Skip to content

datalake_fdw: NUMERIC/DECIMAL in the Parquet format layer #1988

Description

@MisterRaindrop

Summary

contrib/datalake_fdw's Parquet format layer (#1951) refuses numeric columns at CREATE TABLE ... USING iceberg. DECIMAL is the most common column type in real lake tables, so this is the first type to add.

What has to be decided

  • Storage form. Parquet stores DECIMAL four ways (INT32, INT64, FIXED_LEN_BYTE_ARRAY, BYTE_ARRAY). Iceberg's spec fixes the writer to decimal(P,S) with P <= 38 backed by fixed-length bytes; the reader has to accept all four.
  • Unconstrained numeric. atttypmod = -1 has no precision or scale. Computing ((typmod - 4) >> 16) & 65535 without checking gives precision 65535 / scale 65531, so bare numeric columns silently match nothing (the trap in lithium-tech/tea validate.cpp:59-61). Iceberg needs P and S, so the likely answer is to refuse bare numeric at CREATE TABLE and require numeric(P,S).
  • Precision above 38. PostgreSQL allows up to 1000; decimal128 caps at 38. Refuse at CREATE TABLE.
  • Conversion. NumericVar <-> two's-complement int128, the way tea's bridge.cpp:269-277 and numeric_var.cpp do it, with NaN/Infinity refused on write.

Where

  • format/format_types.h / arrow_support.cpp: dl_format_type_refusal() decides what CREATE TABLE accepts; the mapping to arrow::decimal128(P, S) goes next to the other types.
  • arrow_builder.cpp / arrow_decode.c: append and decode. The decoder reads the Arrow C data interface directly; the format string is d:P,S.
  • Tests in test/automation/sqlrepo/smoke/format_parquet/, plus CREATE TABLE refusals in iceberg_am_reject.sql.

Deferred from #1951 on purpose: the framework there is settled, the type work is separate.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Labels

datalakecontrib/datalake_fdw and contrib/datalake_agent: Iceberg lake tables

Type

No type

Projects

No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions