Speed up TaskGroup.topological_sort with int-indexed projected sweep#67288
Open
shahar1 wants to merge 7 commits into
Open
Speed up TaskGroup.topological_sort with int-indexed projected sweep#67288shahar1 wants to merge 7 commits into
shahar1 wants to merge 7 commits into
Conversation
shahar1
force-pushed
the
taskgroup-topo-sort-projection
branch
from
90e10cc to
204a0e0
Compare
Contributor
There was a problem hiding this comment.
Pull request overview
This PR optimizes TaskGroup.topological_sort (and the serialized equivalent) by pre-projecting upstream dependencies to sibling indices once per call, then performing a sweep using an efficient emitted-flag structure. This aims to reduce repeated per-edge work during sorting while preserving the existing emission order.
Changes:
- Replaced the previous modified-Kahn loop in
TaskGroup.topological_sortwith a “project then sweep” implementation, including explicit cycle detection. - Mirrored the same algorithm into
SerializedTaskGroup.topological_sortfor consistency/performance. - Added Task SDK tests that validate topological correctness across multiple DAG shapes, plus a newsfragment entry.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| task-sdk/tests/task_sdk/definitions/test_taskgroup.py | Adds cross-shape correctness tests for TaskGroup.topological_sort. |
| task-sdk/src/airflow/sdk/definitions/taskgroup.py | Implements projected dependency computation and sweep-based topo sort in Task SDK TaskGroup. |
| airflow-core/src/airflow/serialization/definitions/taskgroup.py | Mirrors the new topo sort algorithm for serialized task groups and adds cycle detection behavior. |
| airflow-core/newsfragments/67288.improvement.rst | Documents the TaskGroup.topological_sort performance improvement. |
The previous modified-Kahn implementation re-derived each child's upstream edges (materializing ``upstream_list`` and walking ``parent_group``) on every outer-loop pass. Project per-task upstream IDs onto sibling-level integer indices once up front, then run a greedy multi-pass sweep against that projection with a ``bytearray`` emission flag. Emission order is identical to the previous implementation; existing order-sensitive tests cover the contract. Same change is mirrored in SerializedTaskGroup.
Adds a round-trip test for SerializedTaskGroup.topological_sort (the serialization variant was previously untested), rewrites the newsfragment in user-facing terms, and cleans up a stale reference and type annotation in the task-sdk shape tests.
Per Copilot review feedback, the previous wording said cycles are caught at "deserialization time" — but DAG.check_cycle runs at DAG parse time (via dagbag loading), not during from_dict/from_json. Reword to describe cycles reaching this code path as malformed serialized data, with the defensive ValueError still raised on detection.
Tighten _project_child_deps and the topological_sort body: drop the per-edge child_idx comparison in favour of a single set.discard at the end, inline the get_task/.task_group chain, and replace the pre-alloc loop with a list comprehension. Net -15 lines per file with a small (~10%) speedup on the layered shape and no regression elsewhere. The hot _sweep_projection inner loop is left intact — earlier attempts to extract its duplicated body into a closure cost 10-30% on independent / layered shapes and were reverted.
shahar1
force-pushed
the
taskgroup-topo-sort-projection
branch
from
62325c1 to
411f90e
Compare
Drop n=500 from the parametrize grid. The algorithm has no n-dependent branches, so n=100 covers every code path; n=500 only re-runs the same loops with more iterations and added ~0.7s to the file without exercising new correctness behaviour (per Copilot review feedback on PR apache#67288).
shahar1
requested review from
amoghrajesh,
ashb,
bolkedebruin and
kaxil
as code owners
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.
Summary
Replaces the modified-Kahn body of
TaskGroup.topological_sort(and the mirror inSerializedTaskGroup) with a projected sweep:upstream_task_idsis projected onto sibling-level integer indices once up front (_project_child_deps)._sweep_projection) then runs against the projection with abytearray-backed emission flag - per-edge work happens once per sort instead of once per outer-loop pass.Emission order is identical to the previous implementation. The existing order-sensitive tests (
test_topological_sort1/2,test_topological_nested_groups,test_topological_group_dep) cover the contract.Benchmark
Min of five runs, N=2000 children, ms per call. Source: https://gist.github.com/shahar1/9c61dc9f34f7e77cd29cfb9d67af7ceb
The 7.8x on rev-chain comes from the projection precompute alone (one projection vs N projections); algorithmic complexity is unchanged and remains O(N²) on that shape until the follow-up PR.
The nested row times a recursive walk across all 39 nested groups (depth=3, ~222 tasks each); the root-only timing was dominated by setup overhead and was previously labelled
~noisein this table.Test plan
task-sdk/tests/task_sdk/definitions/test_taskgroup.py+test_dag.py— 145 passed (18 new shape-correctness cases)airflow-core/tests/unit/utils/test_task_group.py— 20 passed (1 new serialization round-trip case)airflow-core/tests/unit/serialization/(task-group / topological filter) — 4 passedprek run mypy-task-sdk— passedprek run mypy-airflow-core— passedprek run --from-ref upstream/main --stage pre-commit— passedWas generative AI tooling used to co-author this PR?
Generated-by: Claude Code (Opus 4.7) following the guidelines