TFX 1.21.0 Release

Major Features and Improvements

• Added dynamic support for ZetaSQL-free MLMD environments across TFX Resolvers and metadata extensions. The system automatically detects missing C++ ZetaSQL engine binaries at runtime and transparently falls back to a highly robust, pure-Python in-memory lineage graph traversal and relation evaluation engine.

Breaking Changes

• Transitioned proto compilation tooling in Bazel workspaces from legacy deprecated py_proto_library rules to custom Starlark provider compilation macros, enabling unified, robust build integration on Bazel 7.x workspaces running with Bzlmod enabled.

For Pipeline Authors

• N/A

For Component Authors

• N/A

Deprecations

• Bypassed legacy testing targets checking deprecated and retired Google Cloud AI Platform (CAIP) integration points, fully migrating Vertex AI-compatible pipeline targets.

Bug Fixes and Other Changes

• Refactored Wide & Deep functional models (taxi_utils.py, templates, and test modules) to slice wide categorical input layers dynamically matching actually wide-encoded category bounds ([:len(_MAX_CATEGORICAL_FEATURE_VALUES)]). This prevents disconnected inputs from triggering Keras 3 inputs not connected to outputs exception under Python 3.10.
• Converted Keras Functional model building methods' Normalization layer instantiation inside list comprehensions to standard procedural for loops, fully securing execution scope connectivity tracking under Python 3.10.
• Implemented dynamic pytest_ignore_collect hooks in conftest.py with static spec checks (importlib.util.find_spec) to dynamically exclude targets of uninstalled optional dependencies (like Airflow, Vertex AI, and Kubeflow). This completely eliminates early logging stream deadlocks and startup import-time test suite collection crashes.
• Upgraded Docker build tools and wheel scripts, configuring internal compilation of TFDV and TFX-BSL source files on a unified conda-GCC 13/binutils toolchain using Bazel 7.7.0.
• Resolved random temporary directory synchronization and write finalizer errors in BulkInferrer (executor.py) when executing flattened PCollections under local runners (DirectRunner/PrismRunner/FnApiRunner) by introducing a dynamic helper mapping local executions to use num_shards=1 while preserving high-performance dynamic sharding for distributed production pipelines.
• Bypassed strict committed/attempted metrics equivalence checks in the Transform ExecutorTest base class (executor_test.py) that crashed under modern versions of Apache Beam utilizing the parallel/multi-process PrismRunner backend due to asynchronous task metric updating limits, ensuring robust and stable local metrics count verifications.
• Monkey-patched PipelineOptions dynamically in the global test conftest (conftest.py) to bypass resource-throttled multi-process PrismRunner delegation for standard local testing jobs, forcing the low-overhead, fast single-threaded in-memory DirectRunner (--direct_running_mode=in_memory) globally. This slashes total unit testing execution time and prevents workflow cancellations/timeouts across Python 3.9, 3.10, 3.11, and 3.12 GHA platforms.

Dependency Updates

• Upgrades target pipeline constraints to support TensorFlow 2.21.0 and Protobuf 6.x across both Python 3.10, 3.11, 3.12 ande 3.13.
• Split SciPy library dependency constraint inside test_constraints.txt using Python target markers to bypass dynamic version conflicts with JAX versions under Python < 3.13.
• Cleanly dropped outdated/incompatible dependencies (tensorflow-decision-forests, tensorflow-ranking, tensorflow-text, tensorflowjs) globally from dependencies list and constraint definitions to prevent PIP backtracking solver storms and secure stable installation on TF 2.21.0.