Writing the same node again no longer reconnects every time

When one part of the portal writes a node that another part owns — recording that you opened a page, appending a line to a running activity, updating a progress counter — it does so through a live connection to the node's owner: the writer holds a copy of the node, computes its change against that copy, and sends only the difference.

That connection used to be thrown away after every single write. The moment a write was saved, the portal discarded the writer's copy so that the next write would be sure to start from fresh state, and the next write paid for a brand-new connection: a new subscription, a new round trip for the node's current state, new plumbing on both ends. On a node written once, that is invisible. On a node written constantly it is not — the record of your own logins is written on every cold page load, and one of them had been rebuilt more than six thousand times.

Each rebuild had to deliver the node within 30 seconds or the write was abandoned. Since 10 August that happened 414 times on the production portal, always on the busiest nodes: the login record, and pages people come back to. The visible effect was quiet — a visit that did not count, a last-accessed time that did not move — and the log said only that no state had arrived in time.

What changed

The connection now stays open. Every saved write announces its version number, and the next write simply waits until its copy has caught up to that version before computing its change — which normally takes no time at all, because the owner sends the new version down the connection that is already open. A copy that is provably behind is still thrown away and rebuilt, once, exactly as before; a delete, a re-creation or a restart of the owner still resets the connection too.

Nothing about what is written changes: a write still starts from the newest saved state, and two writers still cannot overwrite each other's changes. What goes away is the rebuild — measured on a test mesh, six writes to one node used to open five new connections after the first and now open none.

The details, including why the obvious shortcut (keeping any connection that looks healthy) loses writes, are in Live Mirrors and the Change Feed.