The Definitive Guide

What Is Diskless Kafka?

Diskless Kafka removes local broker disks entirely and writes data straight to object storage. It's how modern teams cut Kafka costs by 10x while getting rid of operational overhead.

What Is Diskless Kafka?

The old way: Traditional Kafka writes data to local SSDs on each broker, then replicates cross-zone to followers. In the cloud, 70-90% of a Kafka cluster's cost is just moving bytes between availability zones.

The new way: Diskless Kafka writes directly to cloud object storage (S3, GCS, Azure Blob). The compute layer is fully stateless. No disks, no rebalancing, no babysitting. The object store handles durability.

The result: A system that costs a fraction of what Kafka costs, runs with near-zero operational overhead, and scales without limits.
10x
cheaper than Kafka
0
disks to manage
33ms
median produce latency

How Diskless Kafka Works

Stateless agents replace traditional Kafka brokers. Each agent is a lightweight Go binary that speaks the Apache Kafka protocol but holds no data on disk.

1
Buffer
Agent collects records in memory for a short window (as low as 25ms).
2
Write to S3
Batch uploads directly to object storage as a single file. Nothing touches a local disk.
3
Sequence
Control plane assigns offsets and commits the file to the metadata store.
4
Acknowledge
Producer gets confirmation. Data stored with 99.999999999% durability.
Architecture Diagram: Producer → Agent (stateless) → S3/GCS → Control Plane (metadata) → Acknowledge

On the read side, consumers pull data from object storage through agents. WarpStream's service discovery ensures consumers always talk to an agent in the same availability zone — zero cross-zone costs on reads too.

Why stateless matters

IC
Trivial scaling
Add or remove agents like web servers. No data needs to move.
IC
No hotspots
Load spreads evenly regardless of partition assignment.
IC
Instant recovery
Agent down? Next request goes to another agent. No leader election.
IC
Scale to zero
No traffic, no cost. No idle brokers burning money.

Why Disks Are The Problem

Kafka was designed for LinkedIn's data centers in 2011. Cloud flips every one of those assumptions.

$136K
per month
Inter-AZ networking at 1 GiB/s. Every GiB crossing a zone boundary costs ~$0.02.
18x
storage markup
Triple-replicated EBS: $0.39/GiB/mo. S3: $0.021/GiB/mo. You do the math.
24/7
ops overhead
Disk failures, rebalancing, capacity planning, volume resizing. Needs a dedicated team.
Root cause: Kafka ties compute and storage together on the same nodes. You can't scale them independently, can't auto-scale without risking data loss, and can't recover from a failure without waiting for data to move around. Diskless Kafka separates them completely.

The Economics of Going Diskless

This isn't a marginal improvement. It's a structural cost difference.

At 1 GiB/s sustained write throughput, 7 days retention:

Cost ComponentSelf-Hosted KafkaWarpStream (Diskless)
Inter-AZ networking~$136,000/mo$0
Storage~$30,000/mo (3x replicated EBS)~$12,600/mo (S3, single copy)
Compute~$15,000/mo (dedicated brokers)~$5,000/mo (stateless, auto-scaled)
S3 API costs$0~$3,500/mo
Total~$181,000/mo~$21,100/mo
Savings~8.6x cheaper

Where the savings come from

No inter-AZ networking
Data flows within a single AZ. The object store replicates behind the scenes for free.
Object storage pricing
S3 at $0.021/GiB/mo vs triple-replicated EBS at $0.39/GiB/mo. 18x difference.
No over-provisioning
Stateless agents auto-scale on CPU and network. No sizing for peak. No idle brokers.
WarpStream pricing: starts at $0.01/GiB for write throughput. No per-agent fees. No per-partition fees. No per-core fees.

Diskless vs Tiered Storage

Tiered storage (KIP-405) keeps hot data on local broker disks and offloads cold data to S3. Reduces long-retention costs, but brokers are still stateful and you still pay for cross-zone replication.

Diskless puts everything in object storage from the start. No local disks at all. No inter-zone replication. No partition leaders. A ground-up rethink, not a bolt-on optimization.

Diskless (WarpStream)Tiered Storage (KIP-405)Traditional Kafka
Hot data locationObject storageLocal disksLocal disks
Inter-AZ replicationNoneStill required for hot dataAll data cross-zone
Broker stateStatelessStatefulStateful
Partition rebalancingNot neededStill requiredRequired
Auto-scalingTrivial (add containers)Complex (data migration)Complex (data migration)
Disk managementNoneReduced, not eliminatedExtensive
Cost at scale5-10x cheaper~2x cheaperBaseline
Bottom line: Tiered storage reduces storage costs. Diskless eliminates the entire category of problems — networking, ops, rebalancing, disk management — at 5-10x the savings.

Latency: The Trade-Off That's Over

The old criticism

Writing to object storage takes longer than local SSD. In WarpStream's default config, produce P99 is ~400ms vs single-digit ms for Kafka on NVMe.

What changed

Two breakthroughs closed the gap: S3 Express One Zone (105ms median) and Lightning Topics (decoupled durability from sequencing).

S3
S3 Express One Zone
Significantly faster than standard S3. WarpStream on S3 Express: 105ms median, 170ms P99.
LT
Lightning Topics (Feb 2026)
Durability happens upfront; sequencing happens after the ack. Based on the LazyLog paper.

Combined result:

33ms
median produce latency
<50ms
P99 produce latency
$0
cost increase
For logging, observability, analytics, CDC, ETL, and most production workloads — sub-50ms is more than fast enough. Only HFT and real-time bidding genuinely need single-digit-ms. For everything else, the trade-off is over.

History

Born from WarpStream's founders' experience at Datadog, where they built Husky — a columnar database running directly on S3. The question: what would Kafka look like if you built it from scratch for the cloud, with zero local disks?
July 2023
WarpStream launches as the first Kafka protocol-compatible platform built entirely on object storage.
November 2024
Confluent (founded by Apache Kafka's original creators) acquires WarpStream, validating diskless as the future.
March 2025
Apache Kafka community proposes KIP-1150: Diskless Topics, bringing native diskless support to upstream Kafka.
January 2026
WarpStream Tableflow goes GA — materializing Kafka topics as Apache Iceberg tables using the diskless architecture.
February 2026
Lightning Topics ship. 33ms median produce latency on S3 Express. The latency gap with traditional Kafka effectively closes.

KIP-1150: Native Diskless Topics in Apache Kafka

KIP-1150 proposes adding diskless topic support directly to Apache Kafka — strong validation. But there are key differences:

WarpStream (production since 2023)
  • Fully stateless agents — no local disks at all
  • Proven at scale at Robinhood, Character.AI, Goldsky, Grafana Labs, Cursor, Zomato, and more
  • BYOC model — runs in your VPC, zero cross-account IAM access
  • 33ms latency with Lightning Topics + S3 Express
KIP-1150 (proposal, in voting)
  • Brokers still stateful — local disks used as write cache
  • Partition leaders still required — brokers manage leader election
  • Not yet production software — currently in voting status
  • Self-managed — you operate the cluster
The Apache Kafka community moving toward native diskless support confirms that the industry sees this as the future. WarpStream is that future, available today.

Get Started with Diskless Kafka

WarpStream speaks the Apache Kafka protocol. No rewrites. No proprietary SDKs. Change the bootstrap URL and you're streaming.

$ curl https://console.warpstream.com/install.sh | bash
$ warpstream demo
1
Sign up
console.warpstream.com — $400 in free credits, no credit card.
2
Deploy agents
Docker, Kubernetes, ECS. Any container runtime in your VPC.
3
Point clients
Change the bootstrap URL in your Kafka client. Done.
Works on AWS, GCP, Azure, Oracle Cloud, DigitalOcean, Vultr, Cloudflare R2, MinIO, and anything with an S3-compatible object store.
Trusted in production by
Robinhood
Character.AI
Cursor
PostHog
Grafana Labs
Goldsky
Zomato
DISH

The trade-off is over. Diskless Kafka at 33ms.

Kafka compatibility, 10x lower costs, zero operational overhead, and sub-50ms latency.

Get Started Free ($400 credits)Book a Demo

Frequently Asked Questions

What is Diskless Kafka?
A streaming architecture that removes local broker disks entirely. Data writes go straight to cloud object storage (S3, GCS, Azure Blob). WarpStream pioneered the approach in 2023 as the first Kafka protocol-compatible platform with zero local disks.
How does Diskless Kafka work?
Stateless agents replace Kafka brokers. Producers send data to an agent, which writes it directly to object storage. A separate metadata store handles offset sequencing. Any agent can serve any partition. No leaders, no rebalancing, no disk management.
What's the difference between diskless and tiered storage?
Tiered storage (KIP-405) offloads cold data to object storage but keeps hot data on local disks. Diskless puts everything in object storage from the start. That means no inter-zone replication costs, no disk management, fully stateless brokers, and 5-10x cost savings instead of roughly 2x.
Is Diskless Kafka slower?
In the default configuration, WarpStream P99 is about 400ms. With Lightning Topics and S3 Express One Zone, that drops to 33ms median and under 50ms P99. Good enough for logging, analytics, CDC, ETL, and most production workloads.
How much cheaper is it?
Typically 5-10x cheaper than Kafka in the cloud. Savings come from eliminating inter-AZ networking (70-90% of Kafka costs at scale), replacing triple-replicated SSDs with object storage, and auto-scaling stateless compute instead of over-provisioning brokers.
Is WarpStream compatible with my Kafka clients?
Yes. WarpStream speaks the Apache Kafka protocol. Keep your existing client libraries (librdkafka, confluent-kafka-go, kafka-python, etc.) and just change the bootstrap server URL.
Does my data leave my cloud account?
No. WarpStream's BYOC model runs agents in your VPC, stores data in your object storage buckets, and requires zero cross-account IAM access. The only thing WarpStream sees is metadata. Raw data never leaves your environment.