ADR-0056: Every view is bounded, and the topology canvas degrades by level of detail

• Status: accepted
• Date: 2026-09-06
• Deciders: maintainers

Context

Studio has a virtualised grid (ADR-0020) with paging, sorting, sort announcement, node attribution and a reveal panel for clipped cells. Three views use it: queues, resources and messages. operator-ui already requires new tabular views to use it.

The rest of the product does not, and the gap is not a disagreement — those views predate the requirement:

• events and audit render a whole page of rows into a Mantine <Table>; events additionally merges a live buffer of up to 500 rows on top of it.
• dlq renders every address, every queue beneath it, and every per-node row of every queue, with no paging at any level.
• rr/FlowsTable renders whatever one page returned and offers no way to reach the next, so a busy cluster's flows are silently truncated at the page size.
• topology hands every node to React Flow at every zoom level.

None of this is visible on the two-node dev pair every contributor runs, and all of it is visible on the clusters this product exists for. The failure is not a crash; it is a console that becomes slow and then unusable at exactly the cluster size where an operator most needs it, and — in the flows case — one that omits data without saying so.

Decision

We will bound what every view renders, so that the work a screen does is a function of what the operator can see rather than of how large the cluster is. A view listing rows either presents a bounded page with a way to reach the rest, or renders only the rows in the viewport, or both.

We will reuse VirtualTable rather than introduce a second table. For events and audit this is adoption of an existing requirement, not new design.

We will make every bound visible. An operator must never be unable to tell whether they are looking at everything. A truncated flows list gets a pager; a reduced-detail topology says it has been reduced.

The topology canvas will degrade by level of detail, because a graph cannot be paged. Above a node threshold it collapses each live/backup pair into a single node and drops edge labels; node internals render only above a zoom threshold; and React Flow renders only visible elements.

Consequences

• The DLQ view stops showing every per-node row at once. Per-node detail moves behind an expansion, which is one more interaction for the operator who wanted all of it — and the only version of the view that survives a cluster with hundreds of dead-lettered queues.
• Collapsing a live/backup pair hides real structure at high node counts. It is mitigated by stating that detail was reduced rather than reducing it silently; an operator who cannot tell what they are being shown is worse off than one looking at a slow graph.
• The node threshold and zoom threshold are constants chosen by judgement, not measurement. They are in one place and expected to move once someone runs this against a genuinely large cluster.
• events keeps its live buffer and its existing cap; virtualising the rows does not change how many are held, only how many are in the DOM.

Alternatives considered

Leave it and revisit when someone complains. The complaint arrives during an incident on the largest cluster in the estate, which is the worst possible moment to discover it.

Cap the response server-side and drop the rest. Bounds the DOM and loses data without telling anyone — the specific failure rr/FlowsTable already has.

Virtualise the topology graph by viewport alone. Necessary and not sufficient: the layout pass itself is over every node, and a hundred nodes drawn at full detail are illegible whether or not they are all in the DOM.

A separate "large cluster" mode the operator opts into. Two code paths, one of which is exercised only by the users least able to tolerate a bug in it.