xoverrr (pronounced "crossover")

A tool for cross-database and intra-source data quality checks with detailed discrepancy analysis and reporting.

Supported databases: Oracle, PostgreSQL (+ Greenplum), ClickHouse.

Features

• Five check strategies - row samples, counts grouped by date, whole-table counts, custom SQL, and source-only sniff checks
• Multi-DBMS - tables and views, extensible via adapters
• SQLAlchemy engines - pass any supported source, target, or results connection
• Recent-row exclusion - optionally skip rows that may still be delayed (batch load, replication, or calculation)
• Auto metadata - primary keys and column types from DBMS catalogues (or supply your own primary key)
• Type conversion - application-side normalisation across databases
• Column filters - include / exclude lists; mismatched column names are skipped automatically
• Chunked date ranges - process long periods in N-day windows
• Reports - text or JSON, with example mismatched rows
• Optional persistence - write run results to a third engine for dashboards and audit
• Tests - unit coverage plus Docker-backed integration tests

Quick start

Sample check (Greenplum/PostgreSQL -> Oracle):

from xoverrr import DataQualityChecker, DataReference, CheckResult, CHECK_SUCCESS
from sqlalchemy import create_engine
from datetime import date, timedelta

# 1. Connections
source_engine = create_engine('postgresql://user:pass@localhost:5432/source_db')
target_engine = create_engine('oracle+oracledb://user:pass@localhost:1521/target_db')
results_engine = create_engine('postgresql://user:pass@localhost:5432/dq_audit')

# 2. Checker
checker = DataQualityChecker(
    source_engine=source_engine,
    target_engine=target_engine,
    timezone='Europe/Athens',
    results_engine=results_engine,  # optional
    max_dataframe_size_gb=3,  # per query / chunk from one DB
)

# 3. Tables + date window
source_table = DataReference("employees", schema="hr")
target_table = DataReference("employees", schema="hr")
end_date = date.today()
start_date = end_date - timedelta(days=7)

# 4. Run
result: CheckResult = checker.check_samples(
    source_table=source_table,
    target_table=target_table,
    date_column="hire_date",
    update_column="modified_at",
    date_range=(start_date.strftime('%Y-%m-%d'), end_date.strftime('%Y-%m-%d')),
    chunk_size_days=30,
    custom_primary_key=["employee_id"],
    exclude_columns=["audit_log", "temp_field"],
    tolerance_pct=0.5,
    exclude_recent_hours=3,
    max_examples=5,
    persist_result=DataReference("dq_results", "test"),
    check_name="employees_daily",
    check_tags={"env": "prod", "domain": "hr"},
    report_output_format='text',  # 'json' or 'text'
)

# 5. Result
print(result.run_id)
print(result.report)
if result.status == CHECK_SUCCESS:
    print("Data quality check passed")
else:
    print("Data quality check failed")

Every check method returns a CheckResult:

result = checker.check_samples(...)
result.run_id
result.status
result.report
result.stats
result.details

Which method should I use?

A target engine is required for the first four methods, but it may be the same as source_engine (two tables or schemas in one database). Pass a second engine only when the sides live in different databases.

Check methods

1. Data sample (check_samples)

This method compares row sets and column values over a date range.

result = checker.check_samples(
    source_table=DataReference("table_name", "schema_name"),
    target_table=DataReference("table_name", "schema_name"),
    date_column="created_at",
    update_column="modified_date",
    date_range=("2024-01-01", "2024-12-31"),
    chunk_size_days=30,
    exclude_columns=["audit_timestamp", "internal_id"],
    include_columns=None,
    custom_primary_key=["id", "user_id"],
    tolerance_pct=1.0,
    exclude_recent_hours=24,
    max_examples=3,
)

Main parameters

If custom_primary_key is omitted, the primary key is inferred from metadata (it must exist on at least one side).

2. Counts grouped by date (check_counts_group_by_date)

Compares row counts grouped by a date or timestamp column. Currently only daily aggregates are supported (hour / month / year grouping is not). Suitable for large volumes and for spotting missing or extra rows.

result = checker.check_counts_group_by_date(
    source_table=DataReference("users", "schema1"),
    target_table=DataReference("users", "schema2"),
    date_column="created_at",
    date_range=("2024-01-01", "2024-12-31"),
    chunk_size_days=30,
    tolerance_pct=2.0,
    max_examples=5,
)

Main parameters: source_table, target_table, date_column, date_range, chunk_size_days, tolerance_pct, max_examples, plus the shared persist_result / check_name / check_tags / report_output_format options described above. One-sided date_range is allowed without chunking; chunk_size_days requires both start and end dates.

3. Total counts (check_total_counts)

Whole-table COUNT(*) on each side. Optional date_range / chunk_size_days split the scan into date windows and sum the counts (no per-date breakdown).

result = checker.check_total_counts(
    source_table=DataReference("users", "schema1"),
    target_table=DataReference("users", "schema2"),
    tolerance_pct=2.0,
)

result = checker.check_total_counts(
    source_table=DataReference("users", "schema1"),
    target_table=DataReference("users", "schema2"),
    date_column="created_at",
    date_range=("2024-01-01", "2024-12-31"),
    chunk_size_days=30,
    tolerance_pct=2.0,
)

Main parameters: source_table, target_table, date_column, date_range, chunk_size_days, tolerance_pct, plus the shared persist_result / check_name / check_tags / report_output_format options. date_range and chunk_size_days require date_column. One-sided date_range is allowed without chunking; chunk_size_days requires both start and end dates.

4. Custom query (check_custom_queries)

Compare the results of arbitrary SQL on both sides. A primary key is required.

result = checker.check_custom_queries(
    source_query="""
        SELECT id AS user_id, name AS user_name, created_at AS created_date
        FROM scott.source_table
        WHERE status = :status
    """,
    source_params={'status': 'active'},
    target_query="""
        SELECT user_id, user_name, created_date
        FROM scott.target_table
        WHERE status = :status
    """,
    target_params={'status': 'active'},
    custom_primary_key=["user_id"],
    exclude_columns=["internal_code"],
    tolerance_pct=0.5,
    max_examples=3,
)

Chunking: when both source_params and target_params include start_date and end_date, set chunk_size_days to split the range:

result = checker.check_custom_queries(
    source_query="""
        SELECT id, name, created_at
        FROM scott.source_table
        WHERE created_at >= date_trunc('day', cast(:start_date as date))
          AND created_at < date_trunc('day', cast(:end_date as date)) + interval '1 day'
    """,
    source_params={'start_date': '2024-01-01', 'end_date': '2024-12-31'},
    target_query="""
        SELECT id, name, created_at
        FROM scott.target_table
        WHERE created_at >= date_trunc('day', cast(:start_date as date))
          AND created_at < date_trunc('day', cast(:end_date as date)) + interval '1 day'
    """,
    target_params={'start_date': '2024-01-01', 'end_date': '2024-12-31'},
    custom_primary_key=["id"],
    chunk_size_days=30,
    tolerance_pct=0.5,
)

To skip recently changed rows in custom SQL, add the same flag as that used by the sample check:

CASE WHEN updated_at > (sysdate - 3/24) THEN 'y' END AS xrecently_changed

5. Sniff query (check_sniff_query)

A source-only check. Mark each row with xsniff_passed (y = passed, n = failed).
No target engine or primary key is required:

checker = DataQualityChecker(
    source_engine=source_engine,
    timezone='UTC',
)

Row-level - one flag per row. Use this when you need an issue rate over the full scope; tolerance_pct then means "allow up to N% of rows with xsniff_passed = n":

result = checker.check_sniff_query(
    source_query="""
        SELECT
            order_id,
            amount,
            CASE
                WHEN amount > 0 AND customer_id IS NOT NULL THEN 'y'
                ELSE 'n'
            END AS xsniff_passed
        FROM sales.orders
        WHERE created_at >= :start_date
    """,
    source_params={'start_date': '2024-01-01'},
    tolerance_pct=1.0,
)

Scalar pass/fail - a single xsniff_passed value. The outcome is typically binary (final_diff_score 0 or 100), so leave tolerance_pct at the default 0.0 (fail on any issue):

result = checker.check_sniff_query(
    source_query="""
        SELECT CASE
            WHEN EXISTS (SELECT 1 FROM sales.orders WHERE amount <= 0) THEN 'n'
            ELSE 'y'
        END AS xsniff_passed
    """,
)

Issues-only filter - WHERE keeps only the failing rows and marks every returned row with a literal 'n'.
An empty result means a pass (final_score = 100). Any returned row means a fail (issue_rows_pct = 100 for that result set). As with the scalar pattern, keep the default tolerance_pct=0.0 so that the check fails if any issue is found:

result = checker.check_sniff_query(
    source_query="""
        SELECT
            order_id,
            amount,
            'n' AS xsniff_passed
        FROM sales.orders
        WHERE amount <= 0
          AND created_at >= :start_date
    """,
    source_params={'start_date': '2024-01-01'},
)

Use this when you need only to establish that issue rows exist (and to include their keys and attributes in the report). Prefer the row-level CASE pattern above when you need a rate over the full checked scope.

Main parameters: source_query, source_params, chunk_size_days (when the parameters include dates), tolerance_pct, max_examples, plus the shared persistence, naming, and report-format options.

Useful stats fields:

Metric calculation

All methods expose result.stats.final_diff_score and result.stats.final_score.

Quality score (every method):

final_score = 100 - final_diff_score

Scores range from 0 to 100%. A higher final_score indicates better quality.
Pass or fail is determined by the tolerance:

• final_diff_score > tolerance_pct -> CHECK_FAILED
• otherwise -> CHECK_SUCCESS

check_samples / check_custom_queries

final_diff_score =
    (dup_source_rows_pct * 0.1)
  + (dup_target_rows_pct * 0.1)
  + (source_only_rows_pct * 0.15)
  + (target_only_rows_pct * 0.15)
  + (issue_rows_pct * 0.5)

check_counts_group_by_date / check_total_counts

sum_of_absolute_differences = abs(source_count - target_count)  per date (or one total)
sum_of_common_counts        = min(source_count, target_count)   per date (or one total)

final_diff_score = 100 * sum_of_absolute_differences
                       / (sum_of_absolute_differences + sum_of_common_counts)

check_sniff_query

issue_rows_pct = (rows with xsniff_passed = 'n') / (checked rows) * 100
final_diff_score = issue_rows_pct

An empty result yields final_diff_score = 0 (and a score of 100).

With the issues-only filter pattern (WHERE ... plus a literal 'n' AS xsniff_passed), checked rows are only the filtered issue rows, so any non-empty result yields issue_rows_pct = 100.

Shared behaviour

Chunked processing (chunk_size_days)

Available on check_samples, check_counts_group_by_date, check_total_counts, check_custom_queries, and check_sniff_query. It splits a date range into N-day windows, runs each chunk, then aggregates the metrics and examples. This is useful for long ranges or large tables. Chunking needs both start and end dates; without chunk_size_days, either bound may be omitted.

• check_custom_queries: both sides must supply start_date and end_date in their parameters
• check_sniff_query: chunking uses start_date / end_date in source_params
• check_total_counts: date_column plus date_range; chunk COUNT(*) values are summed

Status values

Result persistence

With results_engine set and persist_result=DataReference(...), one row is written per run to that table. The schema must already exist; the table is created if it is missing. The columns cover status, metadata, statistics, details JSON, and the text report. Persistence is skipped if persist_result is omitted. If persistence was requested and the write fails, the check status becomes failed.

Logging

Each run has a run_id on the returned CheckResult (also stored when persistence is enabled, and included in CheckResult.to_dict() / JSON reports):

2024-01-15 10:30:45 - INFO - xoverrr.core - Check run started: run_id=a3f2c8b91d4e5678 check_name=employees_daily check_type=samples
2024-01-15 10:30:45 - INFO - xoverrr.core._check_samples - Query executed in 2.34s
2024-01-15 10:30:46 - INFO - xoverrr.core._check_samples - Source: 150000 rows, Target: 149950 rows
2024-01-15 10:30:47 - INFO - xoverrr.utils.compare_dataframes - Comparison completed in 1.2s
2024-01-15 10:30:47 - INFO - xoverrr.core - Check run finished: run_id=a3f2c8b91d4e5678 status=CHECK_SUCCESS

Performance notes

• DataFrame size limit: max_dataframe_size_gb on the checker (default 3 GB per query / chunk from one DB)
• Rough benchmark: two samples of about 1 million rows * 10 columns (about 330 MB each) compared in about 3 s (Intel Core i5 / 16 GB RAM)

Example report

Text output (report_output_format='text'):

================================================================================
2025-11-24 20:09:40
run_id: a3f2c8b91d4e5678
version: *.*.*
source db type: postgresql
target db type: oracle
SAMPLES CHECK REPORT:
hr.employees
VS
hr.employees
================================================================================
timezone: Europe/Athens

    SELECT employee_id, first_name, last_name, salary, department_id, hire_date,
           case when updated_at > (now() - INTERVAL '3 hours') then 'y' end as xrecently_changed
    FROM hr.employees
    WHERE 1=1
        AND hire_date >= date_trunc('day', cast(:start_date as date))
        AND hire_date < date_trunc('day', cast(:end_date as date)) + interval '1 day'

    params: {'start_date': '2025-11-17', 'end_date': '2025-11-24'}
----------------------------------------

    SELECT employee_id, first_name, last_name, salary, department_id, hire_date,
           case when updated_at > (sysdate - 3/24) then 'y' end as xrecently_changed
    FROM hr.employees
    WHERE 1=1
        AND hire_date >= trunc(to_date(:start_date, 'YYYY-MM-DD'), 'dd')
        AND hire_date < trunc(to_date(:end_date, 'YYYY-MM-DD'), 'dd') + 1

    params: {'start_date': '2025-11-17', 'end_date': '2025-11-24'}
----------------------------------------

SUMMARY:
  Source rows: 105
  Target rows: 105
  Duplicated source rows: 0
  Duplicated target rows: 0
  Only source rows: 0
  Only target rows: 0
  Comparable rows: 105
  Passed rows: 103
----------------------------------------
  Source only rows %: 0.00000
  Target only rows %: 0.00000
  Duplicated source rows %: 0.00000
  Duplicated target rows %: 0.00000
  Issue rows %: 1.90476
  Final discrepancies score: 0.95238
  Final data quality score: 99.04762
  Source-only key examples:
  Target-only key examples:
  Duplicated source key examples:
  Duplicated target key examples:
  Skipped source columns: audit_log, temp_field
  Skipped target columns:

ISSUE BREAKDOWN:
  Max issue %: 1.90476
  Issue counts by column:

 column_name  issue_count
     salary                2

  Issue examples:

 primary_key column_name source_value target_value
         101      salary        50000        51000
         102      salary        60000        60500

================================================================================

Known limitations

Oracle thin client & TIMESTAMP WITH TIME ZONE

With the Oracle thin client and check_custom_queries, TIMESTAMP WITH TIME ZONE columns lose their time-zone context in the result set.

Workaround - cast to TIMESTAMP in SQL:

source_query = """
    SELECT
        order_id,
        CAST(created_at AT TIME ZONE 'Europe/Paris' AS TIMESTAMP) AS created_at,
        amount
    FROM orders
    WHERE status = 'completed'
"""

CI/CD

GitHub Actions (see .github/workflows/):

The tag must match the version in pyproject.toml and src/xoverrr/version.py (for example tag v1.4.6). Release details and the one-time PyPI Trusted Publishing setup are in COTRIBUTE.md.

Local commands used by CI are in the Makefile (make test-unit, make up-dbs, make build).