Fail-Fast ML Engineering

When to Use

• Designing error handling for training pipelines
• Reviewing code for silent failure modes
• Setting up preflight validation before expensive operations
• Deciding where to put runtime constants (config vs code vs checkpoint)

Core Principles

No Silent Fallbacks

Critical paths must not silently fall back to defaults. If a required resource is missing, raise an explicit error immediately:

# Bad: silent fallback
def load_encoder(path=None):
    if path and Path(path).exists():
        return torch.load(path)
    return RandomEncoder()  # Silently trains garbage

# Good: explicit failure
def load_encoder(path: str):
    if not Path(path).exists():
        raise FileNotFoundError(f"Encoder weights not found: {path}")
    return torch.load(path)

Config as Single Source of Truth

Runtime behavior is determined by config files, not checkpoint metadata, environment guessing, or hardcoded defaults:

# Bad: trusting checkpoint metadata
config = checkpoint["config"]  # May be stale, incomplete, or from different code version

# Good: config file is authoritative
config = OmegaConf.load("conf/training/fullrun.yaml")
# Checkpoint metadata is diagnostic only
checkpoint_meta = checkpoint.get("config", {})

Preflight Pattern

Validate everything that can fail before starting expensive computation:

def preflight(cfg):
    """Run before any GPU work. Fail fast, fail cheap."""
    # Data exists?
    assert Path(cfg.data.path).exists(), f"Data missing: {cfg.data.path}"

    # Model weights exist?
    if cfg.model.pretrained:
        assert Path(cfg.model.pretrained).exists(), f"Weights missing: {cfg.model.pretrained}"

    # GPU available?
    assert torch.cuda.is_available(), "No GPU detected"

    # Config sanity
    assert cfg.training.batch_size > 0
    assert cfg.training.lr > 0

    print("Preflight OK")

Fail Cheap, Not Expensive

Order operations so failures happen before GPU hours are consumed:

1. Config validation
2. Data existence checks
3. Model weight loading
4. GPU allocation
5. Training loop

Patterns

Explicit Error Messages

Include the failing value and what was expected:

# Bad
raise ValueError("Invalid batch size")

# Good
raise ValueError(f"batch_size must be > 0, got {cfg.training.batch_size}")

Guard Clauses Over Nested Ifs

# Bad: deeply nested
def process(sample):
    if sample is not None:
        if "audio" in sample:
            if sample["audio"].shape[0] > 0:
                return transform(sample["audio"])
    return None  # Silent failure

# Good: guard clauses with explicit errors
def process(sample):
    if sample is None:
        raise ValueError("Sample is None")
    if "audio" not in sample:
        raise KeyError(f"Sample missing 'audio' key. Keys: {list(sample.keys())}")
    if sample["audio"].shape[0] == 0:
        raise ValueError("Audio tensor is empty")
    return transform(sample["audio"])

Assertion-Heavy Data Loading

Data pipelines are the #1 source of silent bugs in ML. Assert aggressively:

def collate_fn(batch):
    audio = torch.stack([s["audio"] for s in batch])
    labels = [s["labels"] for s in batch]

    assert audio.dim() == 3, f"Expected 3D audio tensor, got {audio.dim()}D"
    assert audio.shape[0] == len(labels), f"Batch size mismatch: {audio.shape[0]} vs {len(labels)}"
    assert all(len(l) > 0 for l in labels), "Empty label sequence in batch"

    return {"audio": audio, "labels": labels}

Framework Compatibility Guards

When combining frameworks (e.g., Accelerate + WebDataset), verify compatibility upfront:

# Accelerate's prepare() may fail on custom dataloaders
if isinstance(dataloader.dataset, IterableDataset):
    # Skip accelerator.prepare() for external dataloaders
    # Only prepare model + optimizer
    model, optimizer = accelerator.prepare(model, optimizer)
else:
    model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)

Anti-Patterns

• try: ... except: pass: Never swallow exceptions in training code. Every exception is information.
• Default arguments that hide failures: def load(path=None) where None triggers a fallback is a bug waiting to happen.
• Trusting checkpoint config: Checkpoints store config snapshots for diagnostics. The YAML config files are the authority.
• Late validation: Checking data format inside the training loop wastes GPU time. Validate in preflight.
• Magic numbers: fps = 25 hides the derivation. Use fps = sample_rate / hop_samples so the dependency is explicit.
• Optimistic error handling: Don't catch errors just to log and continue. If the error matters, stop. If it doesn't, don't catch it.

See Also

• hydra-experiment-config — Config as single source of truth
• hf-dataset-management — Preflight verification for datasets
• slurm-gpu-training — Preflight before Slurm job submission