[webhanspeter]

Summary

CsvDriver built a complete PhpSpreadsheet workbook in memory in order to emit a plain CSV. Rows are now written straight to a stream.

rows	before	after
10,000	0.338 ms/row, 135 MB	0.0030 ms/row, 4 MB
25,000	0.341 ms/row, 358 MB	0.0033 ms/row, 8 MB
50,000	0.358 ms/row, 685 MB	0.0041 ms/row, 8 MB
100,000	0.523 ms/row, 1,364 MB	0.0047 ms/row, 8 MB
318,000	~165 s, ~4.4 GB (projected)	1.66 s, 8 MB

About 100x less time per row, and memory is flat — the buffer is capped and spills to a temp file beyond it, so it no longer scales with the export. Note the old figures get worse per row as the set grows (0.338 to 0.523 ms) because the working set runs into gigabytes; the new ones do not.

Along the way the driver stopped needing two dependencies, and two correctness bugs fell out.

The header only described the first row

toSpreadsheet() took the header from array_keys($data->first()) but fromArray() wrote each row in that row's own key order. Any row whose keys differed in order or membership silently put values under the wrong headers — wrong data, no error.

The header is now the set of keys any row has, in the order first seen, and every row is written against that list. A missing key writes an empty field instead of shifting everything left.

Values were being coerced

A spreadsheet infers a type per cell. DefaultValueBinder decided a numeric-looking string was a number and Cell::setValueExplicit() did $this->value = 0 + $value, at which point the text was gone. A leading = was parsed as a formula and the calculated result written instead.

value produced	old file	new file
8.00	8	8.00
1234.50	1234.5	1234.50
0.10	0.1	0.10
=SUM(A1)	0	=SUM(A1)
+4512	4512	+4512
1e3	1000	1e3

A CSV holds text, so nothing is inferred any more. Leading zeros and long integers were already safe — the binder has explicit guards for those — but trailing decimal zeros were not.

There is a second coercion site in the same library worth knowing about: the CSV writer separately strips trailing zeros from decimals beyond 15 digits of precision. Also gone.

Quoting is unchanged, deliberately

fputcsv quotes any field containing a space, which the old writer with setEnclosureRequired(false) did not — that would have put quotes around most text columns in every row. So rows are assembled directly, enclosing only on the delimiter, the enclosure, or a newline, which is what Writer\Csv did:

if (!$this->enclosureRequired && strpbrk($element, "$delimiter$enclosure\n") === false) {

A bare carriage return is also enclosed now, since rows are separated by one. Any value containing one previously produced a broken file.

Verified by byte comparison

Ran a realistic export through both writers — 21 columns, values including a comma, embedded quotes, a padded identifier and decimals — and diffed the files. Exactly two fields differ, both the fidelity fix:

duration   old='7.5'   new='7.50'
duration   old='0.1'   new='0.10'

Quoting, column order, line endings and every other value are byte-identical.

Test plan

• Four new tests: value fidelity, header alignment across rows with differing keys, quoting only where required, and empty input.

• Existing driver tests unchanged in expectation, including the exact "name\r\nJohn Doe\r\nJane Doe\r\n" assertion.

• Full suite 24 passed, 73 assertions.

• ECS and PHPStan clean.

• Suite re-run with both dependencies uninstalled, to prove neither is still needed.

• Rector clean. (It used to fail on a clean checkout under coding-style 4.x, which registered the deprecated ArraySpreadInsteadOfArrayMergeRector; the 5.x bump now on main fixed that, so it runs properly again.)

Breaking changes

Needs a major release.

• phpoffice/phpspreadsheet and guzzlehttp/guzzle are no longer required. Neither is referenced anywhere in src/ any more — guzzle existed only for toStream()'s return type.
• toSpreadsheet() and toCsvWriter() are removed. Both took or returned PhpSpreadsheet types, and the package has never produced a spreadsheet: GeneratorType has one case, CSV, and GeneratorManager registers one driver.
• toStream() now takes the processor data and returns a stream resource, rather than taking a writer and returning a PSR stream. It was not called anywhere inside the package before.
• saveToStorage() takes a stream rather than a string, and uses Storage::writeStream().
• Exports containing numeric-looking values change, as above. Anything reading them numerically is unaffected; anything comparing the file as text will see restored decimals.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: 76a6dcf9-1006-4fec-bf21-bb5b8db0d43d

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}