ML Data Pipeline is the system that efficiently loads, preprocesses, and batches training data — a bottleneck that can reduce GPU utilization from 100% to < 30% if poorly implemented, making data loading optimization as important as model architecture.

The I/O Bottleneck Problem

• GPU throughput: Processes a batch in 50ms.
• Naive data loading: Read from disk + decode + augment = 200ms per batch.
• Result: GPU idle 75% of the time — $3,000/month GPU cluster at 25% utilization.
• Solution: Overlap data preparation with GPU compute using prefetching and parallel loading.

PyTorch DataLoader

dataloader = DataLoader(
    dataset,
    batch_size=256,
    num_workers=8,          # Parallel CPU workers
    prefetch_factor=2,      # Batches to prefetch per worker
    pin_memory=True,        # Pinned memory for fast GPU transfer
    persistent_workers=True # Avoid worker restart overhead
)

• num_workers: Spawn N CPU processes for parallel loading. Rule of thumb: 4× number of GPUs.
• prefetch_factor: Each worker prefetches factor× batches ahead.
• pin_memory=True: Required for async GPU transfer.

TensorFlow tf.data Pipeline

dataset = tf.data.Dataset.from_tensor_slices(filenames)
dataset = dataset.interleave(tf.data.TFRecordDataset, num_parallel_calls=8)
dataset = dataset.map(preprocess, num_parallel_calls=tf.data.AUTOTUNE)
dataset = dataset.batch(256)
dataset = dataset.prefetch(tf.data.AUTOTUNE)  # Overlap GPU compute with CPU prep

Storage Optimization

• TFRecord / WebDataset: Sequential binary format → faster disk reads than random file access.
• LMDB: Memory-mapped key-value store — near-RAM speeds for small datasets.
• Petastorm: Distributed dataset format for Spark + PyTorch/TF.

Online Augmentation

• Apply augmentations (crop, flip, color jitter) on CPU workers during loading — free compute.
• GPU augmentation (NVIDIA DALI): Move decode and augment to GPU — further reduces CPU bottleneck.

Efficient data pipeline design is a critical ML engineering skill — well-tuned data loading routinely improves training throughput 2-5x with no changes to model architecture, directly reducing the cost and time of every training run.