Current limitations

This page defines the implemented boundary. The specification and RFCs may describe behavior that is not yet exposed by a released runtime.

Platforms

• CPython 3.11 through 3.14.
• Linux x86_64 with glibc 2.28 or later.
• macOS arm64 11 or later.
• No initial Windows, Linux arm64, macOS x86_64, or free-threaded wheels.

Indexes and Queries

• IVF is the only index family. Postings may use source, LVQ4, or LVQ8 encoding.
• Squared L2 and cosine are supported.
• Query vectors are one-dimensional; batch query execution is not implemented.
• Equal-distance result order is unspecified.
• Source vectors must have one fixed dimension and finite float32 or float64 elements. Computation canonicalizes values to float32.
• Source filters are SQL strings rather than a typed expression API.
• Automatic reclamation of an unused shared centroid artifact is not implemented.

Catalog and Storage

• SQLite is the only catalog implementation.
• Index construction writes Parquet. Iceberg writing is not implemented.
• Storage access supports canonical file, S3, and HDFS locations.
• Parquet sources have no snapshot isolation. Replacing registered files in place is the application’s responsibility.
• Native Iceberg reads validate exact table identity and snapshot. Registration currently starts from a metadata location rather than a logical catalog lookup.

Runtime

• The supported execution runtime is embedded DataFusion.
• Client/server transport is under development and is not a stable deployment mode.
• Distributed engine adapters are not bundled.
• Build submission is process-local. Published metadata and index snapshots are durable, but in-progress jobs do not survive process restart.
• ParqDB does not provide replication, a highly available catalog, tenant isolation, authentication, or a managed service control plane.