UI: Fix Graph layout for TaskGroup tasks wired to external nodes by vatsrahul1001 · Pull Request #67720 · apache/airflow

vatsrahul1001 · 2026-05-29T14:31:11Z

Closes#67714.

#65031 could be cause of it

Fix

Tighten hasUniformExternalConnectivity to require every externally-connected child to share the same full (sources, targets) profile. Mixed entry/exit shapes (like the reporter's repro) now fall back to rendering individual crossing edges, preserving the author's explicit dependency intent. The canonical "cleanup group" fan-in/fan-out pattern that the optimization was designed for (every child has same upstream AND same downstream) still triggers the collapse.
Add a preserveInternal option to rewriteGroupEdges. The open-group call site sets it so the collapse optimization preserves internal edges for the subsequent extraction loop; the closed-group call site keeps the default (drop internals — they're not laid out when the group is collapsed).

Tests

New elkGraphUtils.test.ts covers:

hasUniformExternalConnectivity — vanilla TaskGroup, mixed-profile 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 trigger, canonical fan-in/fan-out
generateElkGraph for the 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 repro: internal edges preserved AND escape edges rendered individually
generateElkGraph for the canonical fan-in/fan-out shape: collapse still fires, internal edges kept (exercises preserveInternal: true)
generateElkGraph regression guards: simple open TaskGroup, closed TaskGroup

Verified the new tests fail against main without the fix and pass with it.

Was generative AI tooling used to co-author this PR?

Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines

Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. Closes: #67714

vatsrahul1001 · 2026-05-29T14:34:52Z

cc: @pierrejeambrun @bbovenzi

pierrejeambrun

Tested locally looks good. Code looks good.

github-actions · 2026-06-01T09:51:36Z

Backport successfully created: v3-2-test

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status	Branch	Result
✅	v3-2-test

…l nodes (apache#67720) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from apache#65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in apache#67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com> Closes: apache#67714

…l nodes (#67720) (#67830) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Closes: #67714 Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com>

vatsrahul1001 · 2026-05-29T14:31:11Z

Closes#67714.

#65031 could be cause of it

Fix

Tighten hasUniformExternalConnectivity to require every externally-connected child to share the same full (sources, targets) profile. Mixed entry/exit shapes (like the reporter's repro) now fall back to rendering individual crossing edges, preserving the author's explicit dependency intent. The canonical "cleanup group" fan-in/fan-out pattern that the optimization was designed for (every child has same upstream AND same downstream) still triggers the collapse.
Add a preserveInternal option to rewriteGroupEdges. The open-group call site sets it so the collapse optimization preserves internal edges for the subsequent extraction loop; the closed-group call site keeps the default (drop internals — they're not laid out when the group is collapsed).

Tests

New elkGraphUtils.test.ts covers:

hasUniformExternalConnectivity — vanilla TaskGroup, mixed-profile 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 trigger, canonical fan-in/fan-out
generateElkGraph for the 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 repro: internal edges preserved AND escape edges rendered individually
generateElkGraph for the canonical fan-in/fan-out shape: collapse still fires, internal edges kept (exercises preserveInternal: true)
generateElkGraph regression guards: simple open TaskGroup, closed TaskGroup

Verified the new tests fail against main without the fix and pass with it.

Was generative AI tooling used to co-author this PR?

Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines

Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. Closes: #67714

vatsrahul1001 · 2026-05-29T14:34:52Z

cc: @pierrejeambrun @bbovenzi

pierrejeambrun

Tested locally looks good. Code looks good.

github-actions · 2026-06-01T09:51:36Z

Backport successfully created: v3-2-test

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status	Branch	Result
✅	v3-2-test

…l nodes (apache#67720) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from apache#65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in apache#67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com> Closes: apache#67714

…l nodes (#67720) (#67830) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Closes: #67714 Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com>

vatsrahul1001 · 2026-05-29T14:31:11Z

Closes#67714.

#65031 could be cause of it

Fix

Tighten hasUniformExternalConnectivity to require every externally-connected child to share the same full (sources, targets) profile. Mixed entry/exit shapes (like the reporter's repro) now fall back to rendering individual crossing edges, preserving the author's explicit dependency intent. The canonical "cleanup group" fan-in/fan-out pattern that the optimization was designed for (every child has same upstream AND same downstream) still triggers the collapse.
Add a preserveInternal option to rewriteGroupEdges. The open-group call site sets it so the collapse optimization preserves internal edges for the subsequent extraction loop; the closed-group call site keeps the default (drop internals — they're not laid out when the group is collapsed).

Tests

New elkGraphUtils.test.ts covers:

hasUniformExternalConnectivity — vanilla TaskGroup, mixed-profile 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 trigger, canonical fan-in/fan-out
generateElkGraph for the 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 repro: internal edges preserved AND escape edges rendered individually
generateElkGraph for the canonical fan-in/fan-out shape: collapse still fires, internal edges kept (exercises preserveInternal: true)
generateElkGraph regression guards: simple open TaskGroup, closed TaskGroup

Verified the new tests fail against main without the fix and pass with it.

Was generative AI tooling used to co-author this PR?

Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines

Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. Closes: #67714

vatsrahul1001 · 2026-05-29T14:34:52Z

cc: @pierrejeambrun @bbovenzi

pierrejeambrun

Tested locally looks good. Code looks good.

github-actions · 2026-06-01T09:51:36Z

Backport successfully created: v3-2-test

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status	Branch	Result
✅	v3-2-test

…l nodes (apache#67720) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from apache#65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in apache#67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com> Closes: apache#67714

…l nodes (#67720) (#67830) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Closes: #67714 Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com>

vatsrahul1001 · 2026-05-29T14:31:11Z

Closes#67714.

#65031 could be cause of it

Fix

Tighten hasUniformExternalConnectivity to require every externally-connected child to share the same full (sources, targets) profile. Mixed entry/exit shapes (like the reporter's repro) now fall back to rendering individual crossing edges, preserving the author's explicit dependency intent. The canonical "cleanup group" fan-in/fan-out pattern that the optimization was designed for (every child has same upstream AND same downstream) still triggers the collapse.
Add a preserveInternal option to rewriteGroupEdges. The open-group call site sets it so the collapse optimization preserves internal edges for the subsequent extraction loop; the closed-group call site keeps the default (drop internals — they're not laid out when the group is collapsed).

Tests

New elkGraphUtils.test.ts covers:

hasUniformExternalConnectivity — vanilla TaskGroup, mixed-profile 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 trigger, canonical fan-in/fan-out
generateElkGraph for the 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 repro: internal edges preserved AND escape edges rendered individually
generateElkGraph for the canonical fan-in/fan-out shape: collapse still fires, internal edges kept (exercises preserveInternal: true)
generateElkGraph regression guards: simple open TaskGroup, closed TaskGroup

Verified the new tests fail against main without the fix and pass with it.

Was generative AI tooling used to co-author this PR?

Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines

Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. Closes: #67714

vatsrahul1001 · 2026-05-29T14:34:52Z

cc: @pierrejeambrun @bbovenzi

pierrejeambrun

Tested locally looks good. Code looks good.

github-actions · 2026-06-01T09:51:36Z

Backport successfully created: v3-2-test

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status	Branch	Result
✅	v3-2-test

…l nodes (apache#67720) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from apache#65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in apache#67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com> Closes: apache#67714

…l nodes (#67720) (#67830) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Closes: #67714 Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com>

vatsrahul1001 · 2026-05-29T14:31:11Z

Closes#67714.

#65031 could be cause of it

Fix

Tighten hasUniformExternalConnectivity to require every externally-connected child to share the same full (sources, targets) profile. Mixed entry/exit shapes (like the reporter's repro) now fall back to rendering individual crossing edges, preserving the author's explicit dependency intent. The canonical "cleanup group" fan-in/fan-out pattern that the optimization was designed for (every child has same upstream AND same downstream) still triggers the collapse.
Add a preserveInternal option to rewriteGroupEdges. The open-group call site sets it so the collapse optimization preserves internal edges for the subsequent extraction loop; the closed-group call site keeps the default (drop internals — they're not laid out when the group is collapsed).

Tests

New elkGraphUtils.test.ts covers:

hasUniformExternalConnectivity — vanilla TaskGroup, mixed-profile 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 trigger, canonical fan-in/fan-out
generateElkGraph for the 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 repro: internal edges preserved AND escape edges rendered individually
generateElkGraph for the canonical fan-in/fan-out shape: collapse still fires, internal edges kept (exercises preserveInternal: true)
generateElkGraph regression guards: simple open TaskGroup, closed TaskGroup

Verified the new tests fail against main without the fix and pass with it.

Was generative AI tooling used to co-author this PR?

Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines

Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. Closes: #67714

vatsrahul1001 · 2026-05-29T14:34:52Z

cc: @pierrejeambrun @bbovenzi

pierrejeambrun

Tested locally looks good. Code looks good.

github-actions · 2026-06-01T09:51:36Z

Backport successfully created: v3-2-test

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status	Branch	Result
✅	v3-2-test

…l nodes (apache#67720) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from apache#65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in apache#67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com> Closes: apache#67714

…l nodes (#67720) (#67830) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Closes: #67714 Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com>

vatsrahul1001 · 2026-05-29T14:31:11Z

Closes#67714.

#65031 could be cause of it

Fix

Tighten hasUniformExternalConnectivity to require every externally-connected child to share the same full (sources, targets) profile. Mixed entry/exit shapes (like the reporter's repro) now fall back to rendering individual crossing edges, preserving the author's explicit dependency intent. The canonical "cleanup group" fan-in/fan-out pattern that the optimization was designed for (every child has same upstream AND same downstream) still triggers the collapse.
Add a preserveInternal option to rewriteGroupEdges. The open-group call site sets it so the collapse optimization preserves internal edges for the subsequent extraction loop; the closed-group call site keeps the default (drop internals — they're not laid out when the group is collapsed).

Tests

New elkGraphUtils.test.ts covers:

hasUniformExternalConnectivity — vanilla TaskGroup, mixed-profile 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 trigger, canonical fan-in/fan-out
generateElkGraph for the 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 repro: internal edges preserved AND escape edges rendered individually
generateElkGraph for the canonical fan-in/fan-out shape: collapse still fires, internal edges kept (exercises preserveInternal: true)
generateElkGraph regression guards: simple open TaskGroup, closed TaskGroup

Verified the new tests fail against main without the fix and pass with it.

Was generative AI tooling used to co-author this PR?

Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines

Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. Closes: #67714

vatsrahul1001 · 2026-05-29T14:34:52Z

cc: @pierrejeambrun @bbovenzi

pierrejeambrun

Tested locally looks good. Code looks good.

github-actions · 2026-06-01T09:51:36Z

Backport successfully created: v3-2-test

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status	Branch	Result
✅	v3-2-test

…l nodes (apache#67720) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from apache#65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in apache#67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com> Closes: apache#67714

…l nodes (#67720) (#67830) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Closes: #67714 Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com>

vatsrahul1001 · 2026-05-29T14:31:11Z

Closes#67714.

#65031 could be cause of it

Fix

Tighten hasUniformExternalConnectivity to require every externally-connected child to share the same full (sources, targets) profile. Mixed entry/exit shapes (like the reporter's repro) now fall back to rendering individual crossing edges, preserving the author's explicit dependency intent. The canonical "cleanup group" fan-in/fan-out pattern that the optimization was designed for (every child has same upstream AND same downstream) still triggers the collapse.
Add a preserveInternal option to rewriteGroupEdges. The open-group call site sets it so the collapse optimization preserves internal edges for the subsequent extraction loop; the closed-group call site keeps the default (drop internals — they're not laid out when the group is collapsed).

Tests

New elkGraphUtils.test.ts covers:

hasUniformExternalConnectivity — vanilla TaskGroup, mixed-profile 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 trigger, canonical fan-in/fan-out
generateElkGraph for the 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 repro: internal edges preserved AND escape edges rendered individually
generateElkGraph for the canonical fan-in/fan-out shape: collapse still fires, internal edges kept (exercises preserveInternal: true)
generateElkGraph regression guards: simple open TaskGroup, closed TaskGroup

Verified the new tests fail against main without the fix and pass with it.

Was generative AI tooling used to co-author this PR?

Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines

Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. Closes: #67714

vatsrahul1001 · 2026-05-29T14:34:52Z

cc: @pierrejeambrun @bbovenzi

pierrejeambrun

Tested locally looks good. Code looks good.

github-actions · 2026-06-01T09:51:36Z

Backport successfully created: v3-2-test

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status	Branch	Result
✅	v3-2-test

…l nodes (apache#67720) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from apache#65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in apache#67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com> Closes: apache#67714

…l nodes (#67720) (#67830) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Closes: #67714 Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com>

vatsrahul1001 · 2026-05-29T14:31:11Z

Closes#67714.

#65031 could be cause of it

Fix

Tighten hasUniformExternalConnectivity to require every externally-connected child to share the same full (sources, targets) profile. Mixed entry/exit shapes (like the reporter's repro) now fall back to rendering individual crossing edges, preserving the author's explicit dependency intent. The canonical "cleanup group" fan-in/fan-out pattern that the optimization was designed for (every child has same upstream AND same downstream) still triggers the collapse.
Add a preserveInternal option to rewriteGroupEdges. The open-group call site sets it so the collapse optimization preserves internal edges for the subsequent extraction loop; the closed-group call site keeps the default (drop internals — they're not laid out when the group is collapsed).

Tests

New elkGraphUtils.test.ts covers:

hasUniformExternalConnectivity — vanilla TaskGroup, mixed-profile 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 trigger, canonical fan-in/fan-out
generateElkGraph for the 3.2.2 Graph View breaks when TaskGroup outputs connect internal tasks to external downstream tasks #67714 repro: internal edges preserved AND escape edges rendered individually
generateElkGraph for the canonical fan-in/fan-out shape: collapse still fires, internal edges kept (exercises preserveInternal: true)
generateElkGraph regression guards: simple open TaskGroup, closed TaskGroup

Verified the new tests fail against main without the fix and pass with it.

Was generative AI tooling used to co-author this PR?

Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines

Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. Closes: #67714

vatsrahul1001 · 2026-05-29T14:34:52Z

cc: @pierrejeambrun @bbovenzi

pierrejeambrun

Tested locally looks good. Code looks good.

github-actions · 2026-06-01T09:51:36Z

Backport successfully created: v3-2-test

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status	Branch	Result
✅	v3-2-test

…l nodes (apache#67720) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from apache#65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in apache#67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com> Closes: apache#67714

…l nodes (#67720) (#67830) Open ``@task_group`` rendered with vertically-stacked internals and edges crossing the boundary whenever an internal task had a direct dependency on a node outside the group (an "escape edge" that bypassed the group's entry/exit interface). Dag execution was unaffected. Two underlying issues, both in the ELK graph-layout refactor from #65031: 1. ``hasUniformExternalConnectivity`` was too lenient — it fired whenever externally-connected children separately shared the same external sources OR the same external targets, instead of the canonical fan-in/fan-out pattern where every child has the same full ``(sources, targets)`` profile. On mixed-profile groups (entry + exits), it incorrectly fired and collapsed the author's deliberately- wired escape edges into a single group-level edge, hiding the intent. 2. When the optimisation did fire on an open group, ``rewriteGroupEdges`` was tuned for closed groups and dropped the group's internal edges too, leaving ELK with no internal-layout information for the children (the visible symptom in #67714). Fix: tighten ``hasUniformExternalConnectivity`` to require the full profile to match across externally-connected children, and add a ``preserveInternal`` option to ``rewriteGroupEdges`` so the canonical fan-in/fan-out path keeps internals intact. (cherry picked from commit 55780f2) Closes: #67714 Co-authored-by: Rahul Vats <43964496+vatsrahul1001@users.noreply.github.com>

vatsrahul1001 requested review from bbovenzi, choo121600, guan404ming, pierrejeambrun, ryanahamilton and shubhamraj-git as code owners May 29, 2026 14:31

vatsrahul1001 added kind:bug This is a clearly a bug area:UI Related to UI/UX. For Frontend Developers. area:TaskGroup backport-to-v3-2-test labels May 29, 2026

pierrejeambrun approved these changes Jun 1, 2026
View reviewed changes

pierrejeambrun added this to the Airflow 3.2.3 milestone Jun 1, 2026

pierrejeambrun merged commit 55780f2 into mainJun 1, 2026
83 checks passed

pierrejeambrun deleted the fix-graph-taskgroup-internal-edges-dropped branch June 1, 2026 09:50

vatsrahul1001 modified the milestones: Airflow 3.2.3, Airflow 3.3.0 Jun 8, 2026