console: add a memory by object card to the redesigned cluster page - #39510

Motivation

Fourth PR in the cluster details redesign stack, on top of #39508. The design replaces the resource intensive objects table with a memory by object chart, so the Metrics tab shows where a replica's memory goes next to the resource usage card. Fixes CNS-175

Description

• On the new page's Metrics tab, a card titled Top 10 Objects by Memory (or Objects by Memory when there are fewer than 10) replaces the resource intensive objects table.
• It reads the same query as the table: the largest objects (up to 10) on the replica the page's Replica select names, or on the cluster's largest replica under All replicas, each as a share of that replica's heap limit. LargestReplicaBoundary and buildLargestClusterReplicaQuery take an optional replica name, which reads that replica with the same heap limit.
• One bar of the heap limit is split by object, with ticks at 0 to 100%. Hovering a segment names the object. The legend gives each object's size and share, formatted as the table did. Objects are colored by rank with the Freshness page's palette (generateRainbowPalette).
• Each object links to its Maintained Objects drawer (maintained-objects/:objectId). While maintained-objects-ui-50 is off that route doesn't exist, so the link goes to the object's workflow graph, as the table's rows did. A dropped object's dataflow keeps its name and has no link.
• Without a heap limit there are no shares to draw, so the card lists each object's size alone, with no bar. That covers a replica that hasn't reported a limit yet, and the local process orchestrator, which never reports one. The older query sorts by size, so the top 10 stay the largest in that case.
• The card's title and frame stay in place while it loads, when the objects fail to load, and when the replica has no objects in memory, which it says.
• The largest replica's loading and error states move from LargestMaintainedQueries into LargestReplicaBoundary, which the table and the card share. The boundary takes an optional wrapper for its fallbacks, which the card uses for its frame.
• The largest replica's heap limit now covers all of its processes: the per-process heap_limit times the size's process count from mz_cluster_replica_sizes. Object sizes are replica-wide, so on a multi-process replica every share was inflated by the process count, which also affected the classic table. The process count comes from the size rather than a sum over mz_cluster_replica_metrics rows, because a process whose metrics fetch failed has no limit to sum. Checked on a staging environment: a 2-process replica's limit doubles, and EXPLAIN adds only a differential join on mz_cluster_replica_sizes_ind.

The design plots memory by object over time. Per-object memory is only reported as a current value by default: mz_object_arrangement_size_history isn't collected unless arrangement_size_history_collection_interval is set, and it samples too coarsely for short windows. So the card shows the current snapshot for now.

Not carried over from the table: the freshness column and the hint that memory grows until hydration finishes. The card follows #39508's Replica select rather than having its own dropdown.

Verification

MemoryByObject.test.tsx covers:

• the title, the replica, and each object's size and share;
• the named replica instead of the largest;
• links to the Maintained Objects drawer, and to the workflow graph while that page is off;
• a dropped object's dataflow shown without a link;
• sizes without a bar when the replica reports no heap limit;
• the card's title while loading, and when the replica has no objects;
• nothing rendered at all for a cluster without replicas, once the replica query has answered;
• the busy-replica message, inside the card, when the objects fail to load;
• objects read from mz_object_arrangement_sizes on v26.35 and later, with a dropped object labelled by its id;
• the bar stopping at the heap limit.

LargestMaintainedQueries.test.tsx passes unchanged after the shared boundary moved out.

largestClusterReplica.test.sql.ts covers a replica whose size runs four processes, and a named replica read with its own heap limit.

🤖 Generated with Claude Code