[quinnj]

String columns copy every cell longer than the 12-byte inline payload into a buffer the chunk owns. That buffer grew by one resize! per cell, reallocating as it went. With several tasks the reallocation copies and the page faults of each new block dominated the parse: in an 8-task profile of a text-heavy file, 65% of the chunk workers' samples sat on that resize! line.

At its first long cell, the loop now sums the field spans of the column's remaining long cells in the chunk (an upper bound on what it can copy, since content never exceeds its span) and reserves that once. Columns whose cells are all inline-length never run the pass. Values are unchanged; only buffer capacity changes. A side effect: finished columns no longer retain the growth slack.

strings.csv (163 MB, 1.2M rows, 10 text columns), interleaved process pairs on a loaded M3 Max, Julia 1.12.6:

	main	this PR
8 tasks, wall (min of 3 per process)	0.19-0.22 s	0.13-0.15 s
8 tasks, total CPU	0.62-0.71 s	0.36-0.61 s
1 task, peak RSS	959-994 MB	841 MB
retained string-buffer capacity / bytes	146 / 106 MiB	106 / 106 MiB

A mixed file whose text cells are mostly inline-length showed no difference beyond noise.

Review: my own pass, then an independent agent review (values, row coverage, copy bounds, sizehint! on 1.10/1.12, thread ownership, 40 cases on 1.10 and 1.12 compared with main), which found the over-reservation below.

• Every string parse path (direct wave, stale re-parse, masked Scan pass, Chunks batches) goes through parsecolchunk!(::StringColumn, ...), and each passes a column whose buffer is private to the call. The transposed reader calls _own! on its own and is unchanged.
• The reservation pass applies the loop's own row filters (mask, report limit) and counts the raw span of every long cell, which bounds all three copy paths: malformed quotes copy the raw span, escaped cells unescape to at most their content, plain cells copy their content.
• Over-reservation: the pass counts every span longer than the inline payload, and some such cells copy nothing (a quoted 12-byte value, a short escaped value, a long sentinel, a padded cell under stripwhitespace). A second review found that the finished column kept that slack (one million quoted phone numbers retained 13.8 MB for 9,500 bytes; main 9,873). Fixed in the second commit: after the loop the buffer shrinks to what it holds when the slack exceeds half its size and 64 KiB. Same input now retains 9,500 bytes.
• sizehint! on a fresh buffer only grows it, on Julia 1.10 and later.

Verification:

• New testset DataString: owned bytes are reserved at their final size: (1) plain, quoted, and quote-escaped long cells over many chunks, serial and parallel: values, every byte owned once, and a capacity bound that main fails (328,601 bytes retained against a 208,728 bound); (2) quoted 12-byte and short escaped values with occasional long ones: a capacity bound that the first commit alone fails (1.40 MB and 1.50 MB against a 70 KB bound). Julia 1.10's summarysize ignores capacity, so the bounds only bind on 1.11+.
• Full suite --check-bounds=yes -t4: 331,788 passed (head).
• 300 seeds each of a ground-truth structural fuzzer and a cross-configuration consistency fuzzer (every reader, task count, and chunk size): no differences.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.81%. Comparing base (2f84f42) to head (93a52ab).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1220      +/-   ##
==========================================
+ Coverage   95.79%   95.81%   +0.01%     
==========================================
  Files           9        9              
  Lines        5566     5591      +25     
==========================================
+ Hits         5332     5357      +25     
  Misses        234      234              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.