2026.10.66 — 2026-10-30

Breaking

• Store each series once a minute (6510fbfa)

  A scrape every 15 seconds was writing six rows per subscription per tick, about 45 GB a day at 4,000 subscriptions. Samples are now stored at a 60-second heartbeat and never faster, charts will not bucket finer than that, and a rate alert reads two heartbeats so a flat series stays a zero rate.

Changed

• Ask a broker for its kinds once per connection (506a77d0)

  A subscription list asked the broker for its kind catalogue on every row, so a few hundred groups took about 18 seconds. The catalogue is now kept for the life of the connection, and one list asks once.

• Store each series once a minute (6510fbfa)

  A scrape every 15 seconds was writing six rows per subscription per tick, about 45 GB a day at 4,000 subscriptions. Samples are now stored at a 60-second heartbeat and never faster, charts will not bucket finer than that, and a rate alert reads two heartbeats so a flat series stays a zero rate.

Fixed

• Install Ubuntu security updates the base image does not carry (20fa1a46)

  The image scan fails on libbrotli1 1.2.0-3build1 (CVE-2026-111497). The fix is in the resolute security pocket, which the base image does not enable, so the build enables it and upgrades before the scan.

• Keep a Kafka screen still while it fills in (b5a7716d)

  The registration form reserves the row where parsed servers appear, so the first server does not push the rest of the form. At 200% zoom the header stays one line and the registration buttons stay on their row. A group's messages open on the partition controls immediately, and the heading keeps its height when the description arrives.

• Confirm a Kafka purge from the records it will skip (2377556a)

  A purge of a group now states the records the move will actually skip, taken from the broker at the moment of the preview, rather than the lag stored by the last scrape. The scraped lag also counts records an open transaction still holds, and the move does not pass those. A count the broker does not report, and a group already at the last record it can read, are refused and nothing moves. The confirmation says that no record is deleted and that records an open transaction still holds stay.

  Opening the abort dialog no longer fails when the preview does not list the partitions it would release.

• Say when a depth was not reported (a7ae032e)

  An index subscription with no held count, and a group browse with no position, now say the figure is unavailable instead of showing zero or nothing. A purge count that is not a number says so, and an abort preview with no released partitions no longer throws.

• Trim a store that is over its row quota, oldest first (923b3856)

  Housekeeping now removes the oldest rows of a store that has grown past its row quota, in bounded batches, and audits that as one TRIM_STORE event. Broker events start with a quota of twenty million rows: an older day's partition is dropped first, and today's partition loses its oldest rows rather than being dropped while it is still being written. A quota of zero still means no cap.

  This is the start of finding S2 for the follow-up change. It is not part of the Kafka pull request.