Divergent Data Pipe

Divergent Data Pipe

Bridge an upstream source and a downstream-project dataset that has diverged from it, then keep them in sync incrementally.

Scenario

• Upstream: a canonical data file (CSV/Parquet/JSON/API) that keeps receiving new rows and the occasional row update.
• Downstream: a project-owned dataset derived from upstream but enriched, renamed, re-typed, restructured. It cannot be rebuilt from upstream alone — the project investments would be lost.
• Problem: naively re-loading upstream would either overwrite the downstream or miss the enrichments. A hand-rolled incremental load is fragile.

This skill's job is to (1) map upstream↔downstream and (2) produce a repeatable script that loads only new/changed upstream rows into downstream while preserving the divergence.

When to invoke

• User says "the source file got new rows, how do I merge them into the project?", "my dataset has drifted from the source", "I need an incremental sync".

Procedure

Phase 1 — Diff and map

1. Locate both sides — upstream file/endpoint and downstream dataset.
2. Profile each independently — columns, types, row counts, primary/natural keys, last-modified column if any.
3. Column mapping:

• Auto-suggest by name similarity, sample-value match, and type compatibility.
• Confirm each mapping with the user: upstream.col_x → downstream.col_y or upstream.col_x → (dropped) or downstream.col_z → (added by project, not in upstream).
• Record renames, splits (one upstream column → multiple downstream), merges, and type casts explicitly.

1. Transformation inventory — for each upstream-to-downstream difference, record the transformation the downstream applied (ISO code enrichment, currency conversion, text parsing, text-to-numeric, categorical remapping). Route to plugin skills where applicable.
2. Identify the key — the upstream↔downstream row identity. Usually a stable upstream primary key preserved verbatim downstream. If there isn't one, build a composite or deterministic hash from stable columns; warn that lack of a key makes updates impossible (append-only).
3. Identify the change-detection signal:

• New rows: upstream key not in downstream.
• Updated rows: same key, differing upstream payload vs the downstream's _source_* shadow columns (see below).
• Deleted rows: key in downstream but not in upstream — soft-delete by default; confirm policy.

1. Write a mapping document divergent_pipe_mapping.md — columns, transformations, key, change-detection strategy, collision rules.

Phase 2 — Shadow columns (one-time augmentation)

To detect updates to rows that were already loaded, the downstream needs to remember the upstream state it last saw. Add shadow columns alongside the enriched downstream fields:

• _source_ingested_at — timestamp of last successful ingest for this row.
• _source_row_hash — hash of the upstream columns that map into downstream (exclude enrichment-only downstream columns).
• _source_primary_key — the upstream key (if different from downstream PK).

If the downstream is missing these, generate them from the current upstream once (best-effort reconstruction) and flag rows where reconstruction is ambiguous.

Phase 3 — Generate the sync script

Emit sync_from_upstream.py (or .sh wrapping SQL for DB-backed downstreams). Structure:

1. Read upstream (full or since-last-sync via a watermark).
2. Compute the row hash per upstream row.
3. Left-join against downstream on the upstream key:
   - no match         -> new row: apply transformation, insert.
   - match, same hash -> skip.
   - match, diff hash -> update: apply transformation to *upstream fields only*,
                         preserve downstream enrichments and manual edits.
4. For keys present downstream but absent upstream:
   - soft-delete (flag a `_source_deleted_at` column) or skip, per policy.
5. Update _source_row_hash and _source_ingested_at on touched rows.
6. Write a run log: counts of inserted / updated / skipped / soft-deleted / errored.

The script must be:

• Idempotent — re-running without new upstream data does nothing.
• Dry-run capable — --dry-run prints the plan without writing.
• Atomic per batch — wrap in a transaction for SQL-backed downstreams; write to a temp file and rename for file-backed downstreams.
• Backed up first — per CONVENTIONS.md, confirm a downstream backup or create one before first real run.

Phase 4 — First run and verify

1. Run in --dry-run mode and show the user the planned counts.
2. Confirm with the user before applying.
3. Run the real sync; compare before/after row counts, spot-check a few inserted and updated rows, review the run log.
4. Record the run in the downstream's CHANGELOG.md (via add-changelog).
5. Update the data dictionary with the new shadow columns and the sync procedure reference.

Dependencies

pip install pandas pyarrow
# if downstream lives in a SQL DB
pip install SQLAlchemy psycopg[binary]  # or mysql / sqlite etc.

Edge cases

• Downstream mutates upstream-mapped columns by hand — the hash comparison will see this as "upstream changed". Before first run, ask whether downstream edits to source-origin columns should be preserved (then lock those rows) or overwritten on upstream change. Offer a per-column policy.
• Schema evolution in upstream — new columns appear. Flag and ask whether to (a) add to downstream with default enrichment, (b) ignore, (c) require explicit mapping update. Do not auto-propagate schema changes.
• Key changes upstream — a row's key changes (upstream renumbered). Detect via content-similarity; ask user to pair manually. Never silently reassign.
• Soft-delete vs hard-delete — hard-deleting downstream rows risks losing enrichments. Default to soft-delete; require explicit confirmation for hard-delete.
• Multiple upstream sources — out of scope for a single pipe; the user should run this skill once per upstream.

Safety

This skill writes into a dataset that represents real project investment. Follow CONVENTIONS.md rigorously:

• Confirm downstream backup before first real run.
• Dry-run first, always.
• The generated sync script must keep the dry-run and backup flags on by default.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: danielrosehill
• Source: danielrosehill/Claude-Data-Wrangler-plugin
• License: MIT

Install and usage instructions live in the source repository linked above.