[conradbzura]

Summary

Add cfdb dump duckdb, which writes the materialized files collection — a denormalized document projection shaped for serving one file's metadata — into a single relational DuckDB database shaped for interrogating the corpus. The dump recovers the C2M2 entity and junction tables the materializer joined away, keyed by natural (id_namespace, local_id) composites throughout, and resolves the cross-consortium enrichment question in two tiers: a field with a common semantic slot occupies a promoted column (genome_assembly, output_type, experiment_type, analyte_class, …), and the residue occupies 1:1 per-consortium extension tables that omit verbatim-promoted fields while retaining any field whose promotion is derived rather than verbatim (replicate_info survives because the replicate columns are regex-extracted from it). The schema of record is the checked-in src/cfdb/dump/schema.sql, versioned by its schema_version header — an artifact in the same sense as schema.graphql.

Three loader policies carry the correctness argument. Entity rows repeated across documents resolve first-write-wins, with every divergent later copy counted and a bounded sample written to dump_conflict — divergence is reported, never silently reconciled. Junction rows take the union over all copies, a deliberate asymmetry: a divergent entity column is observable in dump_conflict, whereas a relationship dropped by first-wins would vanish without trace. A key resolving to two submissions aborts the dump, because the schema's global natural keys rest on id_namespaces being DCC-scoped, and the loader asserts that premise rather than assuming it — on the stub paths as much as the ordinary ones — and duplicate file keys abort on the same grounds. The policy's boundary is uniform: model-valid dirt that cannot satisfy the schema, such as a CV term without an id or a name, is skipped and counted at the decision point rather than surfacing as a constraint crash at finalize, and any residual finalize-time constraint violation is translated into a DumpError naming the table. Document validation reuses FileMetadataModel — the same hydration path the GraphQL resolvers run — so the dump and the API read the corpus through one set of read validators.

The design decisions follow the resolutions recorded against the issue's open questions: natural composite keys rather than surrogates, submission as a non-key column, globally deduped ontology tables, and an on-demand CLI with Parquet export. Scheduled publication of dump artifacts remains out of scope.

Closes #123

Proposed changes

Schema (src/cfdb/dump/schema.sql)

Define 29 tables in four tiers — provenance (dump_info, dump_stat, dump_conflict), reference (dcc, project, five ontology vocabularies), C2M2 entities and junctions under their upstream names, and per-consortium extensions — plus three views: v_file (term names and extension columns inlined, one row per file), v_file_biosample (the file × collection × biosample × subject fan-out grain, named so nobody sums size_in_bytes over it unawares), and v_subject_demographics (sex/race/age coalesced across the C2M2 columns and HuBMAP enrichment, each value paired with a *_source column so provenance survives the coalesce). Gate age_unit to the HuBMAP branch of the age coalesce — the unit belongs to age_value alone, and a C2M2-sourced age is years by definition — carried as schema_version 2. Declare every foreign key in the DDL, since DuckDB has no ALTER TABLE … ADD CONSTRAINT.

Loader (src/cfdb/dump/loader.py)

Stream documents through a two-tier accumulator: low-cardinality entities live in Python registries for column-level divergence comparison, while files, their 1:1 extras, and every junction stream through constraint-free TEMP staging tables, bounding memory by entity cardinality rather than file count. Finalize in foreign-key order — dcc → project → ontology → entities → file → junctions and extensions — with junctions deduplicated by SELECT DISTINCT. Coerce the known string-typed numerics from the C2M2 TSV path and ISO timestamps, recording each failure as a named counter rather than failing the document; skip documents that fail model validation, likewise counted. Route every CV-term foreign key through one guard that skips and counts a term lacking an id or a name, nulling the referencing column in the same decision. Assert the dcc id↔submission mapping 1:1 in both directions, run the two-submission assertion ahead of the stub short-circuits, and translate any residual finalize-time ConstraintException into a DumpError naming the table, so every keying-assumption violation is one legible diagnostic. Register stub project rows for foreign keys whose joined object never materialized, upgradeable by a later document carrying the real row, with both project_stubs_created and project_stubs_surviving reported. Anchor the dump_info commit probe to the package directory with a trackedness check, so only a genuine cfdb checkout ever yields a sha and an analyst's own repository never does. Serialize HuBMAP metadata with sorted keys so key order never registers as divergence, parse SCHEMA_VERSION from the DDL header so the constant cannot drift from the schema, and fail that parse with a message naming the required header line.

CLI (src/cfdb/cli.py)

Add a dump group with the duckdb command: a --submission restriction validated against the DCC registry via click.Choice, --to-parquet export via EXPORT DATABASE, a --force overwrite guard, a lazy duckdb import whose failure message names the installable cfdb[duckdb] @ git+… form, and Mongo connection kwargs reused from cfdb.db.mongo_client_kwargs. Remove the partial output on every abnormal exit path — cleanup runs in a finally, so KeyboardInterrupt is covered — while a completed dump is never destroyed: the Parquet export runs outside the destroy-on-failure scope, and an export failure keeps the .duckdb, reports the error, and exits nonzero naming what survived. Fix the shared --debug callback to honor its port argument; this deliberately changes cfdb sync -d as well, which previously ignored the value and listened on 5678.

Packaging (pyproject.toml)

Gate the sole new dependency behind a duckdb extra so the API image is unaffected, carry it in the dev extra for the test suite, and ship the schema DDL via package-data, which src-layout auto-discovery would otherwise omit from the wheel.

Documentation (README.md)

Document the command, the schema tiers, the per-consortium common-column mapping tables, and the loader's dedup, assertion, and coercion policies under a new "DuckDB dump" section — including that counters measure occurrences rather than entities, and that a database without a dump_info row is a partial dump.

Tests (tests/test_dump.py, tests/test_cli.py)

Mirror the module layout: loader, schema, and view tests live in tests/test_dump.py, while the CLI command tests and their Mongo-patching helper live in tests/test_cli.py, matching cfdb/cli.py.

Test cases

#	Test Suite	Given	When	Then	Coverage Target
1	TestDumpReport	A report with rows, conflicts, and non-zero counters	summary() is called	Renders row lines plus the divergence and counter sections	Populated summary rendering
2	TestDumpReport	A report with no conflicts and zero counters	summary() is called	Omits both section headers	Empty-section suppression
3	dump_files	No documents	A dump runs on a fresh database	Every schema table and view exists	Schema bootstrap
4	dump_files	One fully nested 4DN document	The document is dumped	One row per entity and per C2M2 junction	Flattening and junction recovery
5	dump_files	Two documents embedding identical copies of one collection	Both are dumped	One collection row, two membership rows, no conflicts	Entity dedup across documents
6	dump_files	Two copies of one collection with divergent lab values	Both are dumped	First value kept and the exact divergence row lands in dump_conflict	First-write-wins with divergence sampling
7	dump_files	Copies of one collection carrying different biosample lists	Both are dumped	The union of membership rows survives	Junction union semantics
8	dump_files	One collection key appearing under two submissions	Both are dumped	DumpError naming both submissions	Key-uniqueness assertion
9	dump_files	Two documents sharing one file key	Both are dumped	DumpError on duplicate file keys	Duplicate-file-key assertion
10	dump_files	A completed dump	A junction row naming an absent collection is inserted	ConstraintException is raised	Foreign-key enforcement
11	dump_files	A project stubbed by one submission, materialized by another	Both are dumped	DumpError naming both submissions	Key-uniqueness on the stub-upgrade path
12	dump_files	A real project row, then a cross-submission stub for the same key	Both are dumped	DumpError naming both submissions	Key-uniqueness on the incoming-stub path
13	dump_files	String-typed size_in_bytes and sidecar file_size	The document is dumped	Both stored as integers with no failures counted	TSV-typed numeric coercion
14	dump_files	A non-numeric size_in_bytes	The document is dumped	NULL stored, document kept, failure counted	Coercion-failure accounting
15	dump_files	A file_format term carrying an empty id	The document is dumped	FK NULL, no term row, term_id_missing counted	Term-id guard
16	dump_files	A file_format term with an id and a blank name	The document is dumped	Term skipped, FK NULL, term_name_missing counted	Nameless-term guard
17	dump_files	One document per consortium	All three are dumped	Each promoted file and collection column filled from its consortium's source field	Promoted-column mapping
18	dump_files	The dumped schema	Extension columns are inspected	Verbatim-promoted fields absent, replicate_info present	Extension-table residue rule
19	dump_files	One valid document and one with no dcc	Both are dumped	Valid document loaded, invalid skipped and counted	Invalid-document policy
20	dump_files	An explicit submissions filter	The document is dumped	dump_info carries schema version, filter, and file count	Provenance row
21	dump_files	The process CWD inside an unrelated git repository	The document is dumped	dump_info.cfdb_commit is not that repository's HEAD	Commit-provenance anchor
22	dump_files	A HuBMAP subject with enrichment-only demographics	v_subject_demographics is queried	Enrichment values surfaced with source columns naming the enrichment table	Demographics coalesce with provenance
23	dump_files	A subject with a C2M2 age and divergent HuBMAP age and unit	v_subject_demographics is queried	The C2M2 age paired with a NULL age_unit	Age-unit source consistency
24	dump_files	Generated batches with arbitrary file/collection sharing	Each batch is dumped	Row counts equal distinct keys and memberships	Cardinality invariants (property-based)
25	dump_files	Non-ISO creation_time and date_created strings	The document is dumped	NULL stored in both and each failure counted	Timestamp coercion-failure policy
26	dump_files	A creation_time with a +05:00 offset	The document is dumped	The UTC-equivalent naive timestamp stored	Timezone normalization
27	dump_files	One term id with divergent descriptions across two consortia	Both are dumped	One globally deduped term row, divergence in dump_conflict, no error	Ontology global dedup
28	dump_files	A biosample project FK with no joined project object	The document is dumped	All-NULL stub project row written, counted, FK intact	Project stub policy
29	dump_files	A stub-creating document, then one carrying the real project	Both are dumped in order	Stub upgraded to the real row with no conflict	Stub-upgrade path
30	dump_files	Three divergent collection copies with conflict_sample_cap=1	The documents are dumped	One sample row while the tally counts both divergences	Conflict sample cap
31	dump_files	One document per consortium and no submissions argument	The documents are dumped	dump_info carries the sorted seen submissions	Submissions fallback
32	dump_files	Extra submodels carrying only None fields	The document is dumped	No extension rows written	All-None extension suppression
33	dump_files	A HuBMAP metadata dict and vitessce_hints list	The document is dumped	Dict round-trips through the JSON column, list through VARCHAR[]	JSON and list column fidelity
34	dump_files	A 4DN document with two sidecars	The document is dumped	One row per sidecar with ordinals matching source order	Sidecar ordinal ordering
35	dump_files	Three documents dumped with batch_size=1	The documents are dumped	Row counts identical to the default batch size	Staging flush path
36	dump_files	An empty-string compression_format	The document is dumped	The empty string stored, not NULL	Empty-string compression semantics
37	dump_files	One ENCODE and one HuBMAP document	v_file is queried	Each row inlines its own extension columns and term names, the other consortium's columns NULL	v_file inlining
38	dump_files	One fully nested document	v_file_biosample is queried	The fan-out grain with anatomy and taxonomy names resolved	v_file_biosample grain
39	dump_files	Arbitrary surrogate-free text in the pass-through file columns	The document is dumped	Every value round-trips byte-exact	Text fidelity (property-based)
40	dump_duckdb (CLI)	An environment without the duckdb module	The command is invoked	Nonzero exit naming the cfdb[duckdb] extra	Missing-extra diagnostic
41	dump_duckdb (CLI)	A mocked Mongo yielding one document	The command is invoked with -s 4DN	The filter reaches Mongo lower-cased and a valid database is written	CLI wiring
42	dump_duckdb (CLI)	An existing output file	The command is invoked without --force	Nonzero exit with the file untouched	Overwrite guard
43	dump_duckdb (CLI)	An existing output file	The command is invoked with --force	The file replaced by a valid database	Forced overwrite
44	dump_duckdb (CLI)	A --to-parquet directory	The command is invoked	Parquet files written and the export named in the output	Parquet export
45	dump_duckdb (CLI)	A --to-parquet target inside an unwritable directory	The command is invoked	Nonzero exit naming the export failure with the .duckdb kept on disk	Export-failure isolation
46	dump_duckdb (CLI)	Two documents with duplicate file keys	The command is invoked	Nonzero exit, DumpError message printed, partial output removed	Failure cleanup