Mz Test

Testing Materialize

Cluster sizes

mzcompose-based tests and bin/environmentd take their cluster replica sizes from cluster_replica_size_map() in misc/python/materialize/mzcompose/__init__.py, named scale=<n>,workers=<w>[,mem=<n>GiB], e.g. CREATE CLUSTER c SIZE 'scale=1,workers=1'. Plain sizes like '1' fail there with "unknown cluster replica size". The emulator image run directly knows only cloud-style sizes (25cc, 50cc, ...).

Unit tests

Run unit tests with cargo test. If available, use cargo nextest instead. If these complain about METADATA_BACKEND_URL, then use bin/cargo-test instead. When using -p, our packages typically have an mz- suffix, e.g., mz-adapter. Rewrite datadriven test expectations with REWRITE=1 cargo test .... Do not manually update *.snap files. Use cargo test followed by cargo insta accept to update snapshot files.

Some tests require access to a metadata store. Set environment variables to COCKROACH_URL=postgres://root@localhost:26257 METADATA_BACKEND_URL=postgres://root@localhost:26257.

Control log output during tests with MZ_TEST_LOG_FILTER=<level>, e.g. MZ_TEST_LOG_FILTER=debug cargo test ....

The first run includes a cargo build step that can take several minutes. Use a bash timeout of at least 600000ms (10 min) for test commands.

sqllogictest

Run sqllogictest files with:

bin/sqllogictest -- PATH [PATH ...]

PATH is relative to the repo root, usually a file in test/sqllogictest/. sqllogictest accepts a list of files, not just one, and runs them in order. Prefer batching files into one invocation: each invocation bootstraps a fresh catalog and cluster, which is slow, so batching amortizes that startup cost across files instead of paying it per file. It also matches how CI drives the suite, so goldens generated one-file-per-process can diverge from CI's batched context. Rewrite expected results with bin/sqllogictest -- --rewrite-results PATH.

When a plan changes only because of nondeterministic IDs in the output, normalize it with sqllogictest's replace directive rather than running files isolated to dodge the ID noise. Treat catalog-bloat plan drift as a real signal to investigate, not ID text noise.

Add --optimized when running many or large files: it significantly improves execution speed at the cost of a longer one-time compile. For a single small file, skip --optimized, since the extra compile time outweighs the runtime savings. You can use -v for printing each query before it is run, e.g. to figure out where we are crashing.

When sqllogictest gives InconsistentViewOutcome

CI runs many SLTs with --auto-index-selects, which wraps most successful SELECTs in an indexed view and checks that the view returns the same rows as the one-shot query. A query that behaves differently when wrapped fails this check even though a normal run passes.

You can exempt a file from that mode by adding its path to both configs in test/sqllogictest/mzcompose.py:

• compileFastSltConfig (the sqllogictest-fast PR job): add to tests_without_views.
• compileSlowSltConfig (the Nightly slow-tests job): add to tests_no_auto_index_selects.

The file still runs, just without the indexed-view wrapping.

Testdrive

Run testdrive files with:

bin/mzcompose --find testdrive run default -- FILENAME.td

FILENAME.td is a file in test/testdrive/, relative to that directory (not the repo root).

mz_system and materialize are built-in testdrive connections, so $ postgres-execute connection=mz_system (e.g. for ALTER SYSTEM SET) works without a prior registration. Any other connection name must first be registered:

$ postgres-connect name=conn1 url=postgres://materialize:materialize@${testdrive.materialize-sql-addr}

Using an unregistered name fails with connection 'conn1' not found.

Some compositions (e.g. platform-checks, upgrade, pg-cdc multi-version) run the same .td file against multiple Materialize versions. When a change alters version-sensitive output (e.g. the result of a SHOW, a system-catalog query, or the exact text of an error message), gate the affected lines with version guards so each version asserts on the output it actually produces:

>[version<2602300] SHOW some_setting
old_default

>[version>=2602300] SHOW some_setting
new_default

![version<2602300] SELECT * FROM t1_update;
contains:Invalid data in source, saw retractions

![version>=2602300] SELECT * FROM t1_update;
contains:negative record multiplicity

The [version...] guard goes immediately after the directive sigil (> for a query expecting success, ! for one expecting an error, etc.) and before the SQL. The version is encoded as MMmmmpp (e.g. 26.23.0 → 2602300); see src/build-info/src/lib.rs::version_num. Existing examples: test/pg-cdc-old-syntax/privileges.td, test/pg-cdc/publication-with-publish-option.td.

pgtest

Run pgtest files with:

bin/cargo-test -p mz-environmentd pgtest

Run a specific pgtest file: bin/cargo-test -p mz-environmentd test_pgtest_FILE_NAME.

mzcompose

NAME is any directory containing an mzcompose.py file.

Run mzcompose compositions with:

bin/mzcompose --find NAME run WORKFLOW

default is a valid workflow in many compositions, provided a workflow_default exists that iterates over all workflows. Not all compositions have a workflow_default.

Other useful commands:

• bin/mzcompose --find NAME list — list available compositions
• bin/mzcompose --find NAME list-workflows — list workflows for a composition
• bin/mzcompose --find NAME logs — view container logs
• bin/mzcompose --find NAME down — stop a composition

Many tests expect to start with fresh state. Run bin/mzcompose --find NAME down between test runs, not just at the end of a session.

Instead of manually building with cargo build, use bin/mzcompose directly. It implicitly builds in the target-xcompile directory when no prebuilt image is available from Docker Hub/GHCR, so a manual build beforehand is wasted work that mzcompose never uses.

mzcompose builds optimized binaries by default, which are nearly as fast to run as release and much faster to link than debug builds.

• --release builds full release binaries, which can be slow. Instead of building release locally, let CI build the binary by creating a (draft) pull request and waiting for the build jobs to finish. Afterward, mzcompose will attempt to download the binaries instead of building them locally.
• --dev builds non-optimized debug binaries locally.

Reproducing bugs

Reproducing bugs is a manual process. Ask the user for reproduction instructions, or point them to the GitHub issue or the Buildkite annotation under "Test details & reproducer".

Many tests expect fresh state. Always run bin/mzcompose --find NAME down before attempting to reproduce.

For flaky tests, the user can run in a loop until failure:

while true; do
    bin/mzcompose --find testdrive down && \
    bin/mzcompose --find testdrive run default FILENAME.td || break
done

If a flake cannot be reproduced locally, suggest the user trigger CI runs manually via https://ci.dev.materialize.com/trigger Removing the "Target Branch" generates a random identifier, allowing parallel runs. 10-20 runs to reproduce a flake is fine.

To check whether a failure is already known and chronic in CI, see the mz-debug-ci skill's bin/ci-failures search.

Additional logging for flaky tests

Add logging through log_filter in the test's mzcompose.py:

Materialized(
    additional_system_parameter_defaults={
        # TODO: Remove when SQL-NNN is fixed
        "log_filter": "mz_storage::source::postgres=trace"
    },
)

If the flake might occur anywhere, extend get_minimal_system_parameters in misc/python/materialize/mzcompose/__init__.py with a log_filter. Always add a TODO comment referencing the issue.

Adding a test

Determine the right framework based on what you're testing:

• SQL correctness, types, functions (no external systems, no concurrency): sqllogictest (.slt in test/sqllogictest/). Use mode cockroach, test NULLs and edge cases. Do NOT modify files in test/sqllogictest/sqlite or test/sqllogictest/cockroach (upstream). Do NOT drive data-dependent assertions with a LOAD GENERATOR COUNTER source plus mz_unsafe.mz_sleep(...) to wait for ingestion: the counter emits rows over wall-clock time, so the check races ingestion and flakes in CI. Use a plain CREATE TABLE with deterministic INSERTs. When a statically-monotonic operator needs a FROM SOURCE load generator (whose row timing is nondeterministic), split coverage: test the plan shape with EXPLAIN PHYSICAL PLAN over the FROM SOURCE table (no data, non-flaky), and test runtime row correctness with a one-shot SELECT over a plain CREATE TABLE + INSERT.
• Sources/sinks, Kafka, catalog, pgwire (external systems): testdrive (.td in test/testdrive/). Frameworks like pg-cdc, mysql-cdc, sql-server-cdc, kafka-* have targeted setups but also run testdrive files.
• Raw pgwire messages (COPY, extended protocol): pgtest (.pt in test/pgtest/). Wire new files in src/environmentd/tests/pgwire.rs.
• Pure logic, decoding, pure functions: Rust #[mz_ore::test] (or #[mz_ore::test(tokio::test)] for async).
• Restart/upgrade issues: Platform Checks. See doc/developer/platform-checks.md, especially the "Writing a Check" section. Do not use the old Legacy Upgrade tests.
• Panics or unexpected query errors under concurrency: extend Parallel Workload actions in misc/python/materialize/parallel_workload/action.py.
• Many objects / limits: add a Generator subclass in test/limits/mzcompose.py.
• OOM / bounded memory: add a Scenario in test/bounded-memory/mzcompose.py.
• Performance micro-benchmarks: Feature Benchmark scenarios in misc/python/materialize/feature_benchmark/scenarios. See doc/developer/feature-benchmark.md.

For functional issues, aim for at least two different test frameworks that can independently detect the regression.

Read doc/developer/guide-testing.md for more detail on test frameworks.

Extend an existing file, do not create a new one

It is preferred to extend an existing mzcompose-based test, a .td file or .slt file when appropriate. Write the smallest test that fails without the fix.

A panic is caught by CI automatically. Just run the statement that panics. Do not add an assertion on the panic message.

Prove a regression test is red before you call it done

A regression test that has never failed proves nothing. Whenever the test is meant to demonstrate a bug, verify both directions before reporting:

1. With the fix removed run the test and confirm it fails, for the expected reason. Read the failure output.
2. Restore the fix, run the test again, and confirm it passes.