3 deterministic fixtures
Lines near unique, Words high cardinality, and one single long line; every input is exactly 10,000,000 bytes.
Performance & correctness
Browser-local text deduplication benchmark
A dated, reproducible account of three deterministic 10,000,000-byte workloads on one controlled machine. These results describe the tested build, not every device or every text distribution.
Test snapshot
Tested 2026-07-13 UTC. Completion starts when a real temporary file is selected through the browser file-input path and ends when the first result row is visible and processing has finished.
20 cold + 20 warm runs
Each accepted fixture was measured in isolated Desktop Chromium and Mobile WebKit frame and memory cohorts.
480 accepted-file measurements
Four cohorts produced 16 validated artifacts with one worker, serial execution, and zero retries.
Independent output oracle
Final full-output length and SHA-256 were checked against independently generated expectations for every fixture.
End-to-end completion results
Nearest-rank p95 values from 20 runs. Cold and warm are shown in that order for each browser profile.
| Workload | Input shape | Desktop Chromium cold / warm p95 | Mobile WebKit cold / warm p95 |
|---|---|---|---|
| Lines, near unique | 625,000 generated line units | 922.2 / 892.8 ms | 961.2 / 895.2 ms |
| Words, high cardinality | 625,000 generated ASCII word units | 1,933.6 / 1,920.2 ms | 1,926.8 / 1,923.2 ms |
| Single long line | One 10,000,000-byte ASCII row | 144.4 / 120.5 ms | 152.4 / 124.9 ms |
Responsiveness and completion actions
Frame monitoring and process-level memory sampling ran in separate cohorts so the instruments did not distort the same run.
| Measurement | Desktop Chromium p95 | Mobile WebKit p95 |
|---|---|---|
| Highest per-run worst frame gap p95 | 33.3 ms | 75 ms |
| Cancel acknowledgement | 43 ms | 46.3 ms |
| Copy full result | 55.2 ms | 63 ms |
| Copy with counts | 100.7 ms | 138 ms |
| Start plain-text download | 69.6 ms | 70.4 ms |
Deterministic fixtures and correctness evidence
Each final full-output length and SHA-256 was checked against an independently generated expected result. Input transfer also confirmed that the transferred buffer detached from the main thread.
Lines, near unique
625,000 sixteen-byte units made from item-, a zero-padded 10-digit index, and a newline. The final generated value repeats the first; the terminal newline creates a valid empty line under the tested defaults.
- Input SHA-256
- 9b94e00788544162aebf9c572e18bb25a5fb4c7bcc955a1f68bfe3b65f0518ec
- Expected full output
- 9,999,984 bytes
- Expected output SHA-256
- 17cd4992175c90e8b8a221dd32c33820ff89b5ac64388d170dd0bd9a8427281b
- Expected counts
- 625,001 occurrences; 625,000 distinct; 624,999 once-only; 1 repeated value kind; 1 extra occurrence removed.
Words, high cardinality
625,000 sixteen-byte units made from w, a zero-padded 14-digit index, and one space. The final generated value repeats the first.
- Input SHA-256
- 7baa0deb0a605e7cc12acece8c2997b825b4c84d8ef583ce9b7debd0cf92c779
- Expected full output
- 9,999,983 bytes
- Expected output SHA-256
- 942bbb91e0c38061c455d1e5c28970b73fa3ad79df5426d0de741aba8570ffdf
- Expected counts
- 625,000 occurrences; 624,999 distinct; 624,998 once-only; 1 repeated value kind; 1 extra occurrence removed.
Single long line
Exactly 10,000,000 ASCII a bytes in one row.
- Input SHA-256
- 01f4a87c04b40af59aadc0e812293509709c9a8763a60b7f9e19303322f8b03c
- Expected full output
- 10,000,000 bytes
- Expected output SHA-256
- 01f4a87c04b40af59aadc0e812293509709c9a8763a60b7f9e19303322f8b03c
- Expected counts
- 1 occurrence; 1 distinct; 1 once-only; 0 repeated value kinds; 0 extra occurrences removed.
Test environment
| Item | Value |
|---|---|
| Tested source | 1bcb726b8c0f9a84d4cf65cf04d1e2e46610adca |
| Production build ID | nwPqPW-iNgIVBeM73cXfb |
| Source tree | Clean |
| Build mode | Production Next.js server |
| Host | Apple M3 Pro, 11 logical CPUs, 18 GiB memory |
| Platform | Darwin 24.3.0 arm64 |
| Desktop | Chromium 140.0.7339.186 at 1280 × 720 |
| Mobile profile | Playwright WebKit 26.0 at 390 × 664 |
| Runtime | Node.js 22.22.2; pnpm 9.10.0 |
| Execution | 1 worker; 0 retries; 30-second watchdog |
Memory method and results
Memory figures are deltas from a per-run baseline. p95 summaries and the largest individual observation are both disclosed so outliers are not hidden.
| Boundary | Highest p95 across fixtures and temperatures | Largest individual observation |
|---|---|---|
| Desktop Chromium main-thread heap | 30,666,413 bytes / 29.25 MiB | 59,801,421 bytes / 57.03 MiB absolute heap |
| Desktop Chromium browser-tree RSS | 250,478,592 bytes / 238.88 MiB | 523,436,032 bytes / 499.19 MiB peak delta |
| Mobile WebKit launch-generation RSS | 131,055,616 bytes / 124.98 MiB | 433,537,024 bytes / 413.45 MiB peak delta |
Additional 300 KB input checks
A separate cohort exercised the current 300,000-byte manual-input boundary.
| Input path | Desktop p95 | Mobile p95 |
|---|---|---|
| Manual text input | 46.8 ms | 49 ms |
| Paste-path harness | 57.2 ms | 54 ms |
| Delimiter-mode manual input | 34.9 ms | 44 ms |
One Desktop paste-path run retained a 716.7 ms frame-gap outlier. It was not removed and remained below the fixed 1,000 ms worst-case gate. The paste harness excludes operating-system clipboard latency.
A defect the benchmark found
Known limitations
- One Apple M3 Pro machine was tested.
- Mobile WebKit ran on the desktop host with a 390 × 664 viewport; this was not a physical iPhone benchmark.
- The deterministic ASCII fixtures do not represent every language, encoding, delimiter pattern, or real-world distribution.
- Frame and memory results come from separate cohorts and must not be correlated as if captured in the same run.
- Cold runs reset page, React, and Worker state, not the browser process, operating-system disk cache, or machine.
- Copy used a synthetic clipboard capture and excludes operating-system clipboard latency.
- Download ends at the browser download event and excludes final storage-write time.
- The 25 MB and 50 MB checks only verified quick rejection; those sizes are not supported inputs.
- CSV content was treated as plain text, not as structured CSV.
- This is a product benchmark, not a third-party certification, security audit, or privacy audit.
- Results stay tied to the tested source and should be rerun after a material processing change.
Frequently asked questions
Does every 10 MB file finish in under two seconds?
No. About 0.12–1.94 seconds was the p95 range for these three synthetic fixtures on the documented machine and browsers. Other devices, distributions, browsers, and background load can differ.
Was Mobile WebKit tested on an iPhone?
No. It was Playwright WebKit 26.0 on the Apple M3 Pro host with a mobile viewport. It exercises WebKit and the mobile layout profile, but it is not a physical-device benchmark.
Does the tool support files larger than 10 MB?
No. The supported boundary is 10,000,000 bytes per file and for the complete accepted batch. Larger fixtures were used only to verify quick rejection.
Does Copy timing include the system clipboard?
No. The benchmark captured output through a synthetic clipboard implementation so it could measure product-side output construction without operating-system clipboard variance.
Does the benchmark cover structured CSV files?
No. CSV content is read as plain text. The tool does not parse headers, select columns, interpret quoted delimiters, unescape quotes, or reconstruct multiline fields.
Is this a fully independent reproducibility package?
No. Fixture recipes, hashes, environment, run counts, percentile method, and sanitized summaries are public here. The complete internal harness and raw process-level artifacts are not published.
Is browser-local processing the same as zero telemetry?
No. Text and file contents are processed locally for deduplication. When configured in Production, limited page and interaction telemetry may be processed separately as described in the Privacy Policy and Capabilities page.