Multi-Stage Pipelines

Real production workflows are sequences of stages where each stage is an independent save() that produces a versioned dataset. The next stage reads from the previous one with read_dataset(). Each stage is independently checkpointed, versionable, and resumable.

Stage Boundaries

A save() followed by read_dataset() creates a stage boundary:

import datachain as dc

# Stage 1: Extract chunks from PDFs
dc.read_storage("s3://docs/*.pdf").gen(chunk=split_pdf).save("chunks")

# Stage 2: Generate embeddings
dc.read_dataset("chunks").setup(model=lambda: load_embedding_model()).settings(
    parallel=8
).map(emb=embed_chunk).save("chunk_embeddings")

# Stage 3: Classify with LLM
dc.read_dataset("chunk_embeddings").setup(client=lambda: create_llm_client()).settings(
    parallel=4
).map(category=classify).save("classified_chunks")

If Stage 2 fails at 80% completion, checkpoints preserve completed rows. Fix the bug, re-run, and DataChain resumes from where it left off. Stage 1 is skipped entirely because its chain hash matches.

Comparative Evaluation

Run two models on the same dataset, then merge the results:

import datachain as dc

# Run two models
dc.read_dataset("chunk_embeddings").setup(client=lambda: model_a_client()).map(
    response_a=run_model_a
).save("model_a_results")

dc.read_dataset("chunk_embeddings").setup(client=lambda: model_b_client()).map(
    response_b=run_model_b
).save("model_b_results")

# Merge and compare
a = dc.read_dataset("model_a_results")
b = dc.read_dataset("model_b_results")
comparison = a.merge(b, on="chunk.text")
comparison.save("model_comparison")

Each model run is a versioned dataset with full provenance. The comparison dataset tracks which inputs and model versions produced which results.

Cost Tracking

Each stage's LLM cost is visible through aggregate analytics on nested token fields, with no deserialization and no Python runtime:

import datachain as dc

chain = dc.read_dataset("model_a_results")
cost = (
    chain.sum("response_a.usage.prompt_tokens") * 0.000002
    + chain.sum("response_a.usage.completion_tokens") * 0.000006
)
print(f"Spent ${cost:.2f}")

Why Stages Matter

Single queries work for simple workflows. Multi-stage pipelines build production systems where each step can be independently monitored, rerun, and compared across versions. The pattern: save() to checkpoint, read_dataset() to continue. Each stage deposits a versioned conclusion that the next stage builds on.