[fix][proxy] Avoid blocking the proxy IO thread on a cold broker cache

[merlimat]

Motivation

MetadataStoreCacheLoader.getAvailableBrokers() returns the cached active-broker list, but on a cold/empty cache it fell back to a synchronous getChildren() metadata read (getChildrenAsync(...).get(operationTimeoutMs)). Its clean List signature hid that it could block.

It is reached on the proxy's Netty IO thread via ProxyConnection.completeConnect() (when checkActiveBrokers is enabled) → isBrokerActive() → getAvailableBrokers(), and also from BrokerDiscoveryProvider.nextBroker(). A cold cache could therefore stall a proxy IO thread on a metadata round-trip during connection setup.

Modifications

getAvailableBrokers() no longer performs a synchronous metadata read. It returns the snapshot maintained by init() and the metadata-store children/session listeners, and — only when the snapshot is empty — triggers a single background refresh (guarded by an AtomicBoolean so a burst of calls during an empty-cache window does not queue many reads). The former tryUpdate lambda is extracted into a reloadBrokers() method shared by the listeners, the initial fetch, and the background refresh.

The fail-fast isBrokerActive check (which already expects the client to retry after a back-off) now sees the current snapshot instead of blocking. An empty snapshot means there are genuinely no active brokers, since the cache is populated by init() and kept current by the listener.

Verifying this change

Added MetadataStoreCacheLoaderTest, which verifies that getAvailableBrokers() serves the cache without ever invoking the synchronous getChildren(...), and does not block on an empty cache.

Does this pull request potentially affect one of the following parts:

If the box was checked, please highlight the changes

• Dependencies (add or upgrade a dependency)
• The public API
• The schema
• The default values of configurations
• The threading model
• The binary protocol
• The REST endpoints
• The admin CLI options
• The metrics
• Anything that affects deployment