Array Record Source

Data source for ArrayRecord format (efficient columnar storage).

See Also

• Sources Overview - All data sources
• Data Sources Guide - Full guide
• TFDS Source - TensorFlow Datasets
• Performance Tools - Optimization

---

datarax.sources.array_record_source

Data source for reading from ArrayRecord format files.

logger module-attribute

logger = getLogger(__name__)

ArrayRecordSourceConfig dataclass

ArrayRecordSourceConfig(cacheable: bool = False, batch_stats_fn: Callable | Module | None = None, precomputed_stats: dict[str, Any] | None = None, stochastic: bool = False, stream_name: str | None = None, seed: int = 42, num_epochs: int = -1, shuffle_files: bool = False, local_files_only: bool = False)

Bases: StructuralConfig

Configuration for ArrayRecordSourceModule.

Inherits from StructuralConfig for runtime immutability.

Attributes:

Name	Type	Description
seed	int	Random seed for shuffling (used internally, not by Grain).
num_epochs	int	Number of epochs (-1 for infinite).
shuffle_files	bool	Whether to shuffle file order (handled internally).
local_files_only	bool	If True, validate every path exists at construction time and raise FileNotFoundError with path context if any are missing. ArrayRecord sources never download, so this flag is primarily a UX improvement over Grain's lower-level errors.

seed class-attribute instance-attribute

seed: int = 42

num_epochs class-attribute instance-attribute

num_epochs: int = -1

shuffle_files class-attribute instance-attribute

shuffle_files: bool = False

local_files_only class-attribute instance-attribute

local_files_only: bool = False

cacheable class-attribute instance-attribute

cacheable: bool = False

batch_stats_fn class-attribute instance-attribute

batch_stats_fn: Callable | Module | None = None

precomputed_stats class-attribute instance-attribute

precomputed_stats: dict[str, Any] | None = None

stochastic class-attribute instance-attribute

stochastic: bool = False

stream_name class-attribute instance-attribute

stream_name: str | None = None

ArrayRecordSourceModule

ArrayRecordSourceModule(config: ArrayRecordSourceConfig, paths: str | list[str], *, rngs: Rngs | None = None, name: str | None = None)

Bases: DataSourceModule

Stateful wrapper for Grain's ArrayRecordDataSource.

This module wraps Grain's ArrayRecordDataSource while maintaining stateful iteration through NNX Variables, following TDD principles and critical technical guidelines.

Note: Grain's ArrayRecordDataSource doesn't accept a seed parameter directly. Shuffling is handled at the sampler level or through file ordering.

Parameters:

Name	Type	Description	Default
config	ArrayRecordSourceConfig	Configuration for the source.	required
paths	str | list[str]	Path pattern or list of paths to ArrayRecord files.	required
rngs	Rngs | None	NNX Rngs for additional randomness.	None
name	str | None	Optional name for the module.	None

config instance-attribute

config: ArrayRecordSourceConfig

grain_source instance-attribute

grain_source = ArrayRecordDataSource(paths=paths)

current_index instance-attribute

current_index = Variable(0)

current_epoch instance-attribute

current_epoch = Variable(0)

total_records instance-attribute

total_records = Variable(len(grain_source))

prefetch_cache instance-attribute

prefetch_cache: Variable[dict[str, Any]] = Variable({})

iterator_initialized instance-attribute

iterator_initialized = Variable(False)

shuffled_indices instance-attribute

shuffled_indices: Variable[ndarray | None] = Variable(None)

rngs instance-attribute

rngs = rngs

name instance-attribute

name = static(name)

stochastic instance-attribute

stochastic = stochastic

stream_name instance-attribute

stream_name = stream_name

get_state

get_state() -> dict[str, Any]

Get complete state for checkpointing.

set_state

set_state(state: dict[str, Any]) -> None

Restore state from checkpoint.

get_batch_at

get_batch_at(start: int, size: int, key: Any | None = None) -> list[Any]

Stateless indexed batch access for Pipeline-driven iteration.

Returns size records starting at logical position start, wrapping at the end of the dataset and applying any active shuffle permutation. Does not advance self.current_index or any other internal state.

ArrayRecord records are loaded host-side (Grain is a Python library), so this method requires a concrete Python int for start. Driving an ArrayRecord source under nnx.scan (Tier C of the pipeline integration story) currently requires wrapping the host-side fetch in jax.experimental.io_callback — left as a future enhancement. Tier A (Python iteration) and Tier B (single step()) work today.

Parameters:

Name	Type	Description	Default
start	int	Concrete starting index (Python int).	required
size	int	Number of records to return.	required
key	Any | None	Reserved for future shuffled-mode support; currently ignored (shuffle uses self.shuffled_indices).	None

Returns:

Type	Description
list[Any]	List of size records as returned by the underlying
list[Any]	Grain source. Records are typically Python dicts; callers
list[Any]	(typically a parse / decode operator) handle structure.

Raises:

Type	Description
TypeError	If start is a JAX tracer (not host-side concrete). ArrayRecord cannot be traced through nnx.scan without an io_callback wrapper.

get_operation_stats

get_operation_stats() -> dict[str, int]

Get operation statistics.

Note: This method converts JAX arrays to Python ints for introspection. It is intended for use outside of JIT-compiled functions.

Returns:

Type	Description
dict[str, int]	Dictionary with 'applied_count' and 'skipped_count'

reset_operation_stats

reset_operation_stats() -> None

Reset operation statistics to zero.

Note: Creates new JAX arrays to reset the counters.

compute_statistics

compute_statistics(data: Any) -> dict[str, Any] | None

Compute statistics from data using batch_stats_fn.

If batch_stats_fn is not configured, returns None. Computed statistics are cached in _computed_stats.

Parameters:

Name	Type	Description	Default
data	Any	Input data to compute statistics from	required

Returns:

Type	Description
dict[str, Any] | None	Dictionary of statistics, or None if no batch_stats_fn configured

get_statistics

get_statistics() -> dict[str, Any] | None

Get current statistics.

Returns precomputed_stats if configured (unless reset was called), otherwise returns cached computed statistics, or None if no statistics available.

Returns:

Type	Description
dict[str, Any] | None	Dictionary of statistics, or None if no statistics available

set_statistics

set_statistics(stats: dict[str, Any]) -> None

Manually set statistics.

This overwrites any previously computed statistics and clears reset flag.

Parameters:

Name	Type	Description	Default
stats	dict[str, Any]	Dictionary of statistics to set	required

reset_statistics

reset_statistics() -> None

Reset all statistics to None.

This clears both computed statistics and marks that precomputed_stats should be ignored (via internal flag). After reset, get_statistics() will return None until new statistics are set or computed.

reset_cache

reset_cache() -> None

Clear the cache.

Only has effect if cacheable=True in config.

copy

copy(*, config: DataraxModuleConfig | None = None, rngs: Rngs | None = None, name: str | None = None) -> DataraxModule

Create a copy of this module with optional config/parameter changes.

This allows creating a new module instance with modified configuration while preserving other attributes. Useful for hyperparameter tuning.

Parameters:

Name	Type	Description	Default
config	DataraxModuleConfig | None	New config (if None, uses current config)	None
rngs	Rngs | None	New RNG state (if None, uses current rngs)	None
name	str | None	New name (if None, uses current name)	None

Returns:

Type	Description
DataraxModule	New module instance with updated parameters

Examples:

Change configuration

new_config = DataraxModuleConfig(cacheable=True) new_module = module.copy(config=new_config)

Change name only

renamed = module.copy(name="new_name")

Note

Subclasses can override this method to provide more fine-grained control over copying, such as allowing individual config field updates without requiring dataclass replace().

clone

clone() -> DataraxModule

Create a new instance with the same state as this module.

Uses NNX's clone function for proper deep cloning of all state.

Returns:

Type	Description
DataraxModule	A new module instance with the same state.

requires_rng_streams

requires_rng_streams() -> list[str] | None

Get the list of RNG streams required by this module.

Returns:

Type	Description
list[str] | None	A list of required RNG stream names, or None if no RNG streams
list[str] | None	are required.

ensure_rng_streams

ensure_rng_streams(stream_names: list[str]) -> None

Ensure that the required RNG streams are available.

Parameters:

Name	Type	Description	Default
stream_names	list[str]	A list of available RNG stream names.	required

Raises:

Type	Description
ValueError	If a required RNG stream is not available.

process

process(input: Any, *args: Any, **kwargs: Any) -> Any

Process input structure.

This method transforms the structure/organization of input data without modifying the data values themselves.

Subclasses MUST implement this method.

The input/output types depend on the specific structural processor:

• Batcher: list[Element] -> list[Batch]
• Sampler: int -> list[int]
• Sharder: Batch -> Sharded[Batch]
• Splitter: Dataset -> tuple[Dataset, Dataset]

Parameters:

Name	Type	Description	Default
input	Any	Input to process (type varies by processor)	required
*args	Any	Additional positional arguments (processor-specific)	()
**kwargs	Any	Additional keyword arguments (processor-specific)	{}

Returns:

Type	Description
Any	Processed output (type varies by processor)

Examples:

Batcher implementation:

def process(self, elements: list[Element]) -> list[Batch]:
    batches = []
    for i in range(0, len(elements), self.config.batch_size):
        batch_elements = elements[i:i + self.config.batch_size]
        batches.append(Batch.from_elements(batch_elements))
    return batches

Sampler implementation (deterministic):

def process(self, dataset_size: int) -> list[int]:
    return list(range(min(self.config.num_samples, dataset_size)))

Sampler implementation (stochastic):

def process(self, dataset_size: int) -> list[int]:
    rng = self.rngs[self.config.stream_name]()
    indices = jax.random.choice(
        rng, dataset_size, shape=(self.config.num_samples,),
        replace=self.config.replacement
    )
    return indices.tolist()

element_spec

element_spec() -> Any

Return a PyTree of jax.ShapeDtypeStruct describing per-element output.

Downstream consumers (operators, batchers, models) use this contract to pre-allocate buffers, auto-size learnable layers, and statically validate operator chains. Subclasses MUST override this method.

Returns:

Type	Description
Any	A PyTree (typically a dict) whose leaves are jax.ShapeDtypeStruct
Any	instances describing one emitted element.

Raises:

Type	Description
NotImplementedError	Always, on the base class. Subclasses must override.