DBCenter

A real PostgreSQL, embedded in your process. No server to install, no Docker, no connection string to configure - the binaries are fetched once and the database runs as a child process against a directory you choose.

from dbcenter import DBCenter

db = DBCenter(path="./data")
users = db.collection("users")
users.insert_one({"name": "Ada", "age": 36})
print(users.find({"age": {"$gt": 30}}))

Because it is PostgreSQL rather than an imitation of it, the extensions come too. PostGIS and pgvector are downloaded and installed on first use:

db.install_extension("postgis")
db.query("""
    SELECT ST_Distance(
        ST_SetSRID(ST_MakePoint(29.0, 41.0), 4326)::geography,
        ST_SetSRID(ST_MakePoint(32.85, 39.93), 4326)::geography
    ) / 1000 AS km
""")

store = db.vector_store("documents", dimension=1536)
store.add(ids=["a"], embeddings=[[0.1] * 1536])

Fuzzy text search is pg_trgm, which ships with PostgreSQL:

db.collection("cities").fuzzy_search("Istanbu", "name")

What you get

• A real PostgreSQL. Its planner, its transactions, its types. db.query() takes any SQL, with bound parameters.
• PostGIS and pgvector, installed on demand. The Linux PostGIS build carries its own GEOS and PROJ, so it works on a machine with nothing installed.
• A collection API over ordinary relational tables. Columns are added as documents introduce fields.
• Sync to storage, optionally, with multi_user_sync=True and a SyncEngine: a local directory or Google Drive from Python, and S3 or anything S3-compatible through the dbcenter-serve container. Tables are written as Parquet with a column per field, and columnar=True reads them back through DuckDB for aggregates.

Notes

The PostgreSQL distribution is downloaded to ~/.dbcenter on first use, about 60MB, and reused after that. Extension archives are verified against digests pinned in the source, so a substituted one is refused.

PostgreSQL refuses to run as root, so neither can this. In a container that runs as root - which most base images do - create a user first and start the database as them; the error says so if you forget.

In Google Colab, which runs everything as root, create a user and run your script as them, with a data directory they can write to:

!id -u dbuser >/dev/null 2>&1 || useradd -m dbuser
!sudo -u dbuser python3 /path/to/your_script.py

Extensions are built per PostgreSQL major version. Asking for one with no published build is an error naming the versions that exist, rather than a download that fails later.

MIT licensed. Source at github.com/heyobi/dbcenter.