RTC: Fix divergence when two offline users reconnect

[alecgeatches]

What?

In the "Edit Anyway" PR, we discovered a way to get a user into a divergent state after reconnecting to the sync server:

chaos-merge.mov

Above, the left user stays online and makes some typo fixes, while offline user 1 adds content and offline user 2 replaces the whole paragraph. The purpose of the video was to show that a complex merge would have a weird result, but what I found was that the second user to re-connect (bottom-right in the video) got stuck in a non-convergent state.

"Edit Anyway" mode makes this situation easier to reproduce, but the core issue is reproduceable with tighter timings in core. This will happen when two users who have made changes reconnect to a sync server on the same polling cycle.

This PR stops the RTC sync server from silently dropping a client's compaction when another client has already pushed one for the same room.

Why?

When two users edit the same post offline (via the near-future "Edit Anyway" flow or in the window of time before a disconnect dialog appears) and reconnect at roughly the same time, the second reconnecter's changes silently disappear from the server. Their local state still shows their changes, but no other client ever sees it, and the server-side document is the first reconnecter's state.

The root cause is in process_sync_update()'s compaction handling. Both clients push with the same cursor, the last position they saw before disconnecting. The server applies the first one and then sees has_newer_compaction=true for the second, and drops the second compaction. Because the polling provider resets its queue to a single compaction when updates are present on every poll failure, the second user to reconnect has their local changes dropped and becomes divergent.

How?

The fix is pretty simple. Instead of discarding the bytes when a newer compaction already exists, store them as a regular Yjs update. Y.js's applyUpdateV2 merges state-as-update blobs idempotently, so there is no conflict with the operations the first compaction already encoded. We cannot safely call remove_updates_before_cursor in this branch (the newer compaction has already advanced the cursor), so we skip that and store only.

The tricky part is testing, because class-wp-http-polling-sync-server.php already ships in core WordPress, so this is difficult to test in WordPress 7+. The matching change in core, along with a unit test, will land via a core PR (along with an update to the in-progress WordPress/wordpress-develop#11599 table update, which will work the same). The instructions below use a hacky method for patching wp-env core WordPress files to test the fix.

Testing Instructions

First, in trunk, try a manual reproduction:

1. Enable RTC via Settings -> Writing -> "Enable real-time collaboration" checkbox.

2. Manual reproduction is possible with some quick timing and DevTools offline mode:

   1. Create a post with the content: "Our team gathers every Teusday morning. We share progres on thier projects."

   2. Open the post in three browser sessions logged in as three different users.

   3. Disconnect two of them by going on offline mode in DevTools.

   4. In the first offline user's browser, add the sentence "Mondays are quiet." to the beginning of the paragraph.

   5. In the second offline user's browser, replace the whole paragraph with "Weekly team check-in."

   6. In the online browser, fix the three typos in the original text ("Teusday", "progres", "thier").

   7. Reconnect the first offline user, and then the second:

      chaos-merge-manual-reproduction.mov

3. You should see that the second offline user, on reconnection, has changes from the other two users but their changes ("Weekly team check-in.") are not sent to the first two online users. Changes in the divergent user's document may not show up for other users, depending on the update location (shown above).

To test the fix, we'll need to patch core WordPress files:

1. Patch the wp-env core class so the fix is actually live (since core overrides the Gutenberg copy). Note that this requires a wp-env destroy to undo.

   npm run wp-env -- run wordpress perl -i -pe 'BEGIN{undef $/;} s|// Reaching this point means there.s a newer compaction, so we can\n\s*// silently ignore this one\.\n\s*return true;|return \$this->add_update( \$room, \$client_id, self::UPDATE_TYPE_UPDATE, \$data );|' /var/www/html/wp-includes/collaboration/class-wp-http-polling-sync-server.php
   

   To revert core changes, you'll need do destroy wp-env via:

   npx wp-env destroy && npx wp-env start
   

2. Retry the manual reproduction above. See the users converge correctly:

   chaos-merge-patched.mov

   Note that users converge to a nonsense result (WMondays are quiet. eekly team chesick-in.), but that's okay because the changes made offline were purposefully incompatible. This exercise isn't to show that the merge result is ideal, just that each user sees the same replicated state.

Use of AI Tools

AI assistance: Yes
Tool(s): Claude Code
Used for: Diagnosing the silent-drop branch, drafting the fix.

[github-actions]

The following accounts have interacted with this PR and/or linked issues. I will continue to update these lists as activity occurs. You can also manually ask me to refresh this list by adding the props-bot label.

If you're merging code through a pull request on GitHub, copy and paste the following into the bottom of the merge commit message.

Co-authored-by: alecgeatches <[email protected]>
Co-authored-by: maxschmeling <[email protected]>

To understand the WordPress project's expectations around crediting contributors, please review the Contributor Attribution page in the Core Handbook.

[github-actions]

Flaky tests detected in a33d8a0.
Some tests passed with failed attempts. The failures may not be related to this commit but are still reported for visibility. See the documentation for more information.

🔍 Workflow run URL: https://github.com/WordPress/gutenberg/actions/runs/25399388024
📝 Reported issues:

• [Flaky Test] post title and content are editable and blocks can be inserted #77720 in /test/e2e/specs/editor/various/post-content-focus-mode.spec.js
• [Flaky Test] should move focus back to the Publish panel toggle button when canceling #77721 in /test/e2e/specs/editor/various/publish-panel.spec.js
• [Flaky Test] should show correct homepage actions based on current homepage or posts page #77385 in /test/e2e/specs/site-editor/homepage-settings.spec.js

[github-actions]

I just cherry-picked this PR to the wp/7.0 branch to get it included in the next release: f2a4bfa

[tyxla]

I just cherry-picked this PR to the release/23.1 branch to get it included in the next release: d037dc8