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

[leedqin]

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 on the cluster's largest replica (up to 10), each as a share of that replica's 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.
• The largest replica's loading and error states move from LargestMaintainedQueries into LargestReplicaBoundary, which the table and the card share.
• 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. Not included from the design: the replica dropdown. The card uses the largest replica, as the table did.

Verification

MemoryByObject.test.tsx covers:

• the title, the replica, and each object's size and share;
• 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;
• nothing rendered for a cluster without replicas, once the replica query has answered;
• the busy-replica message when the objects fail to load;
• 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.

🤖 Generated with Claude Code

[linear-code]

CNS-173

CNS-175

[jdonelson]

The count is the number the query returns, which caps at 10, so a cluster with hundreds of objects shows "(10)". Something like "Top 10 objects by memory" would avoid reading it as a total.

[leedqin]

Changed to "Top 10 objects by memory".

[jdonelson]

I have some layout-related feedback:

The legends for these 2 graphs are tough to read. On the freshness graph it overflows to the left. Maybe remove them both and move the color swatches to the Freshness table, like we did on the new Freshness page? You'd have to make each object use the same color on both graphs.

[leedqin]

I have some layout-related feedback: @jdonelson

I have a followup PR coming to remove this Freshness section so you can add your Freshness tab/ graph once your page is shipped and I agree that we can use the Freshness Swatch legend

[leedqin]

Thanks. This PR now colors the card's objects by rank with the Freshness page's palette (generateRainbowPalette), so it matches that page. The legends and the Freshness table are unchanged here.

[def-]

QA LLM Review

1. MEDIUM -- Bar's 100% is one process's heap limit, but object sizes span every process

console/src/platform/clusters/ClusterPage/MemoryByObject.tsx:88

The card treats each object's memoryPercentage as an additive share of one heap-limit bar. On a replica with more than one process, that number is the replica-wide arrangement size divided by a single process's heap limit, so every share is multiplied by the process count. On those replicas the bar fills to 100% while the replica still has headroom, and segmentWidths cuts smaller objects out of the bar even though the legend still lists them.

Details

The denominator is MAX(crm.heap_limit) (console/src/api/materialize/expressionBuilders.ts:229, used via LargestMaintainedQueries.tsx:178). mz_cluster_replica_metrics has one row per (replica_id, process_id), and heap_limit comes from the per-pod memory and disk limits (src/controller/src/clusters.rs:832). The base PR's sharedHeapLimitBytes says the same thing: "One process's heap limit in bytes". The numerator is a total for the whole replica. mz_object_arrangement_sizes is a COUNT(*) over every worker's raw size logs (src/adapter/src/coord/introspection.rs:656), and the pre-v26.35 query sums the same logs.

Example: a 6400cc replica (scale: 2, misc/helm-charts/operator/values.yaml:252) with each process's heap limit H. Objects A, B and C use 0.6H, 0.5H and 0.4H across both processes, so together they hold 37.5% of the replica's 2H heap. The card reports 60%, 50% and 40%. The bar shows A at 60% and B clamped to 40%, so it reads as completely full. C gets no segment at all. These objects are on the largest clusters, which is where this card matters most.

The classic table already showed the inflated percentage for each row. The new part is the stacked 0–100% axis, plus the clamp that hides the overflow. A fix is to use the replica's total heap as the denominator, for example SUM(heap_limit) across processes or the limit times mz_cluster_replica_sizes.processes. That would correct the table too.

[leedqin]

Thanks, fixed in f683add: the replica's heap limit is now the per-process limit times the size's process count, which also corrects the classic table. The SQL test covers a four-process size.