Tiered storage lets a streaming broker keep hot data on local SSD and cold data on object storage (S3, GCS, Azure Blob). Redpanda's implementation is among the most polished — same Kafka API, infinite retention, ~50% cheaper than provisioning enough SSD.
The cost equation
EBS SSD: $0.10/GB/month. S3 Standard: $0.023/GB/month. For 100 TB topic retention: $10K/mo SSD vs $2.3K/mo S3. Tiered storage = pay SSD price for recent days, S3 price for older. Saves 60-80% for log topics with long retention.
How it works
Active segment: written to local disk. Closed segments older than threshold: uploaded to S3 asynchronously. Consumer reading old data: broker streams from S3 transparently. No application change needed.
Read latency tradeoff
Hot reads (last N hours): same as local. Cold reads (from S3): ~50-200ms first byte. For replay or backfill workloads this is fine; for real-time consumers it's invisible because they read hot data.
Configuration
# Cluster config
cloud_storage_enabled: true
cloud_storage_credentials_source: aws_instance_metadata
cloud_storage_bucket: my-redpanda-tier
cloud_storage_region: us-east-1
cloud_storage_segment_max_upload_interval_sec: 3600
retention_local_target_bytes_default: 53687091200 # 50GB on localDisaster recovery
Cluster lost entirely? Recover topics from S3 — just point a fresh cluster at the same bucket. Recovery time: minutes to hours for big topics. This makes 'lost cluster' a tolerable failure, not a catastrophe.