sql: use CachedReader for uses with sqlinstance and the sql builtin

The CachedReader won't block, which in multi-region clusters is good. It will
mean that in some cases, it'll state that a sessions is alive when it most
certainly is not. Currently, nobody needs synchronous semantics.

This is a major part of fixing the TestColdStartLatency as sometimes
distsql planning would block. That's not acceptable -- the idea that
query physical planning can need to wait for a cross-region RPC is
unacceptable.

Release note: None

pkg/server/server_sql.go

@@ -562,7 +562,7 @@ func newSQLServer(ctx context.Context, cfg sqlServerArgs) (*SQLServer, error) {
 		cfg.clock, cfg.rangeFeedFactory,
 	)
 	cfg.sqlInstanceReader = instancestorage.NewReader(
-		cfg.sqlInstanceStorage, cfg.sqlLivenessProvider.BlockingReader(),
+		cfg.sqlInstanceStorage, cfg.sqlLivenessProvider.CachedReader(),
 		cfg.stopper,
 	)
 
@@ -789,7 +789,7 @@ func newSQLServer(ctx context.Context, cfg sqlServerArgs) (*SQLServer, error) {
 		RowMetrics:         &rowMetrics,
 		InternalRowMetrics: &internalRowMetrics,
 
-		SQLLivenessReader: cfg.sqlLivenessProvider.BlockingReader(),
+		SQLLivenessReader: cfg.sqlLivenessProvider.CachedReader(),
 		JobRegistry:       jobRegistry,
 		Gossip:            cfg.gossip,
 		PodNodeDialer:     cfg.podNodeDialer,

pkg/sql/planner.go

@@ -114,7 +114,7 @@ func (evalCtx *extendedEvalContext) copyFromExecCfg(execCfg *ExecutorConfig) {
 	evalCtx.Settings = execCfg.Settings
 	evalCtx.Codec = execCfg.Codec
 	evalCtx.Tracer = execCfg.AmbientCtx.Tracer
-	evalCtx.SQLLivenessReader = execCfg.SQLLiveness.BlockingReader()
+	evalCtx.SQLLivenessReader = execCfg.SQLLiveness.CachedReader()
 	evalCtx.CompactEngineSpan = execCfg.CompactEngineSpanFunc
 	evalCtx.SetCompactionConcurrency = execCfg.CompactionConcurrencyFunc
 	evalCtx.TestingKnobs = execCfg.EvalContextTestingKnobs