Queued PDF job benchmark from April 28, 2026

Study scope

On April 28, 2026, we also measured the queued side of the public tool runtime with the same benchmark harness.

This pass covered:

• repair-pdf
• ocr-pdf
• redact-pdf
• pdf-to-pdfa

Each job used 3 runs and a fixed 250ms mock upstream delay so we could split submit overhead from total end-to-end completion time.

Key findings

• Submit latency split into two bands: repair-pdf and ocr-pdf averaged 138.10ms to 142.69ms, while redact-pdf and pdf-to-pdfa averaged 19.55ms to 23.83ms.
• Total median runtime stayed inside a relatively tight band from 343.21ms to 429.05ms.
• pdf-to-pdfa posted the fastest total p50 at 343.21ms.
• ocr-pdf posted the slowest total p50 at 429.05ms, which is still within the same operational envelope as the rest of the queued group.

Queued benchmark table

Tool	Submit Avg	Total Avg	Total P50
repair-pdf	138.10ms	510.29ms	394.15ms
ocr-pdf	142.69ms	499.41ms	429.05ms
redact-pdf	23.83ms	436.32ms	410.22ms
pdf-to-pdfa	19.55ms	344.28ms	343.21ms

Download the raw data

• Queued benchmark CSV

Methodology notes

• The benchmark measures public Astro submission, queueing, polling, and download behavior as one flow.
• The mock upstream delay is intentionally fixed so we can spot app-layer drift without conflating it with real PDF rendering cost.
• These numbers are operational timing signals, not claims about production file fidelity or full output-quality.

Press-ready angles

• The queued runtime is currently keeping end-to-end medians inside a 343ms to 429ms band under fixed mock conditions.
• Intake overhead is now low enough on redaction and PDF/A that queue and completion time dominate the user-visible wait.
• OCR and repair are the first queued tools to watch as real upstream variance and larger files enter the mix.