PHOENIX-7965 :- ConnectionQueryServicesImpl.getMetaDataCache() returns null after connection close, causing a bare NPE during concurrent query compilation

[lokiore]

What changes were proposed in this pull request?

ConnectionQueryServicesImpl.getMetaDataCache() returned the volatile latestMetaData field directly, with no null guard:

public PMetaData getMetaDataCache() {
  return latestMetaData;
}

This PR adds a guard so a closed connection surfaces a descriptive exception instead of a bare NPE. The read is done once into a local so the null-check and the returned reference are guaranteed to be the same value (no time-of-check/time-of-use re-read window):

public PMetaData getMetaDataCache() {
  PMetaData cache = latestMetaData;
  if (cache == null) {
    throwConnectionClosedException();
  }
  return cache;
}

throwConnectionClosedException() is the same private helper that the existing throwConnectionClosedIfNullMetaData() (used by addTable, removeTable, pruneTables, the internal getTable path, etc.) delegates to — it throws IllegalStateException("Connection to the cluster is closed"). getMetaDataCache() was the one cache accessor missing this guard.

Why are the changes needed?

close() sets latestMetaData = null. When a connection is closed concurrently with query compilation (for example, connection-pool teardown or a cluster role transition), an in-flight compile reaches the cache via PhoenixConnection.getTableRef() / getFunction() / getSchema(), which call getMetaDataCache().getTableRef(key) (and friends) on the null return. The result is a bare, causeless NullPointerException with no message — difficult to attribute to a closed connection during triage.

Every other accessor/mutator on this class already guards this state and throws the descriptive IllegalStateException; getMetaDataCache() was inconsistent. This change makes the failure mode uniform and diagnosable.

Does this PR introduce any user-facing change?

No behavioral change on an open connection: when latestMetaData is present, the method returns exactly the same cache reference as before. The only change is the failure mode when the connection has been closed — an opaque NullPointerException is upgraded to a descriptive IllegalStateException("Connection to the cluster is closed"). No caller relies on a null return (every production caller immediately dereferences the result), so no caller is regressed.

How was this patch tested?

Added unit tests in ConnectionQueryServicesImplTest exercising a real ConnectionQueryServicesImpl instance (not a mock, so the real guard executes):

• testGetMetaDataCacheThrowsWhenClosed — reflectively nulls latestMetaData (closed-connection state) and asserts getMetaDataCache() throws IllegalStateException with message "Connection to the cluster is closed".
• testGetMetaDataCacheReturnsCacheWhenOpen — asserts the method returns the cache unchanged when present.

Both pass (Tests run: 2, Failures: 0, Errors: 0) via a reactor build (mvn -pl phoenix-core -am test). mvn spotless:apply reports the tree clean.

Was this patch authored or co-authored using generative AI tooling?

Generated-by: Claude Code (Opus 4.8 (1M context))