stream being drained before the summaries were built: without loops it is walked once, and the pass that looks for the newest iteration has nothing to find, so it is skipped rather than consuming it. Co-Authored-By: Claude <noreply@anthropic.com> * Take a row's loop iteration from the server, and count every pass The breadcrumb worked out which pass a page was on by reading region_index off the url. For a mapped expansion that index is the map slot rather than the pass, so it searched the pass list for a slot number, found nothing, and dropped the iteration level entirely -- leaving a mapped task inside a loop with no way to see or change its pass. A row already knows its own pass: loop_iteration is resolved server side and holds for an expansion and for a pass a clear forked into a successor region alike. Only a group, which is not one row, still reads the url. Map index and iteration are separate levels and now stay that way, rather than one url field being asked to mean both. State counts cover every pass. A looped task's cell stands for the work the loop did and the UI says how many iterations that covers, so an earlier pass put back to work by a clear is visible instead of hidden behind a finished later one. The coordinate still comes from the newest pass, since a link has to address a row. Co-Authored-By: Claude <noreply@anthropic.com> * Resolve a loop pass to the row it became, instead of guessing a coordinate Switching iteration on a task instance page wrote the loop's region and the pass number as if they addressed a row. They do not: a mapped expansion lives in its own region, and a clear forks later passes into a successor one. The pair named no row, and asking for a looped task instance that way is rejected rather than answered empty, so navigating between a loop's tasks could fail outright. The switcher now writes only what it knows -- the loop and the pass. The page resolves that to a coordinate through the list endpoint, which already takes the pair and returns each row's own region, and writes the answer back into the url. Everything downstream keeps reading the url unchanged, and the resolved link is the one worth sharing. The lookup sits on the page rather than in the shared coordinate hook, which nineteen components use: a query there would have every one of them make a request. The page holds its children while the lookup is in flight. The request has to be withheld rather than made and handled, since there is no scope to make it with. Co-Authored-By: Claude <noreply@anthropic.com> * Open the Gantt highlight where one run's segments meet An outline rings all four sides, so a run drawn as a queued stretch followed by its state showed the highlight down the seam between them -- an edge inside the thing being highlighted rather than its boundary. The sides are drawn individually now, and the two that already know they are touching a neighbour are left off. Co-Authored-By: Claude <noreply@anthropic.com> * Address review: group membership, paging, and the loop a summary is asked for The Gantt group highlight decided membership from the task id, so a group declared with prefix_group_id=False lost it. It reads the structure now, the way the loop scope already does. Switching pass restarts paging. A page of one pass says nothing about where the next one begins, so carrying its cursor across left the table on rows from the pass before. A task instance page no longer offers its own row's region as the loop to summarise. Only a group page's region names an invocation: a row's region is an expansion, or a pass a clear forked, and asking for a run summary by one of those is rejected -- which left the header falling back on a count of itself. Co-Authored-By: Claude <noreply@anthropic.com> * Highlight only a pass the reader chose in the Gantt The highlight fell back to the region index when no iteration was named. That index is the map slot on a mapped expansion, so it rang bars of whatever pass happened to share the number -- and resolving a pass to a row now writes a region index into the url far more often, which made the fallback fire where it used to be rare. Co-Authored-By: Claude <noreply@anthropic.com> * Keep earlier attempts on the Gantt chart after a clear Every attempt is its own row and only the newest stays live, and ORM selects read live rows unless they ask otherwise. The chart did not, so clearing a task took its earlier tries off the chart entirely -- including the attempt that the pending-retry filter just above deliberately drops the live row in favour of, which could never have been there to find. Co-Authored-By: Claude <noreply@anthropic.com> * Name the try in the Gantt tooltip The try sat inside a branch keyed on start_date, but the chart reports min_start_date, so it never reached a bar's tooltip at all. It is read off the shape a caller passes now, rather than from one of the branches below it. Shown on a first try too. Clearing a loop pass regenerates the later ones as fresh executions that start again at try 1, so two neighbouring bars can both be a first try and the number is what tells them apart. Co-Authored-By: Claude <noreply@anthropic.com> * Drop the iteration label from the mapped task instances header The counts beside it are the task's across the whole loop, so naming one pass above them said that pass had every expansion of every pass. This page lists the expansions of a mapped task, which is an aggregate view, and the pass it belongs to is already named in the breadcrumb directly above. Narrowing to a single pass is what the task instances table and its loop filter are for. Co-Authored-By: Claude <noreply@anthropic.com> * Regenerate the UI client and reformat after the review changes The spec changed when the execution route went back to its shape on main and the iteration sort was withdrawn, but the generated client still carried both, so the UI lint hook regenerated it and failed on the difference. Two Python files were left unformatted by the same round of edits. Co-Authored-By: Claude <noreply@anthropic.com> * Fix the type errors the UI static check compiles for The checked-in tsconfig does not typecheck src; the hook builds its own project extending tsconfig.app.json, which turns on noUnusedLocals among other things, so running tsc against the root config said clean while the hook did not. Two unused locals left behind by earlier edits, a key typed as a plain string where its neighbours are cast, and a tooltip reading try_number off a union whose grid summary member does not carry one -- lifting it out of the start_date branch lost the narrowing that had been doing that work. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
1 task
ashb
added a commit
that referenced
this pull request
Oct 17, 2026
* Filter task instances by loop and iteration Regions are how the server locates an execution; they are not something a user should have to know to narrow a table. The loop filter is expressed the way the Dag was written -- the loop's task group id and the pass number -- and the server resolves that to the regions behind it. Loop and iteration are one control rather than two. An iteration index means nothing until you say which loop counts it, so the two could previously be set independently into combinations that matched nothing: a pass without a loop, or a pass the loop never reached. Both now sit behind a single pill whose options come from the loop's own summary, so only passes that exist are offered, and removing the pill clears both params. This follows the Run state pill, which already spreads one value over several URL params. The same options render as a strip above the table, because a loop's shape -- how many passes it took and how they went -- is what you came to the page for, and a menu you have to open is a poor place for it. Each chip is also the control that narrows to it, and clicking the active one clears it. Past ten passes the chips collapse into the rich dropdown, the threshold task tries already settled on. A looped group no longer has a tab of its own. Its page is its task instances, the same as any other group, with the strip and filter carrying the loop vocabulary. The strip is scoped to the loop the group belongs to rather than every loop in the Dag; group ids are dotted paths, so a group nested inside a loop reports the loop enclosing it. Grid links now carry the newest region coordinate. A task inside a loop has one row per iteration, so an unscoped link was ambiguous and the API rejected it with "a loop producer requires an explicit scope". isRegional answers whether a task instance carries a region; it is not the same question as whether a loop produced it, and a mapped expansion is regional without being in a loop. Cross-run scope and the mapped-tasks tab ask isLoopTaskInstance instead, backed by the in_loop the schema already reports. Affected-task selection goes back to being keyed by coordinate. Keying it by row id meant an exclusion the user ticked was lost when a retry came back under a new id -- the case its own test describes. The filter registry no longer fetches a Dag structure for every filter bar in the app; the page that knows which Dag is in view supplies the loops. Drops the translation keys the rewrite orphaned. * Select a loop iteration from the breadcrumb A pass is part of where you are rather than a filter you applied -- the same task id means a different run of it in every pass -- so it becomes its own breadcrumb level, built on the switcher the Dag and run levels already use. A task group can be read whole and keeps an "all" entry; a task instance is one row and must name a pass. Picking one moves the region coordinate as well as the filter, so the page shows that pass rather than merely narrowing to it, and a pass carried over from a different loop is ignored instead of being read as a position in this one. The task instances list loses its iteration strip. The loop pill already offers every loop in scope with its passes, so the strip was a second way to say the same thing. The loop filter now survives navigation. Coordinates pin one execution and must not follow you to another, but a pass is where you are in a loop, so it stays as you move between the loop's tasks; a page outside that loop ignores it. Task instances can be ordered by iteration, and carry a Loop column so a list spanning loops says which one each row is from. The gantt outlines the pinned pass, over the whole group or the single row depending on which is being read, and clicking a bar selects the pass it was drawn for. Graph links carry the newest region coordinate for the same reason grid links do: a task inside a loop has one row per pass, so an unscoped link is ambiguous and the API rejects it. Checkbox options move to the left of a dialog footer, away from Cancel and Confirm. This replaces a margin hack in the clear dialog and gives the other dialogs the same shape. Also fixes clearing a task instance, which had been failing silently: the dialog closed itself as soon as the request was sent, and unmounting took its mutation observer with it, so neither the error toast nor the query invalidation ever ran. Closing is left to the mutation's success path. * Address review: drop the dead execution changes and the broken sort The UI no longer calls getExecution, so the route keeps the shape it had on main rather than gaining order_by, in_loop and region nesting that nothing reaches. Its translations go with it. Ordering by iteration is withdrawn. A row's own region index is its map slot once a mapped task runs inside a loop, so it does not order by what the column shows, and the enclosing pass is an ancestry walk rather than a column. Sorting on region_index while the cursor reported the pass would page wrongly, which is worse than the raise it replaced. Clearing is fixed on main by #74725: the confirmation dialog was dismissed twice over by the layer stack, not by its mutation observer being torn down. This drops the change that guessed otherwise, which also reopened the main dialog mid-clear on the gated path. A link that moves the coordinate now clears the pass it leaves behind, so the breadcrumb and the Gantt cannot name a pass the page is not on. The loop filter is sent only with a run, since the endpoint rejects it without one. The grid structure is not requested without a Dag, and the task instance stream is only materialised for a Dag with loops. Also removes two imports of a module this branch deleted, which were failing static checks. * Address review: structure-based loop scope, polling, and an exact region filter Loop membership is read from the Dag structure rather than the task id. A loop declared with prefix_group_id=False does not name its members after itself, so matching on the prefix lost them the breadcrumb, the filter and the Gantt highlight. region_id goes back to naming one region exactly, and stops rejecting the cross-run form. The logical scope of a pass -- the loop's own region plus the expansions nested in it -- is what loop_id and iteration are for, so the two filters keep one meaning each. The coverage for that pair comes back alongside the region case, including the two guards. The loop options follow the run the way the loop summary does, so a running loop's passes and their states no longer sit stale behind a five minute staleTime. Also fixes the task instance stream being drained before the summaries were built: without loops it is walked once, and the pass that looks for the newest iteration has nothing to find, so it is skipped rather than consuming it. * Take a row's loop iteration from the server, and count every pass The breadcrumb worked out which pass a page was on by reading region_index off the url. For a mapped expansion that index is the map slot rather than the pass, so it searched the pass list for a slot number, found nothing, and dropped the iteration level entirely -- leaving a mapped task inside a loop with no way to see or change its pass. A row already knows its own pass: loop_iteration is resolved server side and holds for an expansion and for a pass a clear forked into a successor region alike. Only a group, which is not one row, still reads the url. Map index and iteration are separate levels and now stay that way, rather than one url field being asked to mean both. State counts cover every pass. A looped task's cell stands for the work the loop did and the UI says how many iterations that covers, so an earlier pass put back to work by a clear is visible instead of hidden behind a finished later one. The coordinate still comes from the newest pass, since a link has to address a row. * Resolve a loop pass to the row it became, instead of guessing a coordinate Switching iteration on a task instance page wrote the loop's region and the pass number as if they addressed a row. They do not: a mapped expansion lives in its own region, and a clear forks later passes into a successor one. The pair named no row, and asking for a looped task instance that way is rejected rather than answered empty, so navigating between a loop's tasks could fail outright. The switcher now writes only what it knows -- the loop and the pass. The page resolves that to a coordinate through the list endpoint, which already takes the pair and returns each row's own region, and writes the answer back into the url. Everything downstream keeps reading the url unchanged, and the resolved link is the one worth sharing. The lookup sits on the page rather than in the shared coordinate hook, which nineteen components use: a query there would have every one of them make a request. The page holds its children while the lookup is in flight. The request has to be withheld rather than made and handled, since there is no scope to make it with. * Open the Gantt highlight where one run's segments meet An outline rings all four sides, so a run drawn as a queued stretch followed by its state showed the highlight down the seam between them -- an edge inside the thing being highlighted rather than its boundary. The sides are drawn individually now, and the two that already know they are touching a neighbour are left off. * Address review: group membership, paging, and the loop a summary is asked for The Gantt group highlight decided membership from the task id, so a group declared with prefix_group_id=False lost it. It reads the structure now, the way the loop scope already does. Switching pass restarts paging. A page of one pass says nothing about where the next one begins, so carrying its cursor across left the table on rows from the pass before. A task instance page no longer offers its own row's region as the loop to summarise. Only a group page's region names an invocation: a row's region is an expansion, or a pass a clear forked, and asking for a run summary by one of those is rejected -- which left the header falling back on a count of itself. * Highlight only a pass the reader chose in the Gantt The highlight fell back to the region index when no iteration was named. That index is the map slot on a mapped expansion, so it rang bars of whatever pass happened to share the number -- and resolving a pass to a row now writes a region index into the url far more often, which made the fallback fire where it used to be rare. * Keep earlier attempts on the Gantt chart after a clear Every attempt is its own row and only the newest stays live, and ORM selects read live rows unless they ask otherwise. The chart did not, so clearing a task took its earlier tries off the chart entirely -- including the attempt that the pending-retry filter just above deliberately drops the live row in favour of, which could never have been there to find. * Name the try in the Gantt tooltip The try sat inside a branch keyed on start_date, but the chart reports min_start_date, so it never reached a bar's tooltip at all. It is read off the shape a caller passes now, rather than from one of the branches below it. Shown on a first try too. Clearing a loop pass regenerates the later ones as fresh executions that start again at try 1, so two neighbouring bars can both be a first try and the number is what tells them apart. * Drop the iteration label from the mapped task instances header The counts beside it are the task's across the whole loop, so naming one pass above them said that pass had every expansion of every pass. This page lists the expansions of a mapped task, which is an aggregate view, and the pass it belongs to is already named in the breadcrumb directly above. Narrowing to a single pass is what the task instances table and its loop filter are for. * Regenerate the UI client and reformat after the review changes The spec changed when the execution route went back to its shape on main and the iteration sort was withdrawn, but the generated client still carried both, so the UI lint hook regenerated it and failed on the difference. Two Python files were left unformatted by the same round of edits. * Fix the type errors the UI static check compiles for The checked-in tsconfig does not typecheck src; the hook builds its own project extending tsconfig.app.json, which turns on noUnusedLocals among other things, so running tsc against the root config said clean while the hook did not. Two unused locals left behind by earlier edits, a key typed as a plain string where its neighbours are cast, and a tooltip reading try_number off a union whose grid summary member does not carry one -- lifting it out of the start_date branch lost the narrowing that had been doing that work. --------- (cherry picked from commit 42945d6) Co-authored-by: Brent Bovenzi <brent@astronomer.io>
vatsrahul1001
pushed a commit
that referenced
this pull request
Oct 20, 2026
…irm (#74725) (#74738) Confirming a task instance clear opens a second dialog that re-runs the dry run and then sends the clear. Since the zag dialog upgrade that came with Chakra 3.37, closing the clear dialog underneath dismisses that second dialog twice over: the layer stack dismisses layers above a removed one, and the clear dialog hands focus back to its trigger. The dry run finished but the clear request was never sent, and the queued-task warning never appeared. The second dialog now ignores both of those dismissals. Outside clicks and Escape still close it. (cherry picked from commit ea805e3) closes: #74712 Co-authored-by: Kaxil Naik <kaxilnaik@gmail.com>
vatsrahul1001
pushed a commit
that referenced
this pull request
Oct 20, 2026
* Filter task instances by loop and iteration Regions are how the server locates an execution; they are not something a user should have to know to narrow a table. The loop filter is expressed the way the Dag was written -- the loop's task group id and the pass number -- and the server resolves that to the regions behind it. Loop and iteration are one control rather than two. An iteration index means nothing until you say which loop counts it, so the two could previously be set independently into combinations that matched nothing: a pass without a loop, or a pass the loop never reached. Both now sit behind a single pill whose options come from the loop's own summary, so only passes that exist are offered, and removing the pill clears both params. This follows the Run state pill, which already spreads one value over several URL params. The same options render as a strip above the table, because a loop's shape -- how many passes it took and how they went -- is what you came to the page for, and a menu you have to open is a poor place for it. Each chip is also the control that narrows to it, and clicking the active one clears it. Past ten passes the chips collapse into the rich dropdown, the threshold task tries already settled on. A looped group no longer has a tab of its own. Its page is its task instances, the same as any other group, with the strip and filter carrying the loop vocabulary. The strip is scoped to the loop the group belongs to rather than every loop in the Dag; group ids are dotted paths, so a group nested inside a loop reports the loop enclosing it. Grid links now carry the newest region coordinate. A task inside a loop has one row per iteration, so an unscoped link was ambiguous and the API rejected it with "a loop producer requires an explicit scope". isRegional answers whether a task instance carries a region; it is not the same question as whether a loop produced it, and a mapped expansion is regional without being in a loop. Cross-run scope and the mapped-tasks tab ask isLoopTaskInstance instead, backed by the in_loop the schema already reports. Affected-task selection goes back to being keyed by coordinate. Keying it by row id meant an exclusion the user ticked was lost when a retry came back under a new id -- the case its own test describes. The filter registry no longer fetches a Dag structure for every filter bar in the app; the page that knows which Dag is in view supplies the loops. Drops the translation keys the rewrite orphaned. * Select a loop iteration from the breadcrumb A pass is part of where you are rather than a filter you applied -- the same task id means a different run of it in every pass -- so it becomes its own breadcrumb level, built on the switcher the Dag and run levels already use. A task group can be read whole and keeps an "all" entry; a task instance is one row and must name a pass. Picking one moves the region coordinate as well as the filter, so the page shows that pass rather than merely narrowing to it, and a pass carried over from a different loop is ignored instead of being read as a position in this one. The task instances list loses its iteration strip. The loop pill already offers every loop in scope with its passes, so the strip was a second way to say the same thing. The loop filter now survives navigation. Coordinates pin one execution and must not follow you to another, but a pass is where you are in a loop, so it stays as you move between the loop's tasks; a page outside that loop ignores it. Task instances can be ordered by iteration, and carry a Loop column so a list spanning loops says which one each row is from. The gantt outlines the pinned pass, over the whole group or the single row depending on which is being read, and clicking a bar selects the pass it was drawn for. Graph links carry the newest region coordinate for the same reason grid links do: a task inside a loop has one row per pass, so an unscoped link is ambiguous and the API rejects it. Checkbox options move to the left of a dialog footer, away from Cancel and Confirm. This replaces a margin hack in the clear dialog and gives the other dialogs the same shape. Also fixes clearing a task instance, which had been failing silently: the dialog closed itself as soon as the request was sent, and unmounting took its mutation observer with it, so neither the error toast nor the query invalidation ever ran. Closing is left to the mutation's success path. * Address review: drop the dead execution changes and the broken sort The UI no longer calls getExecution, so the route keeps the shape it had on main rather than gaining order_by, in_loop and region nesting that nothing reaches. Its translations go with it. Ordering by iteration is withdrawn. A row's own region index is its map slot once a mapped task runs inside a loop, so it does not order by what the column shows, and the enclosing pass is an ancestry walk rather than a column. Sorting on region_index while the cursor reported the pass would page wrongly, which is worse than the raise it replaced. Clearing is fixed on main by #74725: the confirmation dialog was dismissed twice over by the layer stack, not by its mutation observer being torn down. This drops the change that guessed otherwise, which also reopened the main dialog mid-clear on the gated path. A link that moves the coordinate now clears the pass it leaves behind, so the breadcrumb and the Gantt cannot name a pass the page is not on. The loop filter is sent only with a run, since the endpoint rejects it without one. The grid structure is not requested without a Dag, and the task instance stream is only materialised for a Dag with loops. Also removes two imports of a module this branch deleted, which were failing static checks. * Address review: structure-based loop scope, polling, and an exact region filter Loop membership is read from the Dag structure rather than the task id. A loop declared with prefix_group_id=False does not name its members after itself, so matching on the prefix lost them the breadcrumb, the filter and the Gantt highlight. region_id goes back to naming one region exactly, and stops rejecting the cross-run form. The logical scope of a pass -- the loop's own region plus the expansions nested in it -- is what loop_id and iteration are for, so the two filters keep one meaning each. The coverage for that pair comes back alongside the region case, including the two guards. The loop options follow the run the way the loop summary does, so a running loop's passes and their states no longer sit stale behind a five minute staleTime. Also fixes the task instance stream being drained before the summaries were built: without loops it is walked once, and the pass that looks for the newest iteration has nothing to find, so it is skipped rather than consuming it. * Take a row's loop iteration from the server, and count every pass The breadcrumb worked out which pass a page was on by reading region_index off the url. For a mapped expansion that index is the map slot rather than the pass, so it searched the pass list for a slot number, found nothing, and dropped the iteration level entirely -- leaving a mapped task inside a loop with no way to see or change its pass. A row already knows its own pass: loop_iteration is resolved server side and holds for an expansion and for a pass a clear forked into a successor region alike. Only a group, which is not one row, still reads the url. Map index and iteration are separate levels and now stay that way, rather than one url field being asked to mean both. State counts cover every pass. A looped task's cell stands for the work the loop did and the UI says how many iterations that covers, so an earlier pass put back to work by a clear is visible instead of hidden behind a finished later one. The coordinate still comes from the newest pass, since a link has to address a row. * Resolve a loop pass to the row it became, instead of guessing a coordinate Switching iteration on a task instance page wrote the loop's region and the pass number as if they addressed a row. They do not: a mapped expansion lives in its own region, and a clear forks later passes into a successor one. The pair named no row, and asking for a looped task instance that way is rejected rather than answered empty, so navigating between a loop's tasks could fail outright. The switcher now writes only what it knows -- the loop and the pass. The page resolves that to a coordinate through the list endpoint, which already takes the pair and returns each row's own region, and writes the answer back into the url. Everything downstream keeps reading the url unchanged, and the resolved link is the one worth sharing. The lookup sits on the page rather than in the shared coordinate hook, which nineteen components use: a query there would have every one of them make a request. The page holds its children while the lookup is in flight. The request has to be withheld rather than made and handled, since there is no scope to make it with. * Open the Gantt highlight where one run's segments meet An outline rings all four sides, so a run drawn as a queued stretch followed by its state showed the highlight down the seam between them -- an edge inside the thing being highlighted rather than its boundary. The sides are drawn individually now, and the two that already know they are touching a neighbour are left off. * Address review: group membership, paging, and the loop a summary is asked for The Gantt group highlight decided membership from the task id, so a group declared with prefix_group_id=False lost it. It reads the structure now, the way the loop scope already does. Switching pass restarts paging. A page of one pass says nothing about where the next one begins, so carrying its cursor across left the table on rows from the pass before. A task instance page no longer offers its own row's region as the loop to summarise. Only a group page's region names an invocation: a row's region is an expansion, or a pass a clear forked, and asking for a run summary by one of those is rejected -- which left the header falling back on a count of itself. * Highlight only a pass the reader chose in the Gantt The highlight fell back to the region index when no iteration was named. That index is the map slot on a mapped expansion, so it rang bars of whatever pass happened to share the number -- and resolving a pass to a row now writes a region index into the url far more often, which made the fallback fire where it used to be rare. * Keep earlier attempts on the Gantt chart after a clear Every attempt is its own row and only the newest stays live, and ORM selects read live rows unless they ask otherwise. The chart did not, so clearing a task took its earlier tries off the chart entirely -- including the attempt that the pending-retry filter just above deliberately drops the live row in favour of, which could never have been there to find. * Name the try in the Gantt tooltip The try sat inside a branch keyed on start_date, but the chart reports min_start_date, so it never reached a bar's tooltip at all. It is read off the shape a caller passes now, rather than from one of the branches below it. Shown on a first try too. Clearing a loop pass regenerates the later ones as fresh executions that start again at try 1, so two neighbouring bars can both be a first try and the number is what tells them apart. * Drop the iteration label from the mapped task instances header The counts beside it are the task's across the whole loop, so naming one pass above them said that pass had every expansion of every pass. This page lists the expansions of a mapped task, which is an aggregate view, and the pass it belongs to is already named in the breadcrumb directly above. Narrowing to a single pass is what the task instances table and its loop filter are for. * Regenerate the UI client and reformat after the review changes The spec changed when the execution route went back to its shape on main and the iteration sort was withdrawn, but the generated client still carried both, so the UI lint hook regenerated it and failed on the difference. Two Python files were left unformatted by the same round of edits. * Fix the type errors the UI static check compiles for The checked-in tsconfig does not typecheck src; the hook builds its own project extending tsconfig.app.json, which turns on noUnusedLocals among other things, so running tsc against the root config said clean while the hook did not. Two unused locals left behind by earlier edits, a key typed as a plain string where its neighbours are cast, and a tooltip reading try_number off a union whose grid summary member does not carry one -- lifting it out of the start_date branch lost the narrowing that had been doing that work. --------- (cherry picked from commit 42945d6) Co-authored-by: Brent Bovenzi <brent@astronomer.io>
This was referenced Oct 20, 2026
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.
Clearing a task instance from the UI on 3.4.0b1 does nothing after Confirm. The browser sends the
dry_run: truerequest toclearTaskInstancesand never the real one, so the task is not cleared. It happens from the task instance page and from the run's Execution tab. With "Prevent rerun if task is running" checked on a queued task, the "Cannot Clear Task Instance" warning never shows either. 3.3.2 is not affected.closes: #74712
Cause. Confirm hides the clear dialog and opens
ClearTaskInstanceConfirmationDialog, which re-runs the dry run, then either sends the clear or shows the queued-task warning. Agit bisectover the UI puts the break at #72524, the Chakra 3.36.1 to 3.37.0 bump, which moved zag's dialog, dismissable and focus-trap packages from 1.41.2 to 1.43.3 (3.3.2 still ships 1.41.2). With 1.43.3, the clear dialog closing underneath dismisses the new dialog before its dry run returns, through two separate paths:layerStack.remove()callsdismiss()on the nested layers).Fix. The confirmation dialog cancels both:
onRequestDismiss, which zag only fires for the layer-stack case, andonFocusOutside. Outside clicks and Escape still close it. Blocking only the layer-stack dismissal, or only outside interaction, still lost the clear in Chrome.Why guard the dismissals instead of restructuring the flow. Opening the second dialog only once there is a warning to show also works, but then nothing is on screen during the check. The page stays clickable, so on the Task Instances list a selection change during the check could widen the clear. Keeping the clear dialog open under the second one leaves two modals stacked; in the unit tests the lower one then closed itself and reset the note. The guards keep the 3.3 behaviour as it was.
Tests.
Execution.test.tsxmocked the confirmation dialog out because "jsdom dismisses [it] before its dry run settles"; that was this bug. The mock is removed. Two Execution tests fail against the old dialog without it and pass with the fix. The two new tests inClearTaskInstanceDialog.test.tsx(the clear goes out once a slow check resolves; the queued-task warning shows) also fail against the old dialog. happy-dom reproduces the layer-stack path but not the focus one, so I checked the fix in Chrome against a local Airflow. Confirm sendsdry_run: falsefrom the task instance page and the Execution tab. A queued task with "Prevent rerun if task is running" shows the warning, sends nothing, and closes with Escape or an outside click.Before and after. Clearing from the run's Execution tab. The before screenshot is from #74712: Confirm sends only the
dry_run: truerequest and the task stays Success. After the fix, Confirm clears the selected task and its downstream task, both drop to No Status, and the run goes back to Queued.With a queued task and "Prevent rerun if task is running" checked, the warning now shows. Before the fix the clear dialog closed and nothing appeared.
{pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.