Adopt main as the fork's shipping branch - #3
Merged
Merged
Conversation
FORK.md documents the branch model; fetch_client.py unpacks the prebuilt web client from a published wheel, since we have no node toolchain. Neither is upstreamable, so both live on a topic branch rather than on the main mirror. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VYsKhgGuNCFtxqFhSLGtCK
`X.sum(axis=0)` and `X.mean(axis=0)` on a scipy sparse matrix return `np.matrix`,
whose `.flatten()` yields a (1, n) matrix rather than a 1-D array, so building the
DataFrame raised:
ValueError: Data must be 1-dimensional, got ndarray of shape (1, 1)
This breaks every request that carries `stats`, i.e. the composition/summary views.
Dense `X` failed too, on the `.toarray()`/`.getnnz()` calls that only exist on
sparse matrices.
Reduce over each layout explicitly and use `np.asarray(...).ravel()`, which is 1-D
for both `np.matrix` and `ndarray`. Adds `tests/test_stats.py`, covering `X_stats`
and `FeatureAggregator` over sparse and dense inputs -- neither had any coverage,
which is how this survived.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X6uSHPpkfbneXYsprQ3wJ6
Both call sites in `data_processing` pass an AnnData -- the second one even names
the local `df` after `apply_filter(adata, ...)` returns one -- but `execute` was
written against a DataFrame, indexing measures as columns. So every request
carrying `groupedStats` fails:
KeyError: '<grouping column>'
which is the whole dot plot / distribution view.
Gather the requested dimensions from `adata.obs` and the measures from `adata.X`
into a frame first, matching how `FeatureAggregator` already consumes AnnData.
Only the requested genes are densified, so the cost stays proportional to the
request rather than the dataset.
Also restrict the groupby aggregation to the measure columns. With more than one
dimension the frame still holds the individual categorical columns, and
aggregating those raised `TypeError: 'Categorical' with dtype category does not
support reduction 'mean'`, breaking multi-dimension dot plots.
Adds `tests/test_dotplot.py` over sparse and dense inputs, checking the values
against a manual groupby -- the aggregator had no coverage at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X6uSHPpkfbneXYsprQ3wJ6
anndata >= 0.13 exposes ``None`` in ``layers`` as an alias for ``X``, so every call site that iterates the mapping directly sees a layer that is not one. The parquet writer then does os.path.join(..., "layers", None) and raises TypeError, and the zarr writer writes a layer literally named "None" that duplicates X. Route the six call sites through one ``layer_names`` helper so the bug class is gone rather than the instances. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VYsKhgGuNCFtxqFhSLGtCK
anndata labels what it writes with an encoding-type, and several of those are groups holding several nodes -- a categorical (codes + categories since 0.8), a nullable string, a nullable integer. The backed readers assume every obs/var node answers node[...], so such a column could not be read at all, and a var index written as a nullable string took the whole store down. Decode anything carrying an encoding-type with read_elem, which handles all of them for h5py and zarr alike, and drop the two remaining calls into the private anndata._io.zarr.read_attribute for the same reason. Also skip the obs index in dataset_schema (it is not a column) and tolerate DataFrame-valued obsm, which has no .shape when the store is read backed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VYsKhgGuNCFtxqFhSLGtCK
Removing the zarr<3 pin is not enough on its own: zarr 3 dropped attribute access to group members, so ZarrDataset needs a small adapter, and the vendored 2021-era copy of anndata's zarr writer uses the v2 creation API throughout. Delete that copy (250 lines) and write through anndata.io.write_elem instead. It is what anndata itself uses, it tracks the encoding versions the readers expect, and it removes the duplicated encoder entirely. Reads keep going through the caller's fsspec mapper, which zarr 3 still accepts. Writes take the path, because zarr creates the directories it writes into and an fsspec LocalFileSystem mapper does not. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VYsKhgGuNCFtxqFhSLGtCK
An octopus merge refuses the moment any pair of branches conflicts, and these touch the same files. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two independent breakages, both from pandas 3 defaults. pandas 3 backs string columns with ArrowStringArray, which the JSON encoder recurses into until it raises "OverflowError: Maximum recursion level reached". Every categorical obs column hits this, so `prepare_data --format jsonl` cannot write a dataset at all. Coerce extension arrays in `dumps` itself rather than at each call site, so the next serialised column does not reintroduce it. Separately, `test_de_4_groups` builds its fixture with Series.replace on a categorical, which pandas 3 refuses because the new values are not existing categories. rename_categories expresses what the test means. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VYsKhgGuNCFtxqFhSLGtCK
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`cirrocumulus/client` symlinks to gitignored `build/`, so installing from git gives a server with no UI and consumers cannot declare cirro as a dependency. CI now builds the client and attaches a wheel to a GitHub Release. Drops `scripts/fetch_client.py`, which scavenged the client out of a published upstream wheel. That shortcut also pinned us to upstream's client and forbade any change under `src/`. FORK.md's claim that Euler has no node module was wrong: `node-js/22.4.0` under `stack/2025-06` builds it fine. `scripts/check_wheel.py` is the gate -- an empty `build/` otherwise yields a client-less wheel with no error at all, only a 404 on `/` at runtime. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CQTJRwSGPC77SCzqraZ92T
Both restated facts the other already owned -- the release section walked through the workflow file step by step, and the docstring re-explained the silent-empty-build failure that FORK.md now covers. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CQTJRwSGPC77SCzqraZ92T
The same `dict(values=..., categories=series.cat.categories.values)` construct appears twice; only the jsonl_io one was coerced. Under pandas 3 the categories are an ArrowStringArray, so every categorical colour-by returned 500 and the UI just said "Unable to retrieve data". The `dumps` wrapper cannot catch this one: the array sits three levels down and the wrapper only walks the top level. Test fails without the fix on the bundled pbmc3k data. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CQTJRwSGPC77SCzqraZ92T
The row described only the jsonl_io site; the /api/data one is what actually blocked the viewer. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CQTJRwSGPC77SCzqraZ92T
# Conflicts: # cirrocumulus/anndata_util.py # tests/test_prepare_data.py
# Conflicts: # cirrocumulus/zarr_output.py
# Conflicts: # cirrocumulus/jsonl_io.py
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
1 participant
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.
Makes
mainthe fork's shipping branch and retires the force-pushedintegrationbranch.Upstream is maintained but does not take patches: of the 100 most recently merged PRs on
lilab-bcb/cirrocumulus — back to Jan 2023 — every one came from lilab-bcb itself, and the only
outside contributor PR has been open since Oct 2024. The mirror-plus-
integrationmodel was builton the assumption that our PRs would land upstream. They will not, so the fork's own
mainbecomeswhat we build, test and release.
This PR carries the seven topic branches that were already merged into
integration, in mergeorder, with their commits and merge commits intact:
fork/toolingscripts/check_wheel.py, the release workflowfix/x-stats-sparse-denseX_statsreturned 2-D columns for sparseX, 500ing the composition view; denseXunhandledfix/dotplot-aggregator-anndataDotPlotAggregatorwas written for a DataFrame but is passed an AnnData, 500ing the dot plot viewfix/anndata-013-none-layerNoneinlayersas an alias forX; six call sites treated it as a layerfix/zarr-categorical-obsfix/zarr3anndata.io.write_elemfix/pandas3ArrowStringArrayoverflowed the JSON encoder; brokeprepare_data --format jsonland 500'd/api/datafor categoricalobsAll six upstream PRs stay open. Their branches are kept, since deleting a head branch closes its PR.
Merge this with "Create a merge commit" — not squash, not rebase — so the seven merge commits
and their SHAs survive. If upstream ever merges one of those PRs, shared SHAs keep the next
git merge upstream/maintrivial.Follow-ups, in order: retarget
release.ymland rewriteFORK.mdfor the new model; apply ruffformatting (the fork's code predates upstream's black-100 → ruff-88 switch in lilab-bcb#234); add a ruleset
on
main; deleteintegration.