Object Storage & Blob Storage • Storage Tiering (Hot/Warm/Cold)Easy⏱️ ~2 min
What is Storage Tiering and How Does it Differ from Caching?
Storage tiering is the practice of placing data across multiple storage media with different cost and performance characteristics based on access frequency. The system measures how often data is accessed (its temperature) and physically moves it between tiers: Hot tiers use expensive, fast media like NVMe Solid State Drives (SSDs) delivering 100,000 to 1,000,000 Input/Output Operations Per Second (IOPS) with 100 microsecond to 1 millisecond latency. Warm tiers use capacity optimized SSDs or Hard Disk Drives (HDDs) with roughly 5 to 10 millisecond seek times. Cold tiers minimize cost per gigabyte per month using erasure coded object stores, accepting tens to hundreds of milliseconds latency for online classes or minutes to hours for offline archives.
The critical distinction from caching is that tiering moves the canonical copy of data. A file starts on hot storage, and after 30 days of no access, the system physically relocates it to warm storage, then eventually to cold. The hot tier no longer holds that data. With caching, the authoritative copy stays in one place (often a capacity efficient store) while a fast tier keeps replicas of frequently accessed items. Evicting from cache does not affect the source of truth.
Production systems commonly combine both approaches. Amazon S3 might store the gold copy in Standard class, while an application maintains a Redis cache in front for microsecond reads. Separately, S3 Intelligent Tiering moves objects between Standard and Infrequent Access classes based on observed patterns. The cache provides latency wins without moving data, while tiering cuts storage bills by relocating rarely touched objects.
The goal is economic: keep the small fraction of data driving most requests (often under 10 percent) on expensive fast media and push the long tail to cheaper tiers. A log ingestion system handling 1 terabyte per day might keep 7 days hot, 23 days warm, and 150 days in cold archive, reducing storage cost by 40 to 70 percent compared to keeping all 180 days on hot storage, assuming cold data is read less than 1 percent of the time monthly.
💡 Key Takeaways
•Hot tier delivers 100,000 to 1,000,000 IOPS with sub millisecond latency using NVMe SSDs, serving the active working set (often under 10 percent of total data)
•Warm tier balances cost and performance with 5 to 10 millisecond latency using capacity optimized SSDs or HDDs, suitable for interactive but non critical access
•Cold tier minimizes dollars per gigabyte per month with erasure coding, accepting tens to hundreds of milliseconds for online classes or minutes to hours for offline archives like AWS Glacier Flexible Retrieval (3 to 5 hours standard rehydration)
•Tiering moves the canonical copy along a lifecycle (Hot to Warm to Cold), while caching keeps a fast replica with the gold copy unchanged; production systems combine both for latency wins and cost reduction
•Real world log pipelines handling 1 terabyte per day often keep 7 days hot, 23 days warm, and 150 days in cold archive, achieving 40 to 70 percent storage cost reduction versus all hot storage
•Policy decisions depend on measured access patterns (recency, frequency), Service Level Objectives (SLOs), and Total Cost of Ownership (TCO); misclassifying frequently accessed data to cold tiers causes bill shock from high retrieval fees
📌 Examples
AWS S3 Standard (hot) with 99.99% availability and millisecond access versus Glacier Flexible Retrieval (cold) with 3 to 5 hour Standard rehydration and 68% lower storage cost
Elasticsearch hot nodes on NVMe delivering single digit millisecond queries versus frozen tier using searchable snapshots in object storage with up to 90% cost reduction but higher tail latency
Azure Blob Storage Cool tier with 30 day minimum duration and higher access costs versus Archive with 180 day minimum requiring hours of rehydration before read