metadata-crawler 2511.1.2

Crawl, extract and push climate metadata for indexing.

metadata-crawler

Harvest, normalise, and index climate / earth-system metadata from POSIX, S3/MinIO, and OpenStack Swift using configurable DRS dialects (CMIP6, CMIP5, CORDEX, …). Output to a temporary catalogue (JSONLines) and then index into systems such as Solr or MongoDB. Configuration is TOML with inheritance, templating, and computed rules.

TL;DR

• Define datasets + dialects in drs_config.toml
• mdc add → write a temporary catalogue (jsonl.gz)
• mdc config → inspect a the (merged) crawler config.
• mdc walk-intake → inspect the content of an intake catalogue.
• mdc <backend> index → push records from catalogue into your index backend
• mdc <backend> delete → remove records by facet match

Features

• Multi-backend discovery: POSIX, S3/MinIO, Swift (async REST), Intake
• Two-stage pipeline: crawl → catalogue then catalogue → index
• Schema driven: strong types (e.g. string, datetime[2], float[4], string[])
• DRS dialects: packaged CMIP6/CMIP5/CORDEX; build your own via inheritance
• Path specs & data specs: parse directory/filename parts and/or read dataset attributes/vars
• Special rules: conditionals, cache lookups and function calls (e.g. CMIP6 realm, time aggregation)
• Index backends: MongoDB (Motor), Solr
• Sync + Async APIs and a clean CLI
• Docs: Sphinx with pydata_sphinx_theme

Install

   pip install metadata-crawler
   conda install -c conda-forge metadata-crawler

Quickstart (CLI)

   # 1) Crawl → write catalogue
   mdc add \
     cat.yaml \
     --config-file drs_config.toml \
     --dataset cmip6-fs,obs-fs \
     --threads 4 --batch-size 100

   # 2) Index from catalogue → Solr (or Mongo)
   mdc solr index \
     cat.yaml \
     --server localhot:8983

   # 3) Delete by facets (supports globs on values)
   mdc delete \
     --server localhost:8983 \
     --facets "file *.nc" --facets "project CMIP6"

[!NOTE] The CLI is a custom framework inspired by Typer (not Typer itself). Use --help on any subcommand to see all options.

Minimal config (drs_config.toml)

   # === Canonical schema ===
   [drs_settings.schema.file]
   key      = "file"
   type     = "path"
   required = true
   indexed  = true
   unique   = true

   [drs_settings.schema.uri]
   key      = "uri"
   type     = "uri"
   required = true
   indexed  = true

   [drs_settings.schema.variable]
   key          = "variable"
   type         = "string[]"
   multi_valued = true
   indexed      = true

   [drs_settings.schema.time]
   key     = "time"
   type    = "datetime[2]"     # [start, end]
   indexed = true
   default = []

   [drs_settings.schema.bbox]
   key     = "bbox"
   type    = "float[4]"        # [W,E,S,N]
   default = [0, 360, -90, 90]

   # === Dialect: CMIP6 (example) ===
   [drs_settings.dialect.cmip6]
   sources   = ["path","data"]         # path | data | storage
   defaults.grid_label = "gn"
   specs_dir  = ["mip_era","activity_id","institution_id","source_id","experiment_id","member_id","table_id","variable_id","grid_label","version"]
   specs_file = ["variable_id","table_id","source_id","experiment_id","member_id","grid_label","time"]

   [drs_settings.dialect.cmip6.special.realm]
   type   = "method"
   method = "_get_realm"
   args   = ["table_id","variable_id","__file_name__"]

   [drs_settings.dialect.cmip6.special.time_aggregation]
   type   = "method"
   method = "_get_aggregation"
   args   = ["table_id","variable_id","__file_name__"]

   # === Dialect: CORDEX (bbox by domain) ===
   [drs_settings.dialect.cordex]
   sources   = ["path","data"]
   specs_dir = ["project","product","domain","institution","driving_model","experiment","ensemble","rcm_name","rcm_version","time_frequency","variable","version"]
   specs_file= ["variable","domain","driving_model","experiment","ensemble","rcm_name","rcm_version","time_frequency","time"]

   [drs_settings.dialect.cordex.special.bbox]
   type   = "call"
   method = "dialect['cordex']['domains'].get('{{domain | upper }}', [0,360,-90,90])"

   [drs_settings.dialect.cordex.domains]
   EUR-11 = [-44.14, 64.40, 22.20, 72.42]
   AFR-44 = [-24.64, 60.28, -45.76, 42.24]

   # === Datasets ===
   [cmip6-fs]
   root_path  = "/data/model/global/cmip6"
   drs_format = "cmip6"             # dialect name
   fs_type    = "posix"

   [cmip6-s3]
   root_path        = "s3://test-bucket/data/model/global/cmip6"
   drs_format       = "cmip6"
   fs_type          = "s3"
   storage_options.endpoint_url = "http://127.0.0.1:9000"
   storage_options.aws_access_key_id = "minioadmin"
   storage_options.aws_secret_access_key = "minioadmin"
   storage_options.region_name = "us-east-1"
   storage_options.url_style   = "path"
   storage_options.use_ssl     = false

   [obs-fs]
   root_path  = "/arch/observations"
   drs_format = "custom"
   # define your specs_dir/specs_file or inherit from another dialect

Concepts

Schema (facet definitions)

Each canonical facet describes:

• key: where to read value ("project", "variable",)
• type: string, integer, float, datetime, with arrays like float[4], string[], datetime[2], or special types like file, uri, fs_type, dataset, fmt
• required, default, indexed, unique, multi_valued

Dialects

A dialect tells the crawler how to interpret paths and read data:

• sources: which sources to consult (path, data, storage) in priority
• specs_dir / specs_file: ordered facet names encoded in directory and file names
• data_specs: pull values from dataset content (attrs/variables); supports __variable__ and templated specs
• special: computed fields (conditional | method | function)
• Optional lookups (e.g., CORDEX domains for bbox)

Path specs vs data specs

• Path specs parse segments from the path, e.g.: /project/product/institute/model/experiment/.../variable_time.nc
• Data specs read from the dataset itself (e.g., xarray/global attribute, variable attributes, per-var stats). Example: gather all variables __variable__, then their units with a templated selector.

Inheritance

Create new dialects/datasets by inheriting:

   [drs_settings.dialect.reana]
   inherits_from = "cmip5"
   sources       = ["path","data"]
   [drs_settings.dialect.reana.data_specs.read_kws]
   engine = "h5netcdf"

Python API

Async

   import asyncio
   from metadata_crawler.run import async_add, async_index, async_delete

   async def main():
       # crawl → catalogue
       await async_add(
           "cat.yaml",
           config_file="drs_config.toml",
           dataset_names=["cmip6-fs"],
           threads=4,
           batch_size=100,
       )
       # index → backend
       await async_index(
           "solr",
           "cat.yaml",
           config_file="drs_config.toml",
           server="localhost:8983",
       )
       # delete by facets
       await async_delete(
           config_path="drs_config.toml",
           index_store="solr",
           facets=[("file", "*.nc")],
       )

   asyncio.run(main())

Sync (simple wrapper)

   import asyncio
   from metadata_crawler import add

   add(
       store="cat.yaml",
       config_file="drs_config.toml",
       dataset_names=["cmip6-fs"],
   )

Index backends

• MongoDB (Motor): upserts by unique facet (e.g., file), bulk deletes (glob → regex)
• Solr: fields align with managed schema; supports multi-valued facets

Contributing

Development install:

   git clone https://github.com/freva-org/metadata-crawler.git
   cd metadata-crawler
   pip install -e .

PRs and issues welcome. Please add tests and keep examples minimal & reproducible (use the MinIO compose stack). Run:

   python -m pip install tox
   tox -e test lint types

Benchmarks

For benchmarking you can create a directory tree with roughly 1.5 M files by calling the create-cordex.sh script in the dev-env folder:

./dev-env/create-cordex.sh
python dev-env/benchmark.py --num-files 20000

See code-of-conduct.rst and whatsnew.rst for guidelines and changelog.

Use MinIO or LocalStack via docker-compose and seed a bucket (e.g., test-bucket). Then point a dataset’s fs_type = "s3" and set storage_options.

Documentation

Built with Sphinx + pydata_sphinx_theme. Build locally:

   tox -e docs