แคชและข้ามขั้นตอนที่ไม่เปลี่ยนแปลง

If you only tweak the eval script, re-running ingest and train wastes time. DVC caches stage results so unchanged work is skipped. ⏱️

For each stage, DVC computes a hash of its dependencies and command. The hash is a fingerprint that changes the moment any input changes.

When you run dvc repro, DVC compares each stage's current hash to the one in dvc.lock. If they match, it skips the stage.

DVC tells you plainly when nothing changed. A message like didn't change, skipping means the cached output was reused.

Change a prep input and DVC re-runs prep, then every downstream stage that depends on it. Upstream stages stay cached.

Stage outputs are saved in DVC's cache directory, keyed by hash. Skipping a stage means DVC just restores that cached file.

Need to ignore the cache? Pass --force to dvc repro and every stage runs again, even if nothing changed.

You can target one stage by name with dvc repro -s. Handy when you want to rerun just train without touching the rest.

Run dvc status to see which stages are out of date without executing anything. It is a safe dry run of what repro would do.

On big datasets, caching turns a thirty-minute pipeline into seconds when only one stage changed. That fast feedback loop is the whole point. ⚡

Because dvc.lock holds the hashes, committing it lets teammates reuse the same cache decisions and skip work you already did.

Think about what triggers a re-run.

You saw how DVC hashes inputs to skip unchanged stages, cascade changes downstream, and force reruns when needed. Fast, smart pipelines. 🎯