5.10 Deploy / monitor / scale / backup — AdminClient's operational role

This chapter is, quietly, where lag monitoring comes from:

What you're monitoring	Calls	What they give you
Consumer lag, computed correctly	`listConsumerGroupOffsets(group)` `listOffsets(OffsetSpec.latest())`	Committed offset per partition; log end offset per partition — `lag = latest - committed`.
Group health	`describeConsumerGroups(…)`	Members, hosts, assignments, assignment algorithm, coordinator host.
Cluster health	`describeCluster()`	`clusterId`, nodes, controller — controller identity and count is a top-tier alert (see Ch. 2/13).

The important point: this is the supported way. Parsing __consumer_offsets is unsupported and breaks on upgrade (#15).

Incident	AdminClient response
Broker config lost	`describeConfigs` on a surviving broker
App must reprocess (bad output, DR failover)	`alterConsumerGroupOffsets` — stop the group first; reset the state store too
Leadership imbalanced	`electLeaders(PREFERRED)`
Partition leaderless	`electLeaders(UNCLEAN)` ← accepts data loss
Broker overloaded / decommissioning	`alterPartitionReassignments` (+ throttle via quotas)
Reassignment gone wrong	`Optional.empty()` to cancel it
GDPR deletion request	`listOffsets(forTimestamp)` + `deleteRecords(beforeOffset)`
Topic misconfigured	`incrementalAlterConfigs` (SET / DELETE / APPEND / SUBTRACT)
Throughput ceiling hit	`createPartitions` (⚠ breaks keyed apps)

Kafka still has no backup command, but Ch. 5 adds two genuinely useful things:

Configuration is recoverable from any running broker. describeConfigs turns every live broker into a config backup. Consider dumping it to version control on a schedule — cheap insurance, per the war story.
Deletion is not recoverable. deleteTopics is final; deleteRecords is one-way. The only protections are delete.topic.enable=false and your own tooling discipline.

And one anti-pattern to retire: do not back up or restore Kafka state by touching ZooKeeper. That path is being removed.

Your situation	How to call it
Admin operations are rare	Blocking `get()` is fine
You're building an admin service	`whenComplete()` + per-call `timeoutMs`
Admin ops on the critical path	Low timeout + a degraded fallback
Many resources to inspect	Batch them — `describeConfigs` accepts multiple resources of multiple types
Lists you mutate concurrently	`APPEND` / `SUBTRACT`, not `SET`