pyacid 0.3.0a1

Astronomical Catalog Inference Driver: XMATCH SQL over HATS-partitioned Parquet via native Polars

ACID — Astronomical Catalog Inference Driver

Copyright (c) 2026, Mario Juric. BSD 3-Clause License.

Cross-match and query HEALPix-partitioned astronomical catalogs from Python. ACID gives you a fluent Catalog surface (crossmatch, where, select, group_by / aggregate, save) for the common shapes, and a SQL escape hatch (db.sql(...)) with one astronomy extension — XMATCH(radius_arcsec => ...) — for everything else. Each anchor partition runs independently against a boundary-safe margin cache.

Reads and writes the HATS format used by LINCC Frameworks (LSDB, hats-import) and by published catalogs such as Gaia DR3 and Rubin DP1.

---

Quick start

Run a query

acid is module-level and singleton-by-default (like Ray / DuckDB / Polars): acid.open(...) returns a catalog handle, acid.sql(...) is the escape hatch, and the worker pool is built once and reused. acid.init(...) is optional — call it to pin the source / worker count; otherwise the first acid.open() lazy-inits with defaults.

import acid
import astropy.units as u

acid.init("catalogs.yaml", workers=8)       # optional — pins config

gaia = acid.open("gaia_dr3")
twomass = acid.open("twomass_psc")

# Fluent: composable verbs, lazy until materialized.
matches = (gaia
           .crossmatch(twomass, radius=1*u.arcsec)
           .where("phot_g_mean_mag < 16")
           .select("source_id, designation"))

matches.head(10).show()         # pretty-print to stdout
df = matches.to_polars()        # Catalog converters: .to_polars(),
                                #   .to_astropy(), .to_arrow(), .to_pandas()

# SQL escape hatch for aggregates / HAVING / windows / DISTINCT.
r = acid.sql("""
    SELECT g.source_id, t.designation, d
    FROM   gaia_dr3 AS g
    JOIN   twomass_psc AS t ON XMATCH(radius_arcsec => 1.0, dist_col => 'd')
    ORDER BY d
    LIMIT  20
""")
print(r)

Need two simultaneous connections, or full isolation (e.g. in a library or a test)? Construct acid.Connection(...) explicitly and use it as a context manager — it bypasses the module-level default entirely:

with acid.Connection("catalogs.yaml", workers=8) as db:
    df = db.open("gaia_dr3").head(10).to_polars()

Result is a thin wrapper around an Arrow table that comes back from every materialization call. .show() prints; .to_pandas() / .to_polars() / .arrow() / .to_pylist() convert (r.arrow() returns a pyarrow.Table); and .write("results.parquet") / .write_parquet() / .write_csv() / .write_fits() save to disk. For an astropy.table.Table, call Catalog.to_astropy() directly on the query handle — to_astropy() is the one converter that lives on Catalog only, not on Result.

Restrict to a region while you iterate

acid.in_cone(...) (or db.in_cone(...)) is a context manager that scopes a spatial cone to every query executed inside the with block — both the fluent surface and db.sql(...). The cone is applied at execution time, not when the query was built, so you can build a query once and run it scoped inside the block and full-sky outside it. Use it for a "debug small, run big" workflow: keep the block in while you iterate, remove it for the production run.

gaia = acid.open("gaia_dr3")

with acid.in_cone((180, 0), radius=1*u.deg):
    small = gaia.where("phot_g_mean_mag < 16").to_polars()

# Same query, full sky, no edits:
big = gaia.where("phot_g_mean_mag < 16").to_polars()

Cones do not nest; one in_cone block at a time.

Materialize an intermediate

Catalog.save(...) writes a query result as a HATS catalog and registers it on the connection so later queries can reference it by name. This is the canonical EDA pattern: run a heavy crossmatch once, save it, iterate cheaply on the cached output.

nearby = (acid.open("gaia_dr3")
              .crossmatch(acid.open("twomass_psc"), radius=1*u.arcsec)
              .save("./out/gaia_x_2mass", name="nearby"))

# `nearby` is a normal Catalog; "nearby" is also resolvable by name.
r = acid.sql("SELECT COUNT(*) FROM nearby")
print(r)

CLI

# Query execution. The SQL query is required (use '-' to read stdin, or -f).
# --db is a ':'-separated list of HATS dirs / registry YAMLs; it's optional,
# falling back to $ACID_PATH, the acid.conf 'path' setting, then ~/datasets.
acid query "SELECT COUNT(*) FROM object" --db datasets/ --out /tmp/result
acid query -f query.sql --db catalogs.yaml --out results/ --workers 32
echo "SELECT ..." | acid query - --db datasets/ --out results/   # '-' reads stdin
# --format is optional: it's inferred from the --out extension
# (.parquet/.pq, .csv, .fits/.fit, .hats); no extension → HATS tree.
acid query "SELECT ..." --db datasets/ --out results.csv
# --open uses a raw file (parquet/csv/fits/arrow/…) as a table, alongside the
# --db catalogs. The ra/dec column names are required. Two forms: positional
# 'PATH,RA,DEC' (table name = file basename) or named 'NAME=PATH,ra=RA,dec=DEC'.
acid query "SELECT t.id, g.source_id FROM t JOIN gaia ON XMATCH(radius_arcsec => 1.0)" \
    --db datasets/ --open t=candidates.csv,ra=RA,dec=DEC
acid query "SELECT * FROM candidates" --db datasets/ --open candidates.csv,RA,DEC
acid validate "SELECT ..." --db datasets/

# Download catalogs (HTTP, SSH, or local; full or spatial subset)
acid download two_mass                                # resolve name + dest (see below)
acid download https://data.lsdb.io/hats/two_mass/two_mass /data/two_mass
acid download https://data.lsdb.io/hats/two_mass/two_mass /data/two_mass --cone 50,-50,10
acid download user@server:/hats/gaia /data/gaia --columns ra,dec,mag --cone 180,0,5

# Inspect catalogs (local or remote)
acid inspect two_mass                                # bare name → resolved on ACID_PATH
acid inspect /data/two_mass                          # summary
acid inspect schema /data/two_mass                   # column schema
acid inspect https://data.lsdb.io/hats/two_mass/two_mass  # remote

# Build margin caches locally (--margin-arcsec defaults to 10.0)
acid hats build-margin /data/two_mass --margin-arcsec 10.0 --workers 16

results/ is itself a valid HATS catalog (lsdb.open_catalog(...) and hats.read_hats(...) will read it). Downloaded subsets are also valid HATS catalogs with rebuilt _metadata.

acid download name resolution. Give a bare catalog name and both the source and destination are resolved for you:

acid download two_mass
# source ← first ACID_DOWNLOAD_PATH root that has it (collection-aware),
#          e.g. https://data.lsdb.io/hats/two_mass/two_mass
# dest   ← <first writable local ACID_PATH root>/two_mass (created if needed)

The source search path is ACID_DOWNLOAD_PATH → the download_path config setting → the built-in https://data.lsdb.io/hats/. Each candidate <root>/<name> is probed; a directory holding collection.properties is a collection, so its hats_primary_table_url child is downloaded. An explicit source (anything with a /, or a URL) skips this and is used verbatim.

The destination follows the same bare-vs-path rule: omit it and the catalog lands in <first writable local ACID_PATH root>/<catalog name>; pass a bare name (acid download two_mass tm) and it resolves to <ACID_PATH root>/tm; pass a path with a / (./tm, /data/tm) and it's used verbatim. The ACID_PATH root is the same search path acid query uses (URL entries are skipped), created with a notice if it doesn't exist. An explicit source with an omitted destination is an error (there's no name to resolve a destination from).

Catalog registry

The simplest way: point --db (or acid.init(...)) at a directory of HATS catalogs. Each subdirectory with a properties file becomes a table named after the directory. Margin caches (dataproduct_type=margin) are auto-skipped.

For more control, use a YAML file:

catalogs:
  dia_source:
    path: /data/dia_source      # HATS root, or CatalogCollection root
    # Auto-detected from <path>/properties when present:
    #   ra_col            (from hats_col_ra)
    #   dec_col           (from hats_col_dec)
    #   hpix_order        (from <path>/partition_info.csv)
    #   neighbor_path     (from collection.properties or sibling '_margin' dir)
    #   neighbor_margin_arcsec  (from hats_margin_threshold)
    #   npix_suffix       (from hats_npix_suffix; default '.parquet')
    # Any auto-detected value can be overridden here.

  object:
    path: /data/object_collection    # a CatalogCollection root works too

  lightcurve:
    path: /data/lightcurve
    hpix_order: 5                    # explicit when partition_info.csv is absent

# Named MOC footprints for IN_MOC() filtering.
# Each entry is a path to a FITS file (HEALPix image or MOC FITS).
mocs:
  des_dr2: /data/mocs/des_dr2.fits
  known_artifacts: /data/mocs/artifacts.fits
  # If a catalog has a point_map.fits at its root, IN_MOC(<alias>, '<catalog_name>')
  # auto-loads it — no explicit entry needed.

Configuration (acid.conf)

So you don't re-type --db/--workers on every invocation, acid reads an INI config. The first existing file wins, searched highest-priority first: --config FILE / $ACID_CONFIG, then ~/.config/acid/acid.conf ($XDG_CONFIG_HOME), /sdf/data/rubin/user/mjuric/etc/acid.conf, /sdf/home/m/mjuric/etc/acid.conf, $XDG_CONFIG_DIRS, /etc/acid/acid.conf.

# ~/.config/acid/acid.conf
[acid]
path = /data/hats:~/datasets        # ':'-separated HATS dirs / registry YAMLs
download_path = https://data.lsdb.io/hats/   # 'acid download' name search path
workers = 32                        # query worker pool ("auto" = cgroup-aware)
mem_per_worker_gb = 4               # RAM/worker bounding "auto" (CPU and memory)
tmpdir = /scratch/$USER             # base temp dir (a per-run subdir is made + cleaned)
inmem_row_limit = 50_000_000        # spill threshold

Each setting resolves explicit flag/arg → env var → config → built-in. Env overrides: ACID_PATH, ACID_DOWNLOAD_PATH, ACID_WORKERS, ACID_MEM_PER_WORKER_GB, ACID_TMPDIR, ACID_INMEM_ROW_LIMIT. Inspect and edit with acid config:

acid config show                 # values set in the file (--effective: resolved + provenance)
acid config get workers          # file value; exits 1 (prints nothing) if unset
acid config set path /data/hats:~/datasets

With a config in place, --db becomes optional (falls back to the path setting, then ~/datasets). See docs/archive/CONFIG-SYSTEM.md for the full design.

---

What XMATCH does

JOIN  b ON XMATCH(radius_arcsec => 1.0)                   -- nearest, inner
JOIN  b ON XMATCH(r => 1.0)                               -- 'r' is an alias
JOIN  b ON XMATCH(r => 1.0, mode => 'all')                -- every match within r
LEFT JOIN b ON XMATCH(r => 1.0)                           -- keep unmatched anchors

-- Distance is exposed as a named column via dist_col on the XMATCH call.
SELECT a.id, d FROM a JOIN b ON XMATCH(r => 1.0, dist_col => 'd')
WHERE  d < 0.5

-- Ordinary joins, WHERE, GROUP BY, HAVING, ORDER BY, LIMIT/OFFSET,
-- DISTINCT all work; cross-partition reduction is handled internally.
SELECT a.id, COUNT(*) AS n, AVG(d) AS avg_d
FROM a
JOIN  b ON XMATCH(r => 1.0, dist_col => 'd')
JOIN  lightcurve AS lc ON a.id = lc.object_id
GROUP BY a.id
ORDER BY n DESC LIMIT 100

-- Footprint filtering via MOC (Multi-Order Coverage maps):
-- Restrict rows to a survey footprint or sky region.
SELECT a.id, a.ra, a.dec
FROM a JOIN b ON XMATCH(r => 1.0)
WHERE IN_MOC(a, 'des_dr2')              -- anchor inside DES footprint
  AND NOT IN_MOC(b, 'known_artifacts')  -- exclude artifact regions

-- IN_MOC also works in SELECT projections (per-row boolean):
SELECT a.id, IN_MOC(a, 'des_dr2') AS in_des FROM a

The fluent equivalent of the simple shapes:

a.crossmatch(b, radius=1*u.arcsec)                          # nearest, inner
a.crossmatch(b, radius=1*u.arcsec, how="all")               # every match within r
a.crossmatch(b, radius=1*u.arcsec, how="left")              # LEFT XMATCH
a.in_region("des_dr2")                                      # IN_MOC mask, per-receiver

Semantics, in short:

• All XMATCHes in a query use the anchor (first FROM) table's coordinates, even after a mode => 'all' expansion.
• A right-table radius must be ≤ that catalog's neighbor_margin_arcsec. Otherwise we'd silently miss boundary pairs; the analyzer rejects the query.
• ORDER BY ... LIMIT K pushes the top-K to each partition first; the reducer re-sorts the union and applies the global LIMIT/OFFSET.
• Aggregates / GROUP BY / DISTINCT / HAVING run in a phase-2 reducer over the per-partition Parquet output.

---

Python API surface

# Module-level API (singleton-by-default). The first call lazy-inits;
# acid.init(...) pins config; acid.shutdown() tears down.
acid.init(source=None, *, workers="auto", threads=None,
          inmem_row_limit=50_000_000, progress via configure) -> None
acid.shutdown() / acid.is_initialized() / acid.configure(progress=...)
acid.open(name_or_path, *, alias=None, columns=None) -> Catalog
acid.sql(query, *, output=None)                      -> Result
acid.map_partitions_sql(query, *, output=None)       -> ExecutionResult
acid.add_catalog(name, **spec_kwargs) / acid.list_catalogs() / acid.register_moc(...)
acid.in_cone(center, *, radius) / acid.validate(query) / acid.explain(query) / acid.status()

# Explicit, isolated Connection (escape hatch — two connections / two configs)
db = acid.Connection(source, *, workers="auto", threads=None,
                     inmem_row_limit=50_000_000, progress="auto")
# ...then db.open(...) / db.sql(...) / etc. — the same methods, on `db`.
db.in_cone(center, *, radius)                       # ctx manager
db.status() / db.validate(q) / db.explain(q)
db.close()    # or use as a context manager

# Catalog (composable, lazy) — composition verbs return Catalog;
# materialization verbs execute and return Result (or, for to_*, the
# converted type).
cat.where(predicate)        -> Catalog
cat.select(*cols)           -> Catalog
cat.limit(n)                -> Catalog
cat.in_region(moc_or_cat)   -> Catalog
cat.crossmatch(other, *, radius, how="nearest"|"all"|"left",
               dist_col=None, suffix=None)                   -> Catalog
cat.join(other, *, on, how="inner"|"left")                   -> Catalog
# Fluent aggregation — decomposable-only (acid.agg constructors).
cat.group_by(*keys)                        -> Catalog
cat.aggregate(**named_aggs)                -> Catalog
# After .aggregate(), verbs compose over the aggregate output in written
# order: a post-aggregate .where(...) is the old HAVING (and .limit(5).where(...)
# filters the top-5 — fluent composes by position, no separate .having()).
cat.sort(*keys, descending=False, nulls_last=False) -> Catalog
# Reduction shortcuts — one aggregate, no agg.* ceremony. Global (no
# group_by) materializes and returns a bare Python scalar; grouped returns a
# lazy Catalog (column `count` / `mean_<col>` / …, so a following .where() is
# HAVING). For mixed stats / named outputs use .aggregate(...).
cat.count(col=None)                         -> int | Catalog
cat.sum/mean/min/max/std/var(col)           -> scalar | Catalog
# Decomposable aggregate constructors (acid.agg):
#   agg.count(col=None), agg.sum, agg.mean, agg.min, agg.max,
#   agg.std, agg.var, agg.bool_and, agg.bool_or, agg.list.
# (No agg.median / agg.mode — non-decomposable; rejected with
# ValidationError. Drop into Polars after .to_polars() if you need them.)

cat.columns / cat.alias / cat.describe() / cat.explain()
cat.head(n=10)              -> Result
cat.execute()               -> Result
# These convert and return the target type directly (no Result detour):
cat.to_pandas() / cat.to_astropy() / cat.to_polars() / cat.to_arrow()
cat.save(path, *, name=None, overwrite=False) -> Catalog

# Result — comes back from Catalog.head / .execute and from db.sql.
# A thin wrapper around an in-memory pa.Table or a partitioned dir;
# no Astropy / Arrow converters live here (those are on Catalog).
r.num_rows, r.column_names, r.schema
r.column(name)         -> pa.ChunkedArray
r.show(n=20)           # pretty-print to stdout
print(r)               # same formatter via __str__
r.arrow()              -> pa.Table
r.df() / r.to_pandas() -> pandas.DataFrame
r.to_polars()          -> polars.DataFrame
r.to_pylist()          -> list[dict]
r.batches(batch_size=None) -> Iterator[pa.RecordBatch]
r.head(n=10)           -> Result
r.write_parquet(path, layout="hats"|"single") -> Path
r.write_csv(path)      -> Path
r.write_fits(path)     -> Path
r.write(path, format=None) -> Path   # format inferred from extension when omitted
len(r), for batch in r: ...

# Errors (all inherit from acid.AcidError)
acid.RegistryError           # catalog config (missing path, mixed Norder, ...)
acid.ParseError              # SQL parse failures
acid.ValidationError         # unsupported XMATCH constructs
acid.ExecutionError          # per-partition execution failures
acid.ConnectionClosedError   # method called on a closed Connection

---

Layout assumptions

• Catalogs follow the HATS layout: <root>/dataset/Norder=N/Dir=D/Npix=P.parquet (or Npix=P/*.parquet when hats_npix_suffix='/').
• Margin caches live as sibling catalogs (HATS canonical), at <root>/margin_cache/..., or any sibling dir matching <name>_margin*. collection.properties is consumed if present.
• Adaptive (per-pixel) Norder is supported: a catalog's partition_info.csv may list pixels at any orders, and XMATCH/ordinary joins across mixed-Norder catalogs are run via a refinement-tree enumeration that emits one work unit per coarsest cursor pixel where every joined catalog has ≤ 1 partition. Output is itself a valid HATS catalog whose partition_info.csv reflects the refinement.

---

What's the speed story?

• Each partition is independent → embarrassingly parallel across HEALPix pixels.
• Top-K queries push the LIMIT to each partition. Aggregates write partial data to disk and reduce centrally.
• Column pruning: the anchor and right relations are lazy Polars LazyFrames over scan_parquet(), so the final projection only pulls referenced columns from disk. Wide catalogs (150+ columns) don't slow down narrow SELECTs.
• Auto-spill: when output is unset and the running result exceeds inmem_row_limit (default 50M rows), acid spills to a tempdir rather than OOM-ing the parent.
• Allocator tuning: acid ships a jemalloc default that avoids page-purge contention at high worker counts (~2× faster wall, ~20% more RSS). It's a single overridable env var — see MEMORY-TUNING.md if you're memory-constrained or scaling workers on a large machine.

See bench/match_all.py and bench/session_vs_oneshot.py for microbenchmarks.

---

Install

With uv (recommended for development)

uv sync --dev          # creates .venv, installs all deps + test + hats
uv run pytest          # run tests

With pip

pip install -e .
# extras:  pip install -e .[hats,dev]

Requires Python 3.10+. Runtime dependencies (installed automatically): Polars ≥ 1.41, SQLGlot ≥ 27 (< 31), PyArrow ≥ 14, NumPy ≥ 1.24, SciPy ≥ 1.10, cdshealpix ≥ 0.7, fsspec ≥ 2023.1, Astropy ≥ 5, PyYAML ≥ 6, rich ≥ 13. mocpy is not a runtime dependency — ACID ships a dependency-light MOC implementation.

---

Status

• v0 (correctness): XMATCH inner/left, mode 'nearest'/'all', chains, ordinary joins, distance via XMATCH(..., dist_col => '<name>').
• v1 (scale): views + narrow side-tables, vectorized matcher, worker initializer, auto-spill, top-K pushdown, manifest.
• v1.1 (HATS spec): writes valid HATS catalogs, reads canonical property keys, supports hats_npix_suffix='/', auto-discovers margin siblings via collection.properties.
• v2 (EDA): persistent Connection, per-worker Polars engine, Result wrapper, Catalog.save() for materialization.
• v3 (adaptive Norder): per-catalog PartitionIndex, refinement-tree tuple enumeration, integer _healpix_29 range filtering for per-pixel row pruning, LEFT-XMATCH/JOIN over partitions without coverage.
• v4 (Polars-native): single native-Polars engine; DuckDB, the SQL rewriter/reducer, the engine abstraction, and the QueryPlan IR removed (see CHANGELOG.md / ARCHITECTURE.md).
• v4 (MOC footprint filtering): IN_MOC(<alias>, '<name>') in WHERE restricts rows to a named sky region (Multi-Order Coverage map). Supports NOT IN_MOC, multiple predicates (AND-combined via mocpy set ops), and catalog auto-resolution from point_map.fits. IN_MOC is a footprint restriction only — it must sit in conjunctive WHERE position (top-level AND-chain, optionally negated); use in SELECT/ORDER BY/CASE/JOIN ON or inside a disjunction is rejected (see Known limitations). Three-level optimization: catalog-footprint scoping, cursor-pixel intersection, and partition-level pruning — all via the existing _healpix_29 row-group pushdown fast path.
• v5 (catalog ops): acid hats build-margin builds HATS margin caches locally (validated against hats-import). acid download generates point_map.fits, auto-includes HATS RA/Dec/healpix columns. acid query accepts --db <directory> for zero-config usage, fails fast on errors, shows tqdm progress, shuffles work for load balancing. Bare column resolution via schema introspection. LocalFetcher for local I/O.
• v6 (fluent Catalog API): acid.init() builds a process-wide default (or acid.Connection() an explicit) Connection; db.open(name) returns a lazy Catalog; verbs (where, select, crossmatch, join, in_region, save) compose without writing SQL. db.in_cone(...) scopes a cone to every query in a with block. db.sql(...) remains the escape hatch for decomposable aggregates, HAVING, and top-K (ORDER BY ... LIMIT). Window functions, DISTINCT, COUNT(DISTINCT), bare GROUP BY, and unbounded ORDER BY are rejected with a ValidationError.

Tests: ~545 passing (~60s parallel via pytest-xdist) on the native Polars engine. Fixtures cached across runs.

Known limitations

• XMATCH must be the entire ON predicate. Compound predicates like XMATCH(...) AND b.mag < 20 are rejected.
• No CTEs / subqueries in the anchor position.
• RIGHT / FULL / CROSS JOIN XMATCH not supported.
• IN_MOC must be in conjunctive WHERE position (top-level AND-chain, optionally negated). Disjunctive use (IN_MOC(...) OR ...) and IN_MOC in JOIN ON are rejected.
• No nested db.in_cone(...) blocks. The true intersection of two non-concentric cones is not a cone; we refuse rather than silently approximate.