Tiered Storage

Tiered storage is an architectural pattern that organizes data across multiple storage media — typically fast, expensive hot storage (SSD, NVMe) for recent or frequently accessed data, and slow, cheap cold storage (HDD, tape, cloud object storage) for older or rarely accessed data. The tiering is usually transparent to applications: data is promoted or demoted between tiers based on access patterns, age, or administrative policy, without requiring application-level changes.

The economic logic of tiered storage is compelling. Hot storage can cost 50-100x more per gigabyte than cold object storage. A system that retains all data on hot storage is paying premium prices for data that is never accessed. Conversely, a system that archives everything to cold storage imposes unacceptable latency on queries that need recent data. Tiered storage is the compromise: it accepts the complexity of managing multiple storage classes in exchange for cost optimization.

In distributed systems, tiered storage has become particularly important for messaging and streaming platforms. Apache Pulsar, Apache Kafka, and cloud-native data lakes all implement tiered storage to enable long-term retention of event streams without the cost of keeping every message on replicated SSD. The boundary between hot and cold is increasingly dynamic: machine learning models predict which data will be accessed and pre-promote it, blurring the line between tiering and caching.

The operational risk of tiered storage is retrieval latency unpredictability. A query that hits hot storage completes in milliseconds; the same query hitting cold storage may take seconds or minutes. Applications that are not designed for this latency spread can experience timeout cascades and degraded user experience. Tiered storage is not merely a storage decision; it is a service-level agreement that the application architecture must accommodate.