Universal Metadata Format

Universal Metadata Format (UMF) is tablespec’s source-table contract. A UMF spec records the table name, column names, source data types, nullability, keys, relationships, source declaration, and validation expectations.

This page is for readers who need to author or review UMF files. Every tablespec operation starts by loading a UMF spec and validating it with Pydantic models.

Formats

Split directory (canonical). One table.yaml file stores table-level metadata. One file per column lives under columns/. This is the review-friendly editing format because a column change is a one-file diff:

tables/medical_claims/
├── table.yaml            # name, description, keys, relationships
├── expectations.yaml     # table-level expectations
└── columns/
    ├── claim_id.yaml     # one file per column
    ├── member_id.yaml
    └── ...

# table.yaml
canonical_name: Medical Claims
description: Healthcare claims - source-faithful ingested bronze
primary_key:
  - claim_id
table_name: medical_claims
version: '1.0'

# columns/claim_id.yaml
column:
  data_type: VARCHAR
  description: Unique claim identifier
  length: 50
  name: claim_id
  nullable:
    MD: false
    MP: false

JSON (artifact standard). A JSON UMF file stores the whole table contract in one .json document. Compiled pipelines can consume this single-file artifact. tablespec convert translates between split directories and JSON, and UMFLoader auto-detects both formats.

Single-file YAML (legacy). Older whole-UMF YAML documents are loadable from Python via load_umf_from_yaml, but the CLI refuses them and points at the explicit migration helper.

Column types

data_type is one of: VARCHAR, CHAR, TEXT, INTEGER, DECIMAL, FLOAT, DATE, DATETIME, TIMESTAMP, BOOLEAN. Sized types use length (VARCHAR/CHAR), and precision/scale (DECIMAL).

The PySpark schema generator targets the raw stage: the landing table that captures source records before tablespec casts them. In that raw stage, dates land as strings and are cast during ingest. That is why DATE maps to StringType() in the raw PySpark schema but to DATE in the typed SQL DDL. The validation model follows the same raw/typed split.

Nullability per context

nullable maps arbitrary context keys to booleans (the model is extra="allow" — any domain works). In healthcare specs the common keys are MD (Medicaid), MP (Medicare Part D), and ME (Medicare):

nullable:
  MD: false   # required in the Medicaid feed
  MP: true    # sometimes omitted in the Medicare Part D feed

A spec can name the column that determines which context applies to each row via context_column on the table. Context-dependent nullability compiles to conditional not-null expectations instead of blanket constraints.

Keys and relationships

# table.yaml
primary_key:
  - claim_id
unique_constraints:
  - [member_id, service_date]
relationships:
  foreign_keys:
    - column: member_id
      references_table: members
      references_column: member_id

tablespec validate checks relationship integrity automatically when multiple UMF specs are present. The dbt and Lakeflow emitters turn declared keys into generated tests and expectations.

Derivations and survivorship

Generated or report-style tables can describe where each output column comes from with derivation. A derivation can carry ordered source candidates, source SQL expressions, join filters, row filters, window ordering columns, join-via metadata, and a survivorship rule. That metadata feeds gold SQL generation and the guidebook lineage view.

column:
  name: latest_a1c
  data_type: DECIMAL
  precision: 5
  scale: 2
  derivation:
    candidates:
      - table: observations
        column: VALUE
        priority: 1
        row_filter: "DESCRIPTION = 'Hemoglobin A1c/Hemoglobin.total in Blood'"
        order_by:
          - DATE
    survivorship:
      strategy: highest_priority
      explanation: Prefer the latest qualifying A1C observation.
      default_value: 0
      default_condition: No qualifying observation is available.

Excel review workbooks preserve this structure through the Derivations sheet. That sheet is machine-readable and separate from the human-oriented survivorship sheet, so an Excel-authored derivation can regenerate the same gold SQL after import.

Expectations

Quality rules live in an expectation suite. In tablespec, an expectation is a Great Expectations rule plus metadata that says where it runs (stage), how serious failure is (severity), whether it blocks a load (blocking), and where it came from (generated_from). In split format, column-scoped expectations are stored in the column’s file under validations:; table-level expectations live in expectations.yaml:

# columns/claim_id.yaml (continued)
validations:
  - type: expect_column_values_to_match_regex
    kwargs:
      column: claim_id
      regex: ^C[0-9]+$
    meta:
      severity: warning
      stage: raw
      generated_from: llm

The column model does not have a separate validation_rules or allowed_values field. Constraints are expectations. They can be generated from the UMF metadata as the baseline suite, inferred from profiling, added from an LLM response through tablespec apply-response, or written by hand.

Source declaration

A UMF spec may declare where its rows come from with a source: block. The kind field says whether the source is delimited, parquet, or jdbc. When source: is absent, tablespec treats the table as a delimited flat file described by file_format.

source:
  kind: jdbc
  url: jdbc:sqlserver://localhost:1433;databaseName=northwind
  dbtable: dbo.Orders
  driver: com.microsoft.sqlserver.jdbc.SQLServerDriver
  user: reader
  password_secret_ref: NORTHWIND_PASSWORD   # a *reference*, never the secret

Two properties matter here:

• Credentials are never inlined. JdbcSource rejects a literal password field (extra="forbid"); password_secret_ref names a secret in the runtime’s secret store (an env var, a Databricks secret scope).
• The source kind drives validation. Typed sources such as JDBC and Parquet land natively typed, so their generated suites do not include string-shape raw checks. See staged validation.

Ingestion

The optional ingestion block controls how tablespec generates the raw-to-ingested transform. It can declare mode (for example, incremental), order_by for dedup-latest windows, pre-upsert exclusions, and post-upsert rules. tablespec generate -f ingest and the dbt emitter use this block.

Provenance columns

Every pipeline-complete UMF spec carries eight meta_* provenance columns such as meta_source_name, meta_load_dt, and meta_checksum. The ingest pipeline populates these columns on every row so a downstream reader can trace where the row came from and when it was loaded. tablespec validate requires the provenance columns; spec-producing flows such as JDBC discovery append them automatically. The canonical list is tablespec.ingestion.constants.PROVENANCE_COLUMNS.