spark_sql_migrations

Chained, idempotent SQL migrations for Spark and Databricks.

Alembic-shaped schema migrations for a Spark/Delta catalog: your migrations are plain .sql files linked into a chain by revision headers, the applied revision is recorded in a version table inside your own schema, and each run applies only the tail that has not been applied yet.

Why

Spark and Databricks have no migration story of their own. Delta gives you CREATE TABLE IF NOT EXISTS, but nothing that tracks which DDL a given catalog has seen, and nothing that lets you write a change once and have it apply on a laptop and on Databricks alike. spark_sql_migrations is the small amount of machinery that closes that gap, and nothing more: no ORM, no autogeneration, no Python migration scripts. You write SQL.

Installing

spark_sql_migrations declares no Spark of its own — the extras decide which one you get:

pip install "spark_sql_migrations[local]"        # pyspark + delta-spark, for laptops and CI
pip install "spark_sql_migrations[databricks]"   # databricks-connect
pip install spark_sql_migrations                 # no Spark at all

Pick exactly one flavour. databricks-connect ships its own top-level pyspark/ and delta/ packages, so installing it alongside pyspark/delta-spark silently clobbers both.

Unreleased changes can be consumed as the wheel built by CI (the dist artifact on any green run) or built locally with pants package //:dist.

The two kinds of migration

Migrations come from two different owners, and they behave differently:

The bootstrap chain is re-applied on every run, so it must be idempotent. Your own chains are applied once each and their head revision is written to the version table.

Laying out your chains

Point spark_sql_migrations at a directory holding one or both chains:

your_project/
└── migrations/                 <- this is --migrations-dir
    ├── all_spark_migrations/   <- runs everywhere
    └── dbr_only_migrations/    <- runs only when is_dbr()

Choose deliberately. SQL that only Databricks understands must not live in all_spark_migrations/, or your local runs will break.

Running migrations

spark-sql-migrations run     --cat spark_catalog     --schema default     --migrations-dir path/to/migrations

The command configures logging via logging.basicConfig(), so each applied migration is logged to stderr at INFO. It leaves an already-configured root logger alone, and every other entry point into this library inherits the host application's logging untouched. The command needs a Spark flavour installed; it fails to import on a bare install. python -m spark_sql_migrations.spark_sql.spark_sql is equivalent.

Or from Python, which is what a consuming project usually wraps:

import os

from spark_sql_migrations.spark_sql.spark_sql import main


def get_migrations_dir():
    """the parent of this project's own migration chains."""
    return os.path.dirname(__file__)


main(cat="spark_catalog", schema="default", migrations_dir=get_migrations_dir())

main builds its own session via get_spark() and, being library code, leaves logging to you. To supply your own — a session already configured by your job, say — call run_migrations(spark, cat, schema, output_folder, migrations_root) directly.

Every rendered statement is written to a timestamped folder under migrations_out/ before it is executed, so you can always read the exact SQL a run applied.

Writing a migration

Generate one rather than hand-rolling the header:

spark-sql-migrations create_new_migration     --message "add batch id to metrics"     --output-path path/to/migrations/all_spark_migrations

That writes <yymmdd>_<slug>_<revision_id>.sql from the template shipped in the package (override with --template-path), with revision_id filled in:

-- revision_id:e5ce0039b32b;
-- prev_revision_id:;
begin
create table if not exists {{cat}}.{{schema}}.test_table(int_id bigint, stuff string);
end;

Then set prev_revision_id to the revision this one follows. Ordering comes from that chain, not from the filename — the date prefix is for humans. A chain must have exactly one root (empty prev_revision_id) and no forks.

{{cat}} and {{schema}} are Jinja placeholders rendered at apply time; never hardcode a catalog or schema.

Idempotency

Prefer IF NOT EXISTS. Where Databricks doesn't offer it — notably ALTER TABLE ... ADD COLUMN — guard with a SQLSTATE exit handler:

-- revision_id:57037b6c19b4;
-- prev_revision_id:07990e2a101e;
begin
  -- idempotent add: swallow FIELD_ALREADY_EXISTS (SQLSTATE 42710) if the column
  -- is already present. Databricks has no ADD COLUMN IF NOT EXISTS, so use a
  -- SQL-scripting EXIT handler (CONTINUE handlers are unsupported).
  declare exit handler for sqlstate '42710'
  begin end;
  alter table {{cat}}.{{schema}}.metrics add column metric_batch_id string;
end;

Local vs Databricks

get_spark() returns a DatabricksSession when is_dbr() is true and a Delta-configured local SparkSession otherwise, so the same code runs in both places. Local behaviour is tuned by environment variables:

Development

Built with Pants. Two resolves: python-default for the library, py-reqs-dev for tests.

# black, isort, flake8, mypy. `::` and not `/`, or the subtrees go unchecked.
pants fmt lint check spark_sql_migrations:: spark_sql_migrations_test:: scripts::

# unit, then integration as its own invocation
pants test --use-coverage spark_sql_migrations_test/:: -spark_sql_migrations_test/integration::
pants test spark_sql_migrations_test/integration::

# wheel + sdist
pants package //:dist

scripts/run_local.sh runs everything CI runs, including building the distribution and verifying all three install flavours. Branch coverage over spark_sql_migrations/ is gated at 94%.

Dependencies are locked. Edit spark_sql_migrations/requirements.txt or spark_sql_migrations_test/requirements-dev.txt, then pants generate-lockfiles — never hand-edit a lockfile.

License

Apache License 2.0 — see LICENSE.