pinax

Note: This library has been written extensively with AI assistance (Claude Code). Users who are not comfortable with AI-generated code should take that into account before adopting it.

pinax is a Python library for building, managing, and querying statistical metadata catalogs. Named after "bibliographic work composed by Callimachus (310/305–240 BCE) that is popularly considered to be the first library catalog in the West"Pinax

Inspired by DCAT and the broader landscape of metadata standards (SDMX, DDI), pinax provides a general-purpose engine for catalog storage, discovery, and retrieval — designed to be embedded in ETL pipelines, data platforms, and analytical tooling rather than used as a standalone application. It pairs naturally with domain-specific libraries like sdmxlib for standards-aware workflows.

import pinax as pk

What it does

pinax materializes remote catalog and structural metadata into a local DuckDB database and exposes a fluent Python API for discovery queries — filtering by publisher, theme, dimension, code, free text, or provenance lineage.

The catalog is a materialized graph. Rather than federated queries across separate REST endpoints, ingest pulls the full structural model into one database. SQL JOINs are the graph traversal. No network round-trips at query time.

Installation

Requires Python 3.13+. Managed with uv.

uv add pinaxlib
# or: pip install pinaxlib

The distribution is published on PyPI as pinaxlib; the import name remains pinax.

Quick start

import pinax as pk
import pinax.query as q
import sdmxlib as sl

# Creates my_catalog/bundle.duckdb and my_catalog/parquet/
with pk.Catalog("my_catalog").open() as store:
    with sl.RestRegistry(sl.Provider.ESTAT) as reg:
        reg.get(sl.Dataflow.urn_for(agency="ESTAT", id="NAMA_10_GDP"), resolve=True)
        pk.ingest_sdmx(store, reg.registry)

    # Catalog query — no network needed
    results = (
        store.query(pk.AggregateDataset)
        .filter(q.has_code("geo", "DE"))
        .search("GDP")
        .with_facets("themes", "frequency")
        .execute()
    )
    print(results.total, "datasets found")
    print(results.facets["themes"])

Five dataset kinds

pinax uses a discriminated union of dataset types aligned with DCAT-AP and StatDCAT-AP:

# Generic datasets (CKAN, open portals)
pk.OpenDataset(identifier="co2-2024", title=i18n("CO2 Emissions 2024"), ...)

# Statistical tables with SDMX structure (dimensions, codelists)
pk.AggregateDataset(identifier="ESTAT:UNE_RT_M(1.0)", ..., sdmx_dataflow_urn="urn:...")

# Survey microdata with variable-level metadata
pk.MicrodataDataset(identifier="lfs-2023", ..., variables=[...])

# Spatial datasets with bounding box and CRS
pk.GeospatialDataset(identifier="boundaries-2024", ..., crs="EPSG:4326")

# Articles, reports, and analytical publications
pk.PublicationDataset(identifier="pub-71-607-x", ..., doi="10.25318/...", authors=[...])

All five share the same store and query API. Use pk.BaseDataset as the bound for generic code; pk.Dataset is a type alias for the full union.

Source connectors

SDMX (Eurostat, OECD, BIS, ABS, ...)

import sdmxlib as sl
import pinax as pk

with pk.Catalog("my_catalog").open() as store:
    with sl.RestRegistry(sl.Provider.ESTAT) as reg:
        for df_id in ["NAMA_10_GDP", "UNE_RT_M", "PRIC_HPI_IDX"]:
            reg.get(sl.Dataflow.urn_for(agency="ESTAT", id=df_id), resolve=True)
        pk.ingest_sdmx(store, reg.registry)

        # Stream observation data to Parquet, attach as Distribution
        pk.ingest_data(store, reg, "ESTAT:NAMA_10_GDP(latest)", measure_dim="na_item")

Statistics Canada

from pinax.sources.statcan import WDSClient, NDMClient

with pk.Catalog("statcan").open() as store:
    with WDSClient() as wds:
        pk.ingest_statcan_table(store, wds, "14100287")   # Labour Force Survey

    with NDMClient() as ndm:
        pk.ingest_statcan_publications(store, ndm, product_type="82", limit=200)

CKAN (open.canada.ca, data.gov, ...)

from pinax.sources.ckan import CkanClient

with pk.Catalog("open_canada").open() as store:
    with CkanClient("https://open.canada.ca/data") as client:
        pk.ingest_ckan(store, client, organization="statcan", rows=500)

Scope-based graph traversal

pinax exposes a lazy, scope-based API for navigating the catalog graph. Navigation builds scope objects without executing SQL; only terminal methods (.collect(), .count()) hit the database.

import pinax as pk

store = pk.Catalog("my_catalog").open()

# Navigate themes — no SQL until .collect()
concepts = store.themes["statcan"].collect()          # ItemList[Concept]
concept = store.themes["statcan"]["13"].collect()     # Concept

# Cross-entity navigation — .datasets returns a lazy QueryBuilder
datasets = store.themes["statcan"]["13"].datasets.collect()

# Enrich with sub-traversal expressions (like Polars' pl.col())
store.themes["statcan"].enrich(
    n=pk.each("datasets").count(),
    has_data=pk.each("datasets").exists(),
).collect()

# Codelist navigation — pk.urn builds URN strings for you
codes = store.codelist(pk.urn.codelist("SDMX", "CL_GEO")).collect()    # ItemList[Code]
code = store.codelist(pk.urn.codelist("SDMX", "CL_GEO"))["ON"].collect()  # Code

# CodelistsScope — parallel to ThemesScope, supports enrich
# Enriched output includes labels resolved via sdmx.localized_text
store.codelists.lang("en").enrich(n=pk.each("datasets").count()).collect()
# → [{"urn": "...", "label": "Geography", "n": 5}, ...]

# Filter codelists by label text (SQL-level, case-insensitive)
store.codelists.filter(text_contains="geo").enrich(n=pk.each("datasets").count()).collect()

store.codelist(urn).label("en")   # quick name lookup

# Enriched per-code output includes code name labels
store.codelist(urn).lang("en").enrich(n=pk.each("datasets").count()).collect()
# → [{"code_id": "ON", "label": "Ontario", "n": 5}, ...]

# Batch label resolution — single SQL query for many codes
store.codelist(geo_urn).batch_labels(["CA", "US", "DE"])
# → {"CA": "Canada", "US": "United States", "DE": "Germany"}

# Across multiple codelists at once
store.codelists.batch_labels([(geo_urn, "CA"), (freq_urn, "A")], lang="en")
# → {(geo_urn, "CA"): "Canada", (freq_urn, "A"): "Annual"}

# Dimension traversal
dims = store.dimensions(ds).collect()              # ItemList[DimensionInfo]
codelist = store.dimensions(ds)["GEO"].codelist    # CodelistScope (lazy)

Scope classes: ConceptSchemesScope, ConceptSchemeScope, ThemesScope, SchemeScope, ConceptScope, CodelistsScope, CodelistScope, CodeScope, DimensionsScope, DimensionScope.

Expression types: pk.each("edge") creates a context-free sub-traversal expression. Reusable across .enrich(), .filter(), and .sort_by().

Query API

Structured filters

import pinax.query as q

# Field filters (keyword arguments)
store.query(pk.AggregateDataset).filter(publisher="ESTAT", status="current").all()

# Composable filter objects
store.query(pk.AggregateDataset).filter(
    q.has_code("geo", "DE"),          # datasets with GEO=DE in their codelist
    q.has_dimensions(["geo", "freq"]), # datasets with both GEO and FREQ dimensions
).all()

# Hierarchy descent — any geography *under* Canada, strict (CA itself excluded).
# The applicable Hierarchy is resolved per dataflow; add include_self=True to
# match the root as well.
store.query(pk.AggregateDataset).filter(q.within("REF_AREA", "CA")).all()

# Distribution and service filters
store.query(pk.OpenDataset).filter(
    q.distribution(format="CSV"),
    q.has_service(endpoint_url="https://..."),
).all()

Cross-entity queries

# Agents that publish datasets with code GEO=CA
store.query(pk.Agent).filter(
    q.publishes(q.has_code("geo", "CA"), kind="aggregate")
).all()

# Data services serving aggregate datasets
store.query(pk.DataService).filter(
    q.serves(kind="aggregate")
).all()

MAP column filters

# Spatial coverage filter
store.query(pk.BaseDataset).filter(q.has_spatial("Canada")).all()

# Keyword filter
store.query(pk.BaseDataset).filter(q.has_keyword("employment")).all()

# Title/description search (case-insensitive ILIKE)
store.query(pk.BaseDataset).filter(q.title_contains("GDP")).all()
store.query(pk.BaseDataset).filter(q.description_contains("quarterly")).all()

# Sort by multilingual title
store.query(pk.BaseDataset).sort_by("title", lang="en").all()
store.query(pk.BaseDataset).sort_by("title", lang="en", desc=True).all()

Selective relationship loading

By default, querying a list of datasets loads all relationships (~17 queries). Use .include() to declare exactly which relationships to batch-load — the rest are set to an UNLOADED sentinel:

# Only load publisher and themes — 3 queries instead of ~17
results = (
    store.query(pk.BaseDataset)
    .filter(status="published")
    .include("publisher", "themes", "keywords")
    .sort_by("issued", desc=True)
    .limit(20)
    .all()
)

# Explicit full hydration — useful to make the cost visible at the call site
results = store.query(pk.BaseDataset).filter(...).full().all()

# get() always loads everything — no include() needed
ds = store.get(pk.AggregateDataset, "ESTAT:NAMA_10_GDP(1.0)")

Unloaded fields raise pk.NotLoadedError on access. Use pk.is_unloaded(value) to check before accessing:

ds = store.query(pk.BaseDataset).include("publisher").first()
ds.publisher.name    # OK
ds.themes[0]         # raises NotLoadedError: 'themes' was not loaded

if not pk.is_unloaded(ds.themes):
    print(ds.themes)

Valid relationship names: publisher, contact_point, frequency, licence, themes, subject, dataset_type, keywords, spatial_coverage, distributions, conforms_to, quality_annotations, provenance, dimension_names, variables, feature_types, authors.

Lightweight projections

When you need just a few columns (e.g. autocomplete), projection modifiers bypass full object reconstruction — a single SQL query:

# Row projection — .select() + .rows() returns Row objects (dict subclass)
rows = (
    store.query(pk.BaseDataset)
    .filter(q.title_contains("GDP"))
    .sort_by("title", lang="en")
    .limit(3)
    .select("identifier", "title", lang="en")
    .rows()
)
# → [Row({"identifier": "GDP", "title": "GDP Growth"}), ...]
rows[0].identifier   # attribute-style access
rows[0]["title"]     # dict-style access — both work

# Flat value projection — .scalars() + .values() returns bare values
ids = store.query(pk.BaseDataset).filter(status="current").scalars("identifier").values()
# → ["EXR", "M1", "UNEMP", ...]

# Existence check — no object reconstruction
if store.query(pk.AggregateDataset).filter(q.has_code("geo", "DE")).exists():
    ...

Full-text search and facets

# BM25 search across titles, descriptions, keywords, themes, and dimensions
results = store.search("unemployment", limit=20)

# Combined search and filter
results = (
    store.query(pk.AggregateDataset)
    .filter(publisher="ESTAT")
    .search("labour force")
    .with_facets("themes", "frequency")
    .execute()
)
print(results.facets["themes"])   # {"Labour": 18, "Economy": 6, ...}

# Aggregation counts
counts = store.facets("publisher", "themes", "frequency", "status")

Lineage and provenance (PROV-O)

# Record derivation relationships between datasets
store.add_lineage(
    "lfs-microdata",
    "14100287",
    "aggregated_from",
    activity_type="aggregation",
    activity_label="LFS monthly tabulation",
    confidence="asserted",
)

# Transitive upstream/downstream traversal — returns QueryBuilder for chaining
ancestors = store.dataset("14100287").upstream(depth=5).collect()
dependents = store.dataset("CL_GEO").downstream(relationship="uses_classification").collect()

# Chain additional filters after traversal
current = store.dataset("14100287").upstream().filter(status="current").collect()

# Inspect lineage records
rows = store.dataset("14100287").lineage_records(role="target", relationship="aggregated_from")

Architecture

Three layers share one DuckDB database:

Discovery layer  — Catalog            (dataset, agent, concept_scheme, concept, distribution, lineage)
Structural layer — sdmxlib tables     (dataflows, dsd_components, codes, codelists)
Observation layer — Polars / Parquet  (actual time-series data)

pinax owns the discovery layer. sdmxlib owns the structural layer. Both write to the same DuckDB connection — queries JOIN freely across both. Parquet files live alongside the database and are referenced via DCAT Distribution records.

Development

Run once per clone — installs a pre-push hook that runs the full gate, since GitLab CI only builds the docs site:

just hooks-install

just ci                # full gate: lockcheck + lint + format + typecheck + unit tests
just test              # unit tests
just test-integration  # integration tests (no network)
just test-live         # live tests against real SDMX endpoints
just lint              # ruff check
just typecheck         # basedpyright
just lockcheck         # uv.lock matches pyproject and resolves only from public PyPI
just docs              # local docs server

uv.lock is committed, and every dependency must resolve from public PyPI — just lockcheck fails the build if an internal mirror leaks into it.