data.drivers._pelt._codecs

An immutable encode/decode pair for one PostgreSQL type.

decode_* take the raw column bytes; encode_*return the raw column bytes. prefers_binaryselects the wire format the driver requests when it controls the choice.

A lock-guarded OID→Codecmap that hands out immutable snapshots.

The E1 hot-path leaf codecs defined inline in this module (ints, floats, bool, text).

Every non-parametric E2 codec, gathered from each family'sLEAF_CODECStuple.

Array/composite/range/enum families contribute nothing here — they are parametric (the element/field codec is only known at plan time), so theirLEAF_CODECSare empty and they are wired viabuild_codec_plan() instead. The sibling imports are deferred to call time to break the_codecs↔ family import cycle.

The full set of non-parametric codecs the default registry pre-loads (E1 + E2 leaves).

A fresh registry pre-loaded with the E1 hot-path codecs plus the E2 leaf families.

Registration keeps the fail-loud conflict discipline: each OID is registered exactly once, so a duplicate OID across families (a packaging bug) raisesValueErrorat build time rather than silently last-wins.

BuildDEFAULT_REGISTRYlazily on first access.

DEFAULT_REGISTRYcannot be built eagerly at module import: doing so reads each E2 family's LEAF_CODECS while a family-first import is still resolving its from ._codecs import Codec— the family is only half-initialized at that point, so the read would crash with a partial-initAttributeError. Deferring the build to first attribute access keeps the process-wide default a one-liner for callers while staying cycle-safe under any import order.

The build is double-checked-lock guarded so concurrent first accesses on free-threaded workers materialize exactly one registry. Treated read-only after creation; connections register extra codecs on a freshbuild_default_registry() for per-database enums, arrays, ranges, and composites.

Array OID → element OID (lazy import to stay clear of the family import cycle).

Range OID → element OID, so the planner can resolve a parametric range column's element.

Element OIDs cross-referenced from the leaf families: int4 23, int8 20, numeric 1700, timestamp 1114, timestamptz 1184, date 1082 — seepg_type.dat. Lazy-imported to stay clear of the_codecs↔ family import cycle.

Reject binary bytes for an OID without a proven decoder.

Text-format fallback for an unregistered OID: decode UTF-8.

PostgreSQL's text wire format is the type'stypoutputstring in the server encoding (UTF-8 for any modern deployment), so a faithful str is the safe, non-crashing default.

Resolve the non-NULL decoder for one column from its OID +format_code.

format_code is 1 for binary and 0 for text (PostgreSQL's Bind/RowDescription convention). A registered OID uses the matching half of itsCodec; a known array/range OID is resolved parametrically againstsnapshot; an unregistered OID falls back to raw-bytes (binary) or UTF-8 (text) so the row path never crashes on an unknown type.

Choose one explicit result format per described column.

A registered codec opts into binary throughprefers_binary. Known parametric arrays and ranges opt in when their element codec is available. Every unresolved OID stays text, preserving PostgreSQL's lossless output fallback instead of requesting undecodable bytes.

Copy a statement description with the formats selected byBind.

Precompute one(bytes | None) -> Anydecoder per column for a result set.

Each returned decoder maps a raw column value (from aDataRow) to a Python object:None (SQL NULL) passes straight through as None; otherwise the column bytes go through the resolved per-column decoder. Computing the plan once per RowDescription— rather than re-resolving the codec for every cell — keeps the per-row loop a tuple of bound closures, which is what makes free-threaded row decode cheap (epic E6).registry_snapshot is an immutable snapshot() view, so the plan is lock-free.