Repository navigation
Core, Spark: Fix equality deletes non-deterministic schema ordering (#13873)(Backports #15514) - #15605
Merged
amogh-jahagirdar merged 1 commit intoMar 12, 2026
Conversation
…pache#13873) Equality delete schemas constructed in DeleteFilter.applyEqDeletes relied on the field order of requiredSchema, which varies depending on the query's projection. When the SparkExecutorCache returned delete records read with one field ordering to a reader expecting another, StructProjection silently misinterpreted the positional data, causing deletes to be skipped. We fix this by Canonicalize the deleteSchema by sorting fields by field ID. Now every reader produces the same schema for deletes regardless of projection, ensuring cache hits return correctly ordered records. Coded with the help of Cursor and claude-4.6.opus-high
amogh-jahagirdar
approved these changes
Mar 12, 2026
amogh-jahagirdar
left a comment
amogh-jahagirdar
left a comment
Contributor
There was a problem hiding this comment.
Looks like a clean backport, thank you @RussellSpitzer!
singhpk234
approved these changes
Mar 12, 2026
huaxingao
approved these changes
Mar 12, 2026
3 of 8 tasks
wjxiz1992
added a commit
to NVIDIA/cudf-spark
that referenced
this pull request
Sep 15, 2026
…ed-it] [databricks] (#15964) **JaCoCo production line coverage: not fully measurable locally — no compatible Scala 2.13 / shim 411 nightly baseline is published.** Contributes to #12885. ### Description This test-only recovery restores enforced CPU/GPU result checks for mixed Iceberg position and equality deletes on Iceberg 1.10.2 and newer. Before this change, the format-v2 and format-v3 cases were permanent expected failures with fixed data seeds even on runtimes containing the upstream equality-delete cache fix. The affected older releases are now skipped instead of spending test time on known failures; production behavior is unchanged. [Apache Iceberg #15514](apache/iceberg#15514), released in 1.11.0 and [backported to 1.10.2](apache/iceberg#15605), fixes equality-delete records being cached in a query-dependent field order. This PR: - Skips both mixed-delete tests on Iceberg versions older than 1.10.2, removes their expected-failure markers and permanent seed overrides on fixed versions, and fails pytest configuration when `--iceberg` lacks an expected runtime version. - Passes the expected Iceberg version through the Spark 4.0 extra-classpath pre-merge launcher and documents the required environment variable. - Keeps every count, full-row, and GPU-scan assertion. For the format-v3 count query only, disables adaptive query execution so plan validation sees the completed scan, then resets session configuration before fixture cleanup. Older supported dependencies are not fixed by this PR, so #12885 should remain open. Iceberg 1.10.2 was policy-checked but was not GPU-runtime-tested locally. Shared pytest configuration and pre-merge launcher paths changed, so Databricks remains in CI scope. #### Validation - Spark 4.1.1 / Iceberg 1.11.0 / Python 3.10: all six focused format-v2/format-v3 CPU/GPU cases passed across the three reader modes (`6 passed, 45 deselected` in 219.85 seconds). Full-row equality and the format-v3 `GpuBatchScanExec` assertion were exercised. - Spark 4.0.0 / Iceberg 1.10.1 / Python 3.10: all six focused cases were selected and skipped (`6 skipped, 45 deselected` in 1.73 seconds). - Seven release-boundary checks and two decorator-policy checks passed. The checked cases cover missing version, 1.6.1, 1.9.2, 1.10.0, 1.10.1, 1.10.2, and 1.11.0. - Unset `EXPECTED_ICEBERG_VERSION` with `--iceberg` failed during pytest configuration with exit 4 and the expected error. `bash -n` and a focused launcher probe confirmed that the Spark 4.0 extra-classpath path forwards Iceberg 1.10.1. - Python 3.10 compilation and `git diff --check` passed. No production performance change is expected. AI assistance: The change and PR description were prepared with Codex assistance. ### Checklists Documentation - [ ] Updated for new or modified user-facing features or behaviors - [x] No user-facing change Testing - [x] Added or modified tests to cover new code paths - [ ] Covered by existing tests (Please provide the names of the existing tests in the PR description.) - [ ] Not required Performance - [ ] Tests ran and results are added in the PR description - [ ] Issue filed with a link in the PR description - [x] Not required --------- Signed-off-by: Allen Xu <allxu@nvidia.com>
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
4 participants
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.
Equality delete schemas constructed in DeleteFilter.applyEqDeletes relied on the field order of requiredSchema, which varies depending on the query's projection. When the SparkExecutorCache returned delete records read with one field ordering to a reader expecting another, StructProjection silently misinterpreted the positional data, causing deletes to be skipped.
We fix this by Canonicalize the deleteSchema by sorting fields by field ID. Now every reader produces the same schema for deletes regardless of projection, ensuring cache hits return correctly ordered records.
Coded with the help of Cursor and claude-4.6.opus-high