Why does the first Gremlin query take significantly more time in distributed JanusGraph setup on Kubernetes?

Environment Setup

I'm working with a distributed JanusGraph architecture deployed on Azure Kubernetes Service (AKS):

Infrastructure:

• AKS Cluster: 2 nodes (16 vCPU, 64 GB RAM each)
• Cassandra: 2 replicas with sharding enabled (Kubernetes pods)
• Elasticsearch: 2 replicas with sharding enabled (Kubernetes pods)
• JanusGraph: Single replica connected to both backends (Kubernetes pod)
• Mixed index: Created on title and nt property keys

Connection Pool Implementation:

I've implemented thread-safe connection pooling where each keyspace has its cached traversal:

class BaseGremlinClass(View):
    _connections = {}
    _lock = threading.Lock()

    def get_traversal(self, keyspace_name):
        """Get or create a traversal for the given keyspace"""
        if keyspace_name not in settings.JANUSGRAPH_KEYSPACES:
            raise ValueError(f"Keyspace {keyspace_name} not found in settings")

        with self._lock:
            if keyspace_name not in self._connections:
                self._create_connection(keyspace_name)
            print("Getting Connection from Pool")
            return self._connections[keyspace_name]['traversal']

    def _create_connection(self, keyspace_name):
        """Create a new connection and traversal"""
        try:
            config = settings.JANUSGRAPH_KEYSPACES[keyspace_name]
            connection = DriverRemoteConnection(
                config['url'],
                config['graph'],
                message_serializer=serializer.GraphSONSerializersV3d0(),
                timeout=30,
                pool_size=10,
                max_workers=4,
            )
            traversal_g = traversal().withRemote(connection)
            self._connections[keyspace_name] = {
                'connection': connection,
                'traversal': traversal_g
            }
            logger.info(f"Created connection for keyspace {keyspace_name}")
        except Exception as e:
            logger.error(f"Error creating connection to {keyspace_name}: {e}")
            raise

class GremlinQueries:
    def __init__(self, keyspace_name='main'):
        traversal = BaseGremlinClass()
        self.g = traversal.get_traversal(keyspace_name).with_('evaluationTimeout', 0)
        self.keyspace_name = keyspace_name

    def get_all_nodes_label(self):
        """Return list of node types"""
        data = self.g.V().has('nt').values('nt').dedup().order().by(Order.asc).to_list()
        return data

Performance Issue

Despite having connection pooling, indexing, and sharding implemented, I'm observing:

1. First query execution: Takes significantly longer (e.g., 45 seconds)
2. Second query: Runs in almost half the time (~20-25 seconds)
3. Subsequent queries: Maintain the improved performance (i.e. maintains the performance of the 2nd run)

Questions

1. Why does the first Gremlin query take significantly longer than subsequent runs in a Kubernetes environment, even with connection pooling and indexing?
2. What Kubernetes-specific factors might be contributing to this cold start behavior?
3. What optimisations can be implemented to reduce the first-time latency in a containerised distributed setup?
4. Are there specific considerations for sharded Cassandra/Elasticsearch deployments on Kubernetes that could impact initial query performance?

What I've Tried

• Verified that connection pooling is working (connections are reused)
• Confirmed mixed indexes are properly created and being used
• Checked that subsequent queries with same/different parameters show improved performance
• Monitored that the connection pool prevents reconnection overhead