Cold storage in cloud storage is a class of low-cost storage tiers for data that is written once and read rarely, in which the provider charges much less per gigabyte-month than for frequently accessed data but adds retrieval fees, minimum storage durations and, in the coldest variants, delayed access. It sits between hot object storage and archive storage on the temperature scale of storage classes.
Typical cold tiers include Amazon S3 Standard-Infrequent Access and Glacier Instant Retrieval, Azure Blob Storage Cool and Cold, and Google Cloud Storage Nearline and Coldline. The term describes an access pattern and pricing model, not a different kind of hardware visible to the user, and it is not to be confused with offline cryptocurrency wallets, which share the name.
How cold storage pricing and access work
Cold storage works by shifting cost from capacity to access. The provider stores the data with the same high durability as hot tiers, typically designed for eleven nines of annual object durability, but places it on media and in layouts optimized for density rather than request throughput.
Four price components define every cold tier: the per-gigabyte storage price, the per-gigabyte retrieval fee, request charges that are higher than for hot tiers, and a minimum storage duration that bills early deletions as if the object had stayed for the full period. Many tiers also bill small objects as if they were at least 128 KB.
| Component | Compared with hot storage | Practical consequence |
|---|---|---|
| Capacity price | Substantially lower | Large datasets become affordable |
| Retrieval fee | Charged per GB read | Frequent reads erase the savings |
| Request charges | Higher per operation | Many small objects cost more |
| Minimum duration | Typically 30 to 90 days | Early deletion is billed anyway |
| Minimum object size | Often 128 KB billed | Tiny files are inefficient |
Cold, archive and hot tiers compared
The temperature scale runs from hot for active data through cool or cold for occasional access to archive for data that may never be read. The key difference between cold and archive storage is retrieval latency: cold tiers return data in milliseconds, while archive tiers usually require a restore operation lasting minutes to many hours.
| Provider | Cold tiers | Minimum duration | First-byte latency |
|---|---|---|---|
| Amazon S3 | Standard-IA, Glacier Instant Retrieval | 30 and 90 days | Milliseconds |
| Azure Blob Storage | Cool, Cold | 30 and 90 days | Milliseconds |
| Google Cloud Storage | Nearline, Coldline | 30 and 90 days | Milliseconds |
Archive classes such as S3 Glacier Deep Archive, Azure Archive and Google Archive extend the minimum to 180 or 365 days. Some providers also market a mid-way option, such as S3 Glacier Flexible Retrieval, which is priced near archive but restores in minutes to hours. Exact prices change and differ by region, so designs should be based on current price sheets rather than remembered numbers.
| Access pattern | Suitable tier | Reason |
|---|---|---|
| Several reads per day | Hot or standard | No retrieval fees, lowest request cost |
| About once a month | Cool or infrequent access | Savings exceed occasional retrieval cost |
| About once a quarter | Cold | Low capacity cost, instant access kept |
| Once a year or never | Archive | Lowest capacity cost, restore delay accepted |
| Unknown or changing | Automatic tiering | Provider moves objects by observed use |
Moving data into cold storage
Data reaches cold storage in three ways: by writing it directly with the cold storage class, by a storage lifecycle policy that transitions objects after a set age, or by automatic tiering that moves objects based on observed access. Lifecycle policies are the most common because they need no change in the application.
A typical policy keeps objects hot for 30 days, moves them to a cold tier until day 180, then to archive, and deletes them after the retention period required by policy or regulation. Transitions are themselves billed per object, so millions of small files can make a transition more expensive than leaving them in place.
- Aggregate small files into larger archives such as tar or Parquet before tiering.
- Combine tiering with object versioning carefully, because noncurrent versions need their own rules.
- Tag objects by data class so lifecycle rules can target backups, logs and media separately.
- Model total cost including retrieval, not just the monthly capacity price.
Reading cold data back does not require a restore step: the application issues an ordinary GET request and the provider adds the retrieval charge to the bill. This transparency is convenient but also dangerous, because a backup verification job or a search indexer that scans an entire bucket can trigger retrieval fees for every object without anyone noticing until the invoice arrives. Cost alerts and access logs on cold buckets are therefore a sensible default safeguard.
Typical workloads and design pitfalls
Cold storage suits data with predictable, infrequent reads and long life: backup copies, older log files, finished media projects, medical images past their active period and compliance records. It is a poor fit for data whose access pattern is unknown or spiky, because a single bulk restore can cost more than months of savings.
| Mistake | Effect | Better approach |
|---|---|---|
| Tiering active analytics data | Retrieval fees exceed savings | Keep queried data hot |
| Deleting objects within the minimum | Early deletion charges | Align retention with duration rules |
| Tiering millions of tiny objects | Transition and request costs dominate | Bundle files first |
| Confusing cold with archive | Unexpected hours-long restores | Check latency per class |
| Ignoring egress fees | Restores to another region are costly | Restore within the same region |
Cold tiers sometimes advertise a slightly lower availability target than hot tiers, for example 99.9 percent instead of 99.99 percent. Durability remains equivalent, so the risk is temporary unavailability, not data loss.
The idea in brief
Cold storage is like renting a storage unit outside the city instead of keeping everything in the apartment. The space is much cheaper, and the boxes are just as safe there.
The catch is that each trip to fetch something costs a fee, and the agreement requires paying for at least a minimum period. That is why cold storage is ideal for things that must be kept but are rarely needed, such as old backups or records, and a poor choice for files used every day.
A practical scenario
A regional hospital network keeps diagnostic images for many years. Images from the last 60 days are opened often by radiologists, so they stay in the standard storage class. A lifecycle rule then moves each study to a cold tier with millisecond access, where a doctor can still open an older scan during a follow-up visit without waiting.
After two years a second rule moves studies to archive storage, accepting a restore delay of several hours for rare legal or research requests. The capacity bill for older studies falls substantially, while clinicians notice no change in how quickly recent and follow-up images open.