Data Import & Validation (Streaming + Error CSV + Metrics + Idempotency)

When to use this skill

Use when working on:

• ETL / import pipelines (CSV/XLSX/JSON) into a DB
• Parsers that currently load whole files into memory
• Data validation/coercion, error isolation, auditability, and reproducibility
• Any import job that must be safe to re-run (idempotent)

Non-negotiables (contract)

1. Process input files sequentially (no “load everything then insert”) to control memory.
2. Validate and coerce types explicitly (define allowed coercions; reject ambiguous cases).
3. Irreparable records must be skipped and logged to an error CSV containing:
   • all original columns exactly as seen in input
   • extra columns: timestamp, file, line, error
4. Emit metrics at minimum: rows_ok, rows_skipped, parse_errors.
5. DB writes must be idempotent: re-running the import must not duplicate or corrupt data.

Standard workflow

1) Plan-first (before coding)

• Identify input formats, volume, and “row identity” rules (keys/dedup strategy).
• Locate current import entrypoints + DB write layer.
• Define:
  • schema mapping (source -> target columns)
  • validation rules per field
  • coercion rules per field (what is allowed, what is not)
  • error taxonomy (what counts as “irreparable”)

2) Streaming architecture

• Read one file at a time, row by row (or chunked) and write in bounded batches.
• Never accumulate full datasets in memory.
• Ensure progress logging is monotonic (e.g., file + row counters).

3) Validation & coercion

• Treat raw row values as immutable “source of truth”.
• Perform coercions in a controlled layer:
  • return (ok, parsed_record) or (error, reason) per row
• Separate concerns:
  • parsing (raw -> typed)
  • business validation (typed -> acceptable)
  • persistence (acceptable -> DB)

4) Error CSV logging

• On any irreparable row:
  • write original row values (unmodified)
  • add: timestamp, file, line, error
• Avoid partial writes that drop context; every skipped row must be explainable from the CSV alone.

5) Metrics

Maintain counters:

• rows_ok: successfully persisted rows
• rows_skipped: irreparable rows skipped
• parse_errors: count of parse/validation errors (can equal rows_skipped or be a superset if you track recoverable warnings separately)

Emit metrics:

• end-of-file summary
• end-of-run summary (aggregate over files) Optionally persist metrics to a JSON or a DB table for observability.

6) Idempotent persistence

Pick and implement one clear strategy:

• Upsert by natural key / business key
• Insert with unique constraint + conflict handling
• Staging table + merge
• Per-file “ingestion ledger” (hash/checkpoint) + skip already-ingested files

Idempotency must hold across:

• reruns after crashes
• partial completion
• duplicate input files

Review checklist (definition of done)

• Demonstrably streaming: memory does not scale with total rows.
• Every skipped record is present in error CSV with full context.
• Metrics are emitted and consistent with observed behavior.
• Re-running import on same inputs does not create duplicates (verified).
• Failures are isolated: one bad row/file does not poison the full run.