Ever watched a laptop crawl to a stop halfway through a training run?
Nine times out of ten it isn’t your processor. Neither is it your RAM. More often than not it’s your storage. A single drive spinning your OS, paging file and swap file and none of them particularly well.
Here’s the good news:
It’s not more drive capacity. It’s not a fast drive. It’s not a more expensive drive. The fix is storage tiering, and it’s typically cheaper than most people think.
What this guide covers:
- Why single-drive setups fall apart
- What storage tiering actually means
- The three tiers and what belongs in each
- How to move data between tiers safely
- Tiering mistakes that quietly waste money
Why One Drive Always Runs Out Of Road
Data doesn’t sit still. It multiplies.
Projects leave behind raw exports, cleaned copies, feature tables, model checkpoints and several experiment folders that no one ever deleted. Worldwide creation of data is forecast to triple from 2025 to 2029. That stress hits everyday computers years before it hits a data center.
A 2TB SSD handles all of it… for a while. Then it doesn’t.
But here’s the part that really stings…
Storage ceased declining in price. NAND flash contract prices increased nearly 4x through three quarters. “Buy another drive” is no longer the inexpensive solution.
Data storage is why storage planning shows up so far in advance for a data science major. Anyone pursuing a Data Science Bachelor’s Degree soon learns that modelling is the smallest part of the job — most of it is moving, cleaning, versioning and storing datasets that absolutely refuse to be tidy. When you understand where data should live as a data science major you will ship projects quicker than someone who just knows algorithms.
Storage Tiering: What It Actually Means
Storage tiering is easy. Store different data on different tiers based on access frequency.
Hot data — files you’re constantly reading from and writing to — goes on the fastest storage you have. Warm data goes on slower, cheaper storage. Cold data goes on something very cheap and very slow since you rarely access it.
Imagine your kitchen. Knives belong on the counter. The stand mixer belongs in a cupboard. The large roasting tray belongs in the garage until December.
No one leaves the roasting tray on the counter. But leave it there is essentially what a single-drive configuration does with every dataset ever downloaded.
The Hot Tier: Speed Where It Counts
The hot tier is the working set — whatever is open right now.
That means:
- The dataset currently being trained on
- Active notebooks and scripts
- Databases being queried live
- Checkpoints from the current run
This tier should be NVMe SSD. Random read performance is much more important than sheer size since most analytic tasks involve many small reads rather than a single large sequential read.
The simple rule of thumb: Make your hot tier small and make it quick. Once half your drive is allocated to files no one has accessed in the last month, it ceases to be a hot tier. It becomes an overpriced junk drawer.
The Warm Tier: The Middle Ground Nobody Plans For
Warm data is the awkward middle child of any workflow.
It’s no longer used, but it’s needed next week. Last month’s cleaned data. The previous version of that model. Those temporary files that took 6 hours to create and 6 more hours to re-create.
Warm data should reside on high capacity hard drives, a network share, or regular cloud object storage. While there is a noticeable drop in access speeds, it hardly matters when accessing data weekly.
And this is the tier where you save the most money. Studies have found that 60-70% of unstructured data stored in an enterprise has not been accessed in the last 90 days… and much of it is sitting on high-cost flash storage just collecting dust.
The Cold Tier: Cheap, Slow And Perfectly Fine
Cold storage is data that needs to be retained but rarely accessed.
Archived completed projects. Source files saved to allow recreation of the project. Documentation of compliance. Archived datasets from past competitions that you may want to reference.
Cold tiers are extremely inexpensive per TB. The downside is retrieval. It can be slow, and some cloud archive tiers charge to retrieve data again.
Which leads to a mistake worth avoiding…
Teams throw lots of stuff into cold storage to save money only to realize they access it all the time. An industry survey showed only 19% of data in cloud object storage was truly cold. Guess wrong one way or the other and you spend money either way.
Moving Data Between Tiers Without Losing It
A tiering system falls apart the second nobody can find anything. Structure prevents that.
Begin with a naming convention that holds up after reality strikes. project_stage_date will always trump final_2_FINAL_actually_FINAL.
Then set rules instead of relying on memory:
- Anything untouched for 30 days moves from hot to warm
- Anything untouched for 6 months moves from warm to cold
- Raw source data always keeps one cold copy — no exceptions
Maintain a simple index file of what resides where. Use a spreadsheet if you like. The goal isn’t elegance, it’s that you’ll be able to find a dataset some time in the future without hooking up four drives and hoping you guess correctly.
Automate the movement if at all possible. Lifecycle policies are native to cloud platforms. Scheduled scripts on premise while people sleep.
Tiering Mistakes That Quietly Waste Money
The same few patterns show up over and over again:
Treating backups as a tier. They’re not. Backups are a second copy of live data. Cold storage is frequently the only copy of dead data. Mixing them up leads to disasters.
Buying too much flash. Enterprise SSDs are still costing 18.6x more per TB than hard drives today. Throwing archives at premium storage is expensive.
Never deleting anything. Some things don’t warrant a tier. If it’s an intermediate file that can be recreated in five minutes, delete it.
Lossy compression. Moving from CSV to Parquet often reduces the size of your dataset significantly and allows you to read faster simultaneously. Bonus!
Bringing It All Together
Storage tiering is not a concept that flows down from the large data centre to smaller installations. It’s a working practice that scales from a single laptop to a thousand-node cluster unchanged.
A quick recap:
- Hot data on fast NVMe, kept deliberately small
- Warm data on cheap, high-capacity drives
- Cold data archived, indexed and left alone
- Clear rules for when files move between tiers
- Ruthless deletion of anything that can be rebuilt
Workflows exceed a single SSD’s capacity because data expands while drives do not. Tiering embraces this fact rather than battling it — and turns a costly storage issue into an ordinary filing issue, which is far easier to manage.
