Add Git4Data Part 8: ML Continuous Learning — Train Only What Changed (EN + ZH)

[dengn]

What

Adds Part 8 — ML Continuous Learning: Train Only What Changed (EN + ZH), opening the AI-training arc.

• EN: matrixorigin/git4data-part8-ml-incremental/
• ZH: matrixorigin/git4data-part8-ml-incremental-zh/

The data changes every day — why retrain on everything? Pin the training set with a SNAPSHOT; next round, one DATA BRANCH DIFF gives the exact changed rows, so you train only the delta.

Content (detailed, scenario-driven, strong comparison)

• When you do continuous learning — risk/fraud, recommendation/CTR, moderation, any growing labeled set.
• The three-step loop — pin → train + register → DIFF the delta; a model ↔ data chain that unlocks exact reproduce / attribute / rollback.
• Cost experiment — 6 rounds: incremental 6,004 rows vs full retrain 21,000; the gap grows quadratically.
• "How the alternatives do it — and where they break" — updated_at watermark, CDC/Debezium, two full copies + EXCEPT, DVC / lakeFS, Delta Lake CDF — each with concrete failure modes and a system-by-system table.

Figures (2 per language)

• fig_incremental-loop — the pin → train → DIFF-delta cycle.
• fig_cost-curve — full-retrain (quadratic) vs incremental (linear) over 6 rounds.

Verification

Every statement verified on MatrixOne 4.0.0-rc3: DIFF → INSERTED 3000 / UPDATED 200; net delta (value anti-join) = 3200 vs full table 103000; snapshot reproduces m1's set = 100000; 6-round experiment per-round delta stays 1000 + a couple even as the table grows 6×. pnpm validate → 369/369.

🤖 Generated with Claude Code