2026.10.23 — 2026-10-07

Breaking

• Compare every node's configuration against the majority (f056a029)

  GET /clusters/{clusterId}/config-diff and the config_diff MCP tool now compare every node of the cluster at once, instead of two chosen nodes. For each key the response gives each node's value (or that the key is missing there), the majority value (more than half of the nodes that return the key), the outliers that differ from it, and a state of SAME, DIFFERENT or MISSING_ON_SOME. When no value holds a majority it lists every distinct value with its nodes. Each node costs one batched read through the rate limiter, as before.

  A node that cannot be read is listed with its classified reason and left out of every majority, so its absent keys never read as missing; with fewer than two answering nodes the response says so and gives each reason. A passive backup whose management surface is reduced contributes only the keys it exposes, and the response states that. The summary counts the drift keys, the nodes that differ and the expected differences set aside.

Added

• Open Config diff on the keys that drift, across every node (75250648)

  Config diff now compares every node of the cluster against the majority instead of two chosen nodes. It opens on one sentence ("3 keys drift on 2 nodes", or that no key drifts and how many expected differences were set aside) and one table of the drifting keys: the section, the key, the majority value, and each node that differs with its value or that it is missing the key, plus the key's state and class as words.

  A switch shows the expected differences (a broker's name, its node-local paths) or every key; a search covers keys and values, and a filter keeps the keys that differ on chosen nodes. The choice, the search and the nodes are in the address, so a link shows the same view. A node that did not answer is named with its reason and the cluster is not called clean without it. The page links to the declared configuration and says how the two differ.

  The left and right node pickers and the Drift only switch are gone.

• Compare every node's configuration against the majority (f056a029)

  GET /clusters/{clusterId}/config-diff and the config_diff MCP tool now compare every node of the cluster at once, instead of two chosen nodes. For each key the response gives each node's value (or that the key is missing there), the majority value (more than half of the nodes that return the key), the outliers that differ from it, and a state of SAME, DIFFERENT or MISSING_ON_SOME. When no value holds a majority it lists every distinct value with its nodes. Each node costs one batched read through the rate limiter, as before.

  A node that cannot be read is listed with its classified reason and left out of every majority, so its absent keys never read as missing; with fewer than two answering nodes the response says so and gives each reason. A passive backup whose management surface is reduced contributes only the keys it exposes, and the response states that. The summary counts the drift keys, the nodes that differ and the expected differences set aside.