Day 88: the join, on the column you asked for

Date: 2026-10-20 · Week 14 · Phase 5 Benchmark and optimize

What I added today

Day 87 ended a booking’s --table with the sweeps joined by load, and the join printed one metric, itl_p50_speedup. The median is not where graphs earn their keep under load. A replay removes the launch stalls that make the tail, so the p99 is the column a reader comparing days wants next, in the same shape.

--join COLUMN picks the column the block joins on:

--table d74.csv d79.csv --join itl_p99_speedup

itl_p99_speedup, joined by load:
      load  sweep 1  sweep 2
----------------------------
     1 rps    2.00x    3.25x
     2 rps        -    1.75x
sweep 1: d74.csv
sweep 2: d79.csv
  • Three columns, from JOIN_COLUMNS: itl_p50_speedup (the default), itl_p99_speedup, and split_itl_p50_speedup. Ratios only. A latency in ms moves with the card and the day; a ratio was measured against its own control.
  • The default is Day 87’s block, down to the line. --join itl_p50_speedup and no --join print the same stdout, and a test says so.
  • The heading names the column, so a p50 block and a p99 block pasted side by side cannot be mistaken for each other.
  • A name off the list is argparse’s refusal (exit 2, invalid choice), before any file is read.
  • A column some sweep does not have is refused before anything prints, naming the path that lacks it (check_join_column). That is the whole of the “split column only where every sweep was a split boot” rule: a two-arm sweep has no split_itl_p50_speedup, and nothing else has to special-case it.
  • --join without --table, or with one CSV, is refused. One sweep prints no block, so the option would do nothing, silently.
  • A rerun or ? cell still prints its mark, in every column.

reprint_sweeps takes join= and checks the column next to Day 87’s duplicate-load check, in the same try, so every refusal is still “a sentence on stderr, nothing on stdout”. print_join and _join_cell take the column, defaulting to the p50.

tests/test_table_join_column.py has 11 tests: the list, the default being Day 87’s block, the p99 block’s lines with a gap, the split column on split boots and its refusal naming only the two-arm path, an off-list name through the CLI and through reprint_sweeps, --join without --table, --join with one CSV, the rerun mark in the p99 column, and reprint_sweeps(join=...). Day 87’s 8 tests pass unchanged.

Suite 2531 green (34 GPU-gated skips), ruff clean.

Why it matters

A booking is meant to answer more than one question. Whether the median moved and whether the tail moved are different claims, and Day 74’s gates check them separately. A join that could only show one of them made the other a scroll back up through five tables, which is the problem Day 87 set out to remove.

A hole has to mean one thing. In the join, - means “this sweep did not run this load”. If a two-arm sweep printed - under the split column, the same mark would mean “this sweep could not have measured this at all”, and a reader would go looking for a missing load that was never missing. Refusing the column keeps - honest.

What I learned

  1. Three tests passed before I wrote any code. --join was an unknown argument, so argparse exited 2 with --join in its usage line, which is exactly what my “refused” tests asserted. They were green for the wrong reason. I tightened each to the sentence the code actually says (invalid choice, one sweep prints no join, --join chooses the column), so they now fail without the feature.
  2. The header is the schema. read_sweep gives every row the file’s header as keys, so “does this sweep have the split column” is column in rows[0], with no flag recording which arms booted. The file already said it.
  3. A default you can name is a default you can test. Making itl_p50_speedup the first of JOIN_COLUMNS and comparing the two stdouts byte for byte is a better regression check than re-asserting Day 87’s lines a second time.

Diagram

join-column.png. Left: the three columns on offer and the refusals that happen before anything prints. Right: a p99 join with a gap and a rerun mark, and the heading that names its column.

Tomorrow

A reader with a booking now runs --table up to three times to see all three columns. Next: --join taking more than one column (--join itl_p50_speedup itl_p99_speedup), printing one block per column in the order given, with each column checked against every sweep before the first table prints, and a column named twice refused like a path given twice.

The card run still comes first when hardware is booked.

Post angle

Day 88 of building an LLM inference engine from scratch. Yesterday’s join showed the p50 speedup across a booking’s sweeps, but CUDA graphs earn their keep in the tail. --join itl_p99_speedup now picks the column, and the split column is refused (by path) if any sweep was a two-arm boot, because a - there would lie about why the cell is empty. 2531 green.