Dqx Apply Checks

DQX — Apply quality checks

Apply checks with DQEngine. There are six method families — pick by (1) rule form and (2) what output you want.

For the end-to-end methods see the dqx-end-to-end skill. Streaming uses the same methods as batch.

Examples

Using DQX Classes

from databricks.labs.dqx import check_funcs
from databricks.labs.dqx.engine import DQEngine
from databricks.labs.dqx.rule import DQRowRule, DQDatasetRule
from databricks.sdk import WorkspaceClient
# `spark` is available in Databricks notebooks / jobs. Locally, create it with
# `from pyspark.sql import SparkSession; spark = SparkSession.builder.getOrCreate()`.

dq = DQEngine(WorkspaceClient())

checks = [
    DQRowRule(criticality="warn",  check_func=check_funcs.is_not_null,   column="col3"),
    DQDatasetRule(criticality="error", check_func=check_funcs.is_unique, columns=["col1", "col2"]),
]

df = spark.read.table("catalog.schema.input")

# Option A — one output DataFrame with _warnings / _errors columns appended.
annotated = dq.apply_checks(df, checks)

# Option B — split on error severity. Warnings ride along with valid_df.
valid_df, invalid_df = dq.apply_checks_and_split(df, checks)

Using Metadata (loaded from storage)

dq = DQEngine(WorkspaceClient())

checks_metadata = """
- criticality: warn
  check:
    function: is_not_null
    arguments:
      column: col3

- criticality: error
  check:
    function: is_unique
    arguments:
      columns: [col1, col2]
"""

# Option A — one output DataFrame with _warnings / _errors columns appended.
annotated = dq.apply_checks_by_metadata(df, checks_metadata)

# Option B — split on error severity. Warnings ride along with valid_df.
valid_df, invalid_df = dq.apply_checks_by_metadata_and_split(df, checks_metadata)

Result shape

By default two struct columns are appended:

• _errors — an array of result structs, one per failing error-criticality rule.
• _warnings — same, for warn-criticality rules.

Each result struct contains: name, message, columns, filter, function, run_time, run_id, user_metadata, rule_fingerprint, rule_set_fingerprint, and (for rules that couldn't be evaluated) skipped=true with a reason.

Column names and the skipped-entry behavior are configurable — see Additional Configuration for ExtraParams, result_column_names, and suppress_skipped.

Streaming

Use spark.readStream / writeStream as normal; the same apply_checks* calls work on streaming DataFrames. For incremental batch-style execution out of the box, use apply_checks_and_save_in_table with the default AvailableNow trigger (see dqx-end-to-end).

Multiple inputs

• apply_checks_and_save_in_tables(run_configs) — one call processes many (input, output) pairs defined by RunConfig entries.
• apply_checks_and_save_in_tables_for_patterns(patterns=["catalog.schema.*"], checks_location="catalog.schema.checks", run_config_template=RunConfig(...)) — expand wildcards against Unity Catalog. patterns is a list[str]; checks_location is required and is used as a template for per-table file lookups or as the shared Delta table. See Applying Checks for the full signature.

Do / Don't

• Do fail fast: DQEngine.validate_checks(checks) raises on bad definitions before you touch data.
• Do persist the split result (invalid_df) to a quarantine table — not just the annotated DataFrame — so downstream consumers see only good rows.
• Don't mix DQX classes and metadata in one call; pick one form per invocation.
• Don't drop the _errors / _warnings columns unless you've materialised the result elsewhere — they're how downstream tooling sees what failed.

Canonical docs: https://databrickslabs.github.io/dqx/docs/guide/quality_checks_apply.