Kafka 4.0 for AI Call Analytics Fan-Out: KRaft, Share Groups, Eligible Leader Replicas

TL;DR — Kafka 4.0 is the first major release with zero ZooKeeper (KRaft default), KIP-848 next-gen consumer rebalance GA, KIP-932 share groups (queue semantics on a log), and KIP-966 Eligible Leader Replicas in preview. For AI call analytics, this is the difference between a 30-second rebalance stall and an invisible reassignment.

The pattern

When an AI voice agent finishes a call, ten downstream systems want to know: billing, CRM, the embedding pipeline, the QA dashboard, the recording archiver, the ML training set, the affiliate attribution job, the SLA monitor, the fraud detector, and the founder's Slack. Kafka is the canonical fan-out backbone: produce once, consume independently, replay on demand.

Kafka 4.0 (March 2025, mainline through 2026) closed the last operator-pain gaps: KRaft is default, KIP-848 makes consumer rebalances incremental rather than stop-the-world, and ELR keeps a min.in-sync subset that's safe to elect even when the ISR shrinks.

How it works (architecture)

flowchart LR
  Voice[Voice agent] -->|produce call.completed| K[(Kafka 4.0<br/>KRaft cluster<br/>3 brokers)]
  K --> CG1[Consumer group: billing]
  K --> CG2[Consumer group: crm-sync]
  K --> CG3[Consumer group: embeddings]
  K --> CG4[Consumer group: qa-dashboard]
  K --> CG5[Share group: ml-training]
  CG3 --> Vec[(Vector DB)]
  CG2 --> CRM[(HubSpot/Salesforce)]
  CG5 --> Train[Training pipeline]

Each consumer group reads independently with its own offset. Share groups (KIP-932) layer queue semantics on top of the log so workloads that don't want partition-pinned ordering get cooperative consumption.

CallSphere implementation

Explore a live demo and compare current plans to find the right fit for your business.

Build steps with code

1. Bootstrap KRaft cluster: bin/kafka-storage.sh format -t $UUID -c kraft.properties.
2. Create the topic with proper retention: bin/kafka-topics.sh --create --topic call.completed --partitions 24 --replication-factor 3 --config retention.ms=604800000.
3. Set min.insync.replicas=2 and acks=all on producers for ELR safety.
4. Use the new consumer group protocol: group.protocol=consumer (KIP-848).
5. Schema-registry the payload (CloudEvents 1.0 envelope, see post #12).
6. Wire a Kafka Connect sink for ClickHouse and S3 archival.
7. Monitor under-replicated partitions and consumer lag in Grafana.

from confluent_kafka import Producer, Consumer

p = Producer({
    "bootstrap.servers": "kafka1:9092,kafka2:9092,kafka3:9092",
    "acks": "all",
    "enable.idempotence": True,
    "compression.type": "zstd",
})

p.produce(
    topic="call.completed",
    key=call_id.encode(),
    value=cloudevent_envelope_json,
    headers=[("ce-type", b"com.callsphere.call.completed.v1")],
)
p.flush()

c = Consumer({
    "bootstrap.servers": "kafka1:9092",
    "group.id": "embeddings",
    "group.protocol": "consumer",   # KIP-848 next-gen
    "auto.offset.reset": "earliest",
})
c.subscribe(["call.completed"])
while True:
    msg = c.poll(1.0)
    if msg and not msg.error():
        upsert_embedding(msg.value())
        c.commit(msg)

Common pitfalls

• Partition count too low — 6 partitions caps you at 6 parallel consumers per group; pick partitions for your peak parallelism, not your current load.
• Acks=1 + idempotence=false — silent data loss on broker failover.
• One topic per event type explosion — group related events under call.* topics with header-based filtering downstream.
• Skipping the new consumer protocol — old protocol still works but rebalances stall on every deploy.
• Treating retention.ms as backup — Kafka is a buffer, not a database; tier to S3 with KIP-405.

FAQ

Do we still need ZooKeeper? No. Kafka 4.0 only runs in KRaft mode. Migrating clusters: do the migration on 3.x first, then upgrade to 4.0.

Share groups vs consumer groups? Share groups (KIP-932) are queue-like — multiple consumers ack individual messages, no partition pinning. Use for unordered work like ML inference.

When are ELRs GA? Preview in 4.0, expected GA in a 4.x release. Until then, treat them as a safety preview.

How does CallSphere price this? Kafka costs land in our infra, not customer pricing — see /pricing for $49/$99/$149 plans.

Can I demo a Kafka-fed analytics dashboard? Yes — book a demo.

Sources

• Apache Kafka 4.0.0 Release Announcement
• Confluent: Kafka 4.0 — Default KRaft, Queues, Faster Rebalances
• Apache Kafka 4.2.0 Release Announcement
• Kafka 4.0 & KRaft: The End of ZooKeeper