Data Join

Steps

1. Index both files: Run mcp__qsv__qsv_index on both files for fast random access.

2. Inspect both files: Run mcp__qsv__qsv_headers on both files to identify column names. Determine which columns to join on.

3. Profile join columns: Run mcp__qsv__qsv_stats with cardinality: true, stats_jsonl: true on both files. Check the cardinality of join columns to determine optimal table order.

4. Choose strategy:

   • If cardinality of join column in file1 > file2, put file1 on the left
   • For joinp: smaller cardinality table should be on the right for best performance
   • If join condition is complex (non-equi), use mcp__qsv__qsv_sqlp
   • If join involves date/time matching where exact dates won't align (e.g., quarterly to monthly, event dates to nearest reporting period), use mcp__qsv__qsv_joinp with asof: true

5. Execute join: Use mcp__qsv__qsv_joinp for standard joins:

   joinp
     columns1: "id"
     input1: "file1.csv"
     columns2: "id"
     input2: "file2.csv"
     # Join type: omit for inner (default), or set one of:
     # left: true, full: true, cross: true
   

   Or use mcp__qsv__qsv_sqlp for complex joins:

   SELECT a.*, b.col1, b.col2
   FROM file1 a
   JOIN file2 b ON a.id = b.id AND a.date BETWEEN b.start_date AND b.end_date
   

   For ASOF (nearest-match) joins, use mcp__qsv__qsv_joinp with asof: true:

   joinp
     columns1: "date"
     input1: "events.csv"
     columns2: "date"
     input2: "reference.csv"
     asof: true
     strategy: "backward"
     allow_exact_matches: true
   

   • strategy: "backward" (default) — match to the last right row with key < left key
   • strategy: "forward" — match to the first right row with key > left key
   • strategy: "nearest" — match to the numerically closest row (supports tolerance parameter)
   • Add left_by/right_by parameters to restrict matching within subgroups (e.g., per jurisdiction)
   • Add allow_exact_matches: true to include equal keys (<=, >=); default is strict inequality (<, >)

6. Clean up result: Use mcp__qsv__qsv_select to remove duplicate join columns or unnecessary columns from the result.

7. Verify: Run mcp__qsv__qsv_count on the result. Compare with input counts to validate join behavior:

   • Inner join: result <= min(left, right)
   • Left join: result >= left count
   • Full outer: result >= max(left, right)
   • ASOF: result = left count (every left row gets a match or null, like a left join)

Join Column Validation Checklist

Before executing a join, read .stats.csv for both files and validate:

Join Types

Notes

• joinp uses the Polars engine and is significantly faster than join for large files
• The stats cache helps joinp optimize join execution
• For joining on multiple columns, separate column names with commas: columns1: "col1,col2"
• Column names must match exactly (case-sensitive)
• If join columns have different names, specify separately: columns1: "id", columns2: "customer_id"
• For one-to-many joins, the result will have more rows than either input
• joinp handles null values in join columns (nulls don't match by default)
• ASOF joins implicitly enable --try-parsedates — no need to pass it explicitly
• For ASOF joins with subgroups, use --left_by and --right_by (e.g., match nearest date per jurisdiction)
• The --tolerance option (nearest strategy only) limits how far the nearest match can be: use duration strings for dates (1d, 30d, 365d) or positive integers for numeric keys
• ASOF joins require sorted join columns; both datasets are auto-sorted unless --no-sort is set