Skip to content

chore: version packages - #264

Open
github-actions[bot] wants to merge 1 commit into
mainfrom
changeset-release/main
Open

github-actions[bot] wants to merge 1 commit into
mainfrom
changeset-release/main

Conversation

@github-actions

@github-actions github-actions Bot commented Nov 20, 2026 •

Copy link
Copy Markdown

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and publish to npm yourself or setup this action to publish automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

zero-rust@0.5.1

Patch Changes

  • 8d84a26: Opening a query with joins takes less time. zbugs' issue list at a limit of 1,000, on a replica of 8,000 rows, opens and closes in 6.3 ms instead of 6.6 ms on a Mac, and 21.3 ms instead of 24.2 ms on Linux. A join's index of its parent rows keeps no set for a join key with one parent, and none for a join on the parent's primary key.
  • cbea3dc: The README suggests jemalloc as a Linux app's global allocator. glibc's malloc is slow at the client's many small allocations: with jemalloc, Zero::new on 8,097 stored rows takes 25 ms instead of 34 ms, and a poke of 1,000 rows on SQLite 4.9 ms instead of 6.1 ms, with the same memory. mimalloc is about as fast but holds a quarter more memory for 100,000 rows.
  • d29ff10: Zero::from_prepared and Zero::new take less of the app's thread to load a stored replica. On a Mac, from_prepared takes 3.5 ms instead of 5.4 ms for a replica of 8,000 rows, and 2.0 ms instead of 2.6 ms for 2,000; Zero::new takes 22.8 ms instead of 23.7 ms for 8,000. On Linux, from_prepared takes 9.9 ms instead of 13.1 ms for 8,000 rows. The replica's map is built a level at a time from the stored rows, which come in key order, and the query sources load straight from it instead of from an add made for every row.
  • 7088c8b: Applying pokes takes less time. A first sync of 2,000 rows takes 3.2 ms instead of 3.6 ms on a Mac, and 6.6 ms instead of 7.0 ms with zbugs' issue list; on Linux, 7.8 ms instead of 8.8 ms, and 14.9 ms instead of 16.0 ms. A poke of 1,000 changed rows takes 2.33 ms instead of 2.45 ms on a Mac. A patch's operations are no longer copied to keep the last of each key, a one-column primary key is written straight into its row's key, and a source's index shares its sort order instead of copying it on every lookup.
  • e79729e: Zero::from_prepared takes less of the app's thread. On a Mac, it takes 2.1 ms instead of 3.8 ms for a replica of 8,000 rows, and 1.6 ms instead of 2.2 ms for 2,000; on Linux, 6.2 ms instead of 9.6 ms for 8,000. Zero::prepare now also sorts each table's rows for the query sources on its own thread, so the app's thread only puts them into the sources and hydrates the first views.
  • f3b25e1: On Linux, a client on SQLite storage applies pokes and mutations faster. A first sync of 2,000 rows with zbugs' issue list takes 17.5 ms of the app's thread instead of 26.4 ms, a poke of 1,000 changed rows 6.1 ms instead of 8.2 ms, and a mutation's round trip 51 µs instead of 107 µs. The store's thread freed the writes it was handed, which glibc does under the lock of the thread that allocated them, so the client's thread stalled on its own allocations; the writes now go back to be dropped on the client's thread. macOS didn't stall this way, and doesn't change.

zero-swift@0.10.1

Patch Changes

  • 85c3325: Big pokes apply faster. A poke's rows are no longer sorted before they go into the replica, where their order doesn't matter, and a poke that changes a large share of the replica, such as the first sync, is diffed in one walk of the replica instead of sorting its keys and looking each one up. On a Mac, zbugs' first sync of 2,058 rows with the issue list open takes 11.7 ms instead of 13.8 ms, 5.7 ms instead of 8.0 ms with no view, and a poke of 1,000 rows into a 1,000-row view 3.3 ms instead of 3.8 ms.
  • eecc8aa: The replica takes a fifth less memory: the rows it keeps as JSON objects no longer keep room for up to twice their columns. Parsed objects start as big as the last one parsed like them was and give up much spare room when they close, and rows turned into JSON objects start at their size. On a Mac, 100,000 rows of 10 columns with a view of them take 129 MB instead of 161 MB. Opening a stored replica and big pokes take 2% less time.
  • 90241b6: A LiveQuery of many rows (and @ZeroQuery) updates faster: the rows it reads again copy the unchanged rows around them in runs instead of one at a time. On a Mac, an optimistic update with a live query of 1,000 rows takes 31 µs instead of 51 µs.
  • 18b4bba: Queries fetch and push their rows faster: a query source finds the index a fetch reads, and writes each index on a change, without writing the index's sort out as JSON to look it up. On a Mac, an optimistic update of an issue in zbugs' issue list takes 12% less time, and its first sync 6% less.
  • 93d8c4a: VirtualList scrolls with less work on the Mac's main thread: it tells SwiftUI its size (what it's offered, as before) instead of SwiftUI asking Auto Layout for it, over every row's views, each time the stack around the list laid out, which it did for each snapshot. In the TestHost's issue list on a Mac, the main thread's work per frame while scrolling falls by 15–30%. The list's size is the same in every layout.

@github-actions
github-actions Bot force-pushed the changeset-release/main branch 11 times, most recently from 16e3672 to 6656032 Compare November 20, 2026 22:07
@github-actions
github-actions Bot force-pushed the changeset-release/main branch from 6656032 to 14d9f55 Compare November 20, 2026 23:53

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

0 participants