Skip to content
Laurel Plate

Journal

Notes from the engineers and support team. Release notes live in the changelog.

Why a verified checksum beats a successful upload

An upload that finishes is not an upload that worked. Over the last 12 months our servers rejected 0.011 percent of completed uploads because the SHA-256 computed on arrival did not match the one the loader sent. That is about one file in 9,000, and every one of them looked fine in a file browser.

The causes were boring. Catalog software rewrote a sidecar mid-transfer. A card reader returned stale data on a second pass. Once, a hotel gateway rewrote a response header and truncated a 61 MB file by 4 KB. A gallery service that does not hash would have shown the first frames and hidden the damage until a print lab asked why a bride's veil had a stripe through it.

So we hash before the first byte leaves, hash again on arrival and hold the file in quarantine until both agree. From version 3.4 the loader also re-reads the part it just sent if the server reports a mismatch, which cut retry traffic by 31 percent.

A 42 GB wedding for a grandmother on hotel wifi

We do a lot of measuring around the unglamorous case: the recipient is a grandmother, the file is 42 GB, the connection is a hotel's. A plain download of one big file fails somewhere around the third hour and starts again from zero, because the browser throws away the partial file.

Splitting the delivery into 4 GB parts fixes the failure mode. A failed part costs at most 4 GB of retry, and the parts resume with range requests. In a test run on a throttled 20 Mbps link with a forced disconnect every 20 minutes, a monolithic 42 GB download never completed in 10 attempts; the split version completed in 5 hours 21 minutes every time.

The helper script that rejoins parts on desktop went through four rewrites. The final one is 38 lines and does one thing: concatenate, then verify against the manifest.

What happened in Frankfurt on August 19

On August 19 at 14:05 UTC we shipped a change to the part index service, the component that remembers which 64 MB parts of an upload have been acknowledged. The change made the index smaller. It also stopped answering for uploads older than 6 hours, which is exactly the case when someone resumes an overnight job.

New uploads worked, resumed uploads failed with a generic error, and our health checks only exercised new uploads. A photographer in Gdansk told us at 14:16. We rolled back at 14:27. Nothing was lost, since failed resumes simply retried later.

We have added a resume check to the health suite, and the part index now has a test that replays a 20-hour-old upload on every build.

Publishing restore times, not promises

Most archive services describe cold storage as "retrieval within hours". We would rather you could plan. Between March and May 2026 we logged 118,400 restores from settled storage. Median time to first byte was 61 seconds, the 95th percentile was 3 minutes 40 seconds and the slowest was 4 minutes 52 seconds.

That is why the docs say "up to five minutes", and why the number will change on this page if the data does. Restores of 5 TB or more are scheduled rather than interactive, and finish in about four hours.