Skip to content
Smol.ai
← All engineering notes

Engineering · Oct 5, 2026 · 5 min read

A dropped connection shouldn’t restart a big upload

Chrome had sent 18 parts of a file when we reloaded the tab. Here’s how Smol knew which two to send next and why it asked us for the file again.

We reloaded Chrome with two parts of a 200 MiB upload left to go. When we picked the file again, the browser sent parts 19 and 20. That was the whole retry. We downloaded the finished file and its SHA-256 hash matched the original. The browser run

The bit you notice as a user is that Smol asks for the file a second time. That’s deliberate. Two edits of the same recording can have the same name, and stitching the beginning of one to the end of the other would make for a very bad handoff.

Keep the parts that arrived

The browser cuts a file into 10 MiB parts. Each time storage acknowledges a part, the browser saves its identifier and a hash of its contents alongside the upload ID. That little record in browser storage is the checkpoint. It survives a reload, so a new tab can find the upload and see how much work is left.

Scale the example up to 100 equal parts. If the first 90 are in the checkpoint, you have ten left to send. If part 91 reached storage just as the connection failed, the browser might never have saved the acknowledgement. It can send that part again.

What gets sent after an interruption at 90%?Full restart sends all 100 parts again. Resuming reuses 90 acknowledged parts and sends the remaining 10. Calculated example, not measured transfer speed.Full restart100 parts sent againResume90 reused10 sent
In this 100-part example, green marks the 90 parts Smol already has. Only the ten blue parts travel again.

The uploader starts with two parts in flight. It can work up to four when the connection holds up, or back off to one when it slows. Pausing keeps the checkpoint. Smol shows 100% after the server has assembled the file and confirmed it’s ready.

Reselecting the file is a check, too

Imagine an editor has exported a second episode.mov since the upload started. The name still fits. Even the size might match. Which version did you just pick?

Smol samples the file to find the likely upload, then hashes each completed part of your selection and checks it against the saved hash. If the bytes have changed, it asks for the original. That’s why resuming can take a moment before the progress bar moves. Your computer is reading bytes it may never need to send.

The server checks your workspace and folder access again, too. Losing access means you can’t use an old checkpoint to finish the upload. In a separate test, we restarted the upload service partway through a 101 MiB file. The parts it had already accepted still counted when it came back up. Recovery checks

What the size numbers mean

Keep the original recording until the upload finishes, and leave the browser’s site data in place if you plan to come back to it. Clearing that data erases the checkpoint.

Smol accepts files up to 25 GiB. We checked a 101 MiB original on the live service on October 5 and verified its hash after downloading it. The older 12.25 GiB result came from Smol’s predecessor, which took a different path into storage. Our live check at 101 MiB leaves performance at the 25 GiB ceiling open. Live download check · Transfer data from both systems

Comparing Google Drive and Dropbox

Both services already have ways to keep parts of an upload. Google Drive has a resumable upload API. Dropbox lets a client keep an upload session open for up to seven days. In our Chrome test, we compared the 20 MiB Smol actually resent with the 200 MiB a restart from zero would require. We didn’t reload either competitor’s browser uploader in that test. Google Drive uploads · Dropbox upload sessions

If you’re carrying a much bigger archive, check how you’ll upload it. In our October 5 comparison, Google listed a 5 TB file limit and a 750 GB daily upload and copy allowance per user. Dropbox listed 350 GB on the web and 2 TB from its desktop app. A second Dropbox guide warned of browser timeouts above 375 GB. For a file near either browser figure, we’d use the desktop app. Google Drive limits · Dropbox uploads · Dropbox’s browser guidance

A podcast team also needs to know what happens after the recording arrives. A two-hour interview, for example, runs 15 minutes over the 105-minute transcription limit Dropbox documented when we compared the services on October 5. Google Drive can search transcripts for captioned videos. Smol’s media processor takes recordings up to six hours and works through them in five-minute chunks under a daily budget. Check the limit for the job you need, whether that’s storing the original, transcribing it, or reviewing a particular cut. Dropbox transcripts · Google Drive transcripts

Storage at cost and the team-plan comparison

An editor who wants to keep the original and its backup needs space for both. At the R2 Standard storage rate we used in our October 5 comparison, 100 GB held for a month costs about $1.50. Keep a full backup and the 200 GB costs about $3 for storage, before allowances, requests or taxes. We plan to pass storage and transfer costs through without markup. Processing costs extra in that model, and we haven’t turned on billing. R2 pricing

On the same date, three annually billed Google Workspace Business Standard seats came to $42 a month, and three Dropbox Standard seats came to $45. Those prices included a suite of tools and pooled storage, 6 TB for the three Google seats and 5 TB for the three Dropbox seats. If your team already pays for Workspace and has room in Drive, putting another episode there may cost you nothing extra. At 2,000 GB, R2 storage alone came to about $30 a month, while a single $14 Workspace Standard seat included 2 TB. The homepage calculator lets you try the comparison with your own team and library size. Smol’s storage comparison · Google Workspace plans · Dropbox plans

More from the workbenchFollow via RSS ↗

Smol.ai
Opening your workspace…