Reviewing Pull Requests#

When reviewing PRs, our primary goal is to improve DataFusion and its community together. PR feedback should be constructive and help improve the code as well as the understanding of the contributor.

Review bandwidth is currently our most limited resource, and reviews from the broader community are both welcomed and encouraged. Reviewing PRs is a great way to learn the codebase, and you do not need to be a committer to leave valuable review feedback. In fact, one of the best ways to become a committer is to thoughtfully review other PRs.

Please ensure any comments you leave contain a rationale and suggested alternative – it is frustrating to be told “don’t do it this way” without any clear reason or alternative provided.

The criteria in this guide are also a useful checklist when preparing your own PR for review.

PR Review Mechanics#

Some helpful links:

The overall PR lifecycle (CI triggering, approval, the 24-hour rule for “major” PRs, and merging) is described in the Pull Request Overview section of the contributor guide.

Practical tips:

  1. Check out the changes locally to explore them in your IDE or with an agent, e.g. gh pr checkout <PR number> using the GitHub CLI.

  2. There is normally no need to rerun locally any tests that CI has already run.

  3. Leave comments on specific lines of the diff where possible, so the discussion has context.

  4. If you review a PR but don’t feel confident approving it, leaving comments is still valuable: a partial review (e.g. “I reviewed the tests and they look good”) helps the next reviewer focus their time.

  5. Anything that does not need to block the current PR can be noted as a potential follow-up (ideally by filing an issue), keeping the PR focused and quick to merge.

Review the PR Description#

The PR description is often what users and contributors will find when they have a question about the intention behind a change, or when the code itself is not clear. The PR description also becomes the extended commit message.

Check that the description:

  1. Concisely describes the problem being solved from the user’s point of view.

  2. Follows the PR template, and answers the template’s questions.

  3. Accurately describes the content of the PR, including any relevant context or background. Great descriptions have a high signal-to-noise ratio, summarizing important implementation changes without repeating technical minutiae that are already present in the code itself.

  4. Explicitly calls out any user-facing or API changes (see Review the Code below).

Review the Code Comments#

The goal of code comments is to help future readers of the code understand what is not obvious from reading the code itself. Great comments make the code easier to reason about for readers with the expected background, and help future maintainers.

Some practical guidelines for reviewing comments:

  1. The code has adequate comments focused on the rationale for any non-obvious change (the “why”), not a restatement of what the code does (the “what”), which is typically clear from reading the code itself.

  2. Comments do not narrate irrelevant internal implementation details or the history of how the change was developed (this is common in LLM-assisted code, e.g. “// changed to use a HashMap” or “// this handles the case mentioned above”). Such comments become irrelevant as soon as the PR merges.

  3. When comments refer to other structs, functions, or modules, they should use rustdoc intra-doc links (e.g. [`SessionContext`]) rather than plain text names, so that cargo doc link checking ensures the references stay valid as the code evolves.

  4. New public APIs have doc comments, including examples where appropriate (doc examples are also tested by CI, so they double as test coverage).

  5. When documenting modules, functions, or fields, start with simple examples and intuitive explanations, and optionally add formal, math-like definitions when necessary. This makes the implementation easier to reason about.

  6. When something is confusing on first read, treat that as a good opportunity to improve the comments.

Review the Test Coverage#

Check that the feature or fix is covered sufficiently with tests (see the Testing guide for more details): the PR should include tests for any new functionality, and a bug fix should include a test that reproduces the reported problem.

Guidelines for evaluating tests:

  1. Prefer sqllogictest (.slt) tests or DataFrame API tests where possible, as they exercise user-visible behavior and are less coupled to internal implementation details than unit tests.

  2. Verify tests cover edge cases and common failure scenarios, not just the common successful path. However, it is NOT necessary to test every possible error path, especially if it is difficult to trigger or unlikely to occur in practice.

  3. Verify test coverage of changed code using the codecov check on the PR, or by running cargo llvm-cov locally for an HTML report. Use judgment about any uncovered lines – the goal is confidence in the change, not slavishly hitting some coverage number.

  4. Avoid tests with lots of repeated boilerplate: when many tests share near-identical setup, it is hard to understand what is different (and thus what is actually being tested) between them. Make the difference between cases obvious.

  5. Check that tests assert on specific expected values or plans (e.g. via insta snapshots or .slt expected output) rather than merely checking “no error occurred”.

  6. Verify tests actually cover the bug (“Ablation Testing”): For bug fixes, revert the fix locally and check that the new test fails without it (i.e. the test actually reproduces the bug or covers the new feature).

Review the Code#

Check that:

  1. The code is clear and fits the style of the existing codebase.

  2. New functions and tests are placed near similar functions and tests. For example, helper functions should be defined close to where they are used, and new tests should be placed in the same module as the code they test. SLT tests should be placed in an existing .slt file with related functionality, unless the new tests are large enough to justify their own file.

  3. New APIs are consistent with existing public APIs and patterns; where a similar mechanism already exists, the PR should extend it rather than introduce a parallel one.

  4. Any changes to the public API follow the API health policy.

  5. The change is appropriately scoped: unrelated refactoring, formatting churn, or drive-by changes make review longer and are better as separate PRs.

  6. New errors are actionable, mention the offending item, and use the right error variant (e.g. plan_err! for user-triggerable errors vs internal_err! for invariant violations).

Review the Performance#

Performance is a key feature of DataFusion. See Performance Improvements for the project policy: an improvement should be “enough” to justify any added code complexity, and performance PRs should come with benchmark results.

When reviewing:

  1. Find any relevant existing benchmarks and run them against main: the system-level SQL benchmarks are run with bench.sh (see the benchmarks README), and microbenchmarks (e.g. in datafusion/functions/benches) are run with cargo bench.

  2. Be aware that benchmarking on a machine where other work is being done will make results hard to reproduce. Prefer a quiet, dedicated machine and repeated runs.

  3. If the PR claims a performance improvement, check that the reported results are reproducible and that the benchmark exercises the changed code path.

Best Practices for Reviewers#

Here are some suggested best practices to follow when reviewing PRs.

Review Tone: Thank Contributors and Praise Good Work Specifically#

Open reviews by thanking the author by name, and when a PR is well done, say specifically what makes it good – positive feedback encourages people to keep contributing and helps them understand what is valued in the project.

State Approval Conditions Explicitly#

If you are not ready to approve, list concretely what you would need to see before approving (e.g. “benchmark results and an upgrade guide entry”) so the author has a clear path to merge.

Defer Non-Blocking Work to Follow-On Issues#

Explicitly defer non-critical suggestions to a follow-on PR and file (or ask the author to file) issues for them, so good PRs merge quickly without scope creep.

Similarly, when a PR mixes refactoring with behavior changes or fixes a narrow problem with a broad mechanism, ask for it to be split or scoped down rather than reviewing it as-is.

Narrate What You Verified When Approving#

Rather than a bare “LGTM”, say what you actually checked (“traced the state transitions by hand”, “confirmed the hasher change cannot affect ordering”) so it is clear what was verified and what was not.

Invite Additional Committers on Core Changes#

For changes to core, widely shared code, leave the PR open for other committers to look at and cc those who know the area, even after you have approved.