Store.md

Store — Persistence for Attribute Functions

The Store class (store/store.py) provides SQLite-backed persistence for attribute functions. It uses SqliteDict as a key/blob store, where each AF is serialized (pickled) and stored by its UUID.

Basic usage

from fdm.attribute_functions import TF, RF
from store.store import Store

# Create a store (backed by a SQLite file):
store = Store(file_name="my_store.sqlite")

# Create some AFs:
users = RF({
    1: TF({"name": "Alice"}),
    2: TF({"name": "Bob"}),
})

# Register (persist) the AF:
store.register(users)

# Retrieve by UUID:
loaded = store.get(users.uuid)
loaded[1].name  # → "Alice"

Swizzling and unswizzling

When an AF contains references to other AFs (e.g. an RF whose values are TFs), these references are swizzled during serialization: each nested AF is replaced by an AttributeFunctionSentinel containing only its UUID. On deserialization, sentinels are unswizzled back to full AF objects on first access via the store.

This enables lazy loading: only the AFs you actually access are loaded from disk.

Ephemeral state

The following AF state is stripped during pickling and must be re-attached after deserialization:

• computed — computed attribute value definitions (lambdas)
• default — default fallback function (lambda)
• domain — active domain set
• store — store reference (re-attached automatically on load)

This is by design: lambdas and closures are not safely serializable with standard pickle.

Limitations

• Read-only swizzling: unswizzling currently works for reads only; writes through swizzled references are a known TODO.
• No query pushdown: the store treats values as opaque blobs — filtering and projection happen in memory after loading.
• Security: unpickling untrusted data is a remote code execution vector. Only load from trusted sources.