On October 21 we shipped faster databases: tables with 50,000 rows now open in under a second on any device. Plus workspaces can hold 50,000 rows per database and Business 250,000, so this was not an edge case. It was the main case for our biggest teams.
Where the time went
We traced 400 slow opens from real workspaces (with permission, and with the content stripped). The pattern was the same every time. The client asked for every row, waited for all of them, sorted them, filtered them, then drew the 30 that fit on screen.
| Step | Before | Share |
|---|---|---|
| Fetch all rows | 1.9 s | 56% |
| Sort and filter in the browser | 0.9 s | 26% |
| Compute formula columns | 0.4 s | 12% |
| Draw the first screen | 0.2 s | 6% |
Draw what fits, then stream
The fix is simple to say. The server now sorts and filters, sends the first 200 rows with the view, and streams the rest in the background while you read. Scrolling past the loaded rows asks for the next page before you reach it.
// Simplified: the view request now carries its own sort and filter.
const page = await db.view(viewId).rows({
sort: view.sort, // e.g. [{ property: "due", direction: "asc" }]
filter: view.filter, // evaluated server-side, permission-checked
limit: 200,
after: cursor, // null for the first page
});The hard part was keeping edits instant. When you change a status on row 30, the row has to move to its new place in the sort before the server answers. We kept the client-side sort for rows already on screen, and let the server correct the order quietly if it disagrees.
The boring cache
Most opens are re-opens: the same person, the same view, an hour later. We now keep the first page of every view you opened in the last week on your device. A re-open draws from the cache in about 120 ms, then checks for changes.
The result
| Rows | Before | After |
|---|---|---|
| 1,000 | 0.4 s | 0.2 s |
| 10,000 | 1.1 s | 0.4 s |
| 50,000 | 3.4 s | 0.8 s |
| 250,000 | 14.2 s | 1.3 s |
Four times faster at 50,000 rows, ten times at 250,000. Sorting by a formula column no longer recalculates the whole table either. If you find a view that still feels slow, the support team will want the link.