Skip to content

// data lakes and query engines · 13 October 2026 · 3 min read

Why your data lake isn’t as slow as you think

Across the AWS, GCP, and Azure deployments I’ve worked on, I keep running into the same assumption: “data lakes are cheap because you don’t access them often.” It’s a tidy explanation. It’s also wrong, and the mix-up leads people to the wrong conclusion about whether a lake will actually be fast enough for what they’re building.

Two different questions, collapsed into one

“Data lake” answers one question: where does the data live, and in what format. It means plain files sitting in object storage, usually in an open format like Parquet or Delta Lake. That’s the whole definition. It says nothing about speed.

“Hot, cool, archive” answers a completely different question: how often do you plan to touch this data. These are access tiers within the storage service itself, and the same split exists on every major cloud, not just Azure. You pick one per file, independent of whether you call the overall thing “a data lake.”

Two separate dials. Treating them as one is the whole myth.

The three tiers, and what each actually gets you

Hot is for data you access often. Highest storage cost per gigabyte, lowest access cost, and it’s the default: fast, no restore step, read it whenever.

Cool is for data you touch occasionally, maybe monthly. Cheaper to store, a small fee each time you read it, but still quick: no waiting required.

Archive is where things actually slow down. Microsoft’s own documentation puts first-byte latency in hours, not milliseconds. AWS’s equivalent, S3 Glacier, is explicit about it too: a standard Glacier retrieval typically takes 3 to 5 hours, and a bulk retrieval can take up to 12. A file on archive tier isn’t just “slower to read.” It isn’t readable at all until you’ve restored it first.

None of these three tiers is “the data lake tier.” A lake can sit entirely on hot. Most production ones do.

So why do lakes have a cheap reputation at all?

Not because the data goes untouched. Because storage and compute aren’t bundled together the way they are in a traditional database.

A database’s storage price has a standing compute engine baked into it, something running continuously, ready to answer a query the instant it arrives, with a low-latency guarantee on every single byte. You pay for that readiness whether anyone ever asks a question or not.

Object storage carries none of that. It’s disks holding bytes, with a simple read/write API, no compute reserved next to it, no latency promise attached to every file. That’s what makes it cheap, on any tier, hot included. It has nothing to do with how rarely you read it.

When archive is actually the right call

Archive tier earns its reputation honestly for certain workloads:

  • Old backups you’re keeping for compliance, not for reading. Retention requirements, not active use.
  • Superseded snapshots behind a version you already trust. You’d only reach for them to debug something, not to serve a live query.
  • Historical data nobody’s asking about right now. A customer’s early years of records, say, once only the recent ones come up in practice.

If you’re running something that answers questions in real time, a BI tool, a chat agent, anything with a live user on the other end, that data has no business sitting on archive. It belongs on hot, the same as it would in a database, and it’ll perform like it.

The actual mental model

Stop asking “is this a data lake, so it must be slow.” Ask two separate questions instead:

  • Where does the data live, and in what format? That’s the lake-versus-database decision, an architecture choice.
  • How often is this specific file actually accessed? That’s hot-versus-cool-versus-archive, a per-file cost decision, ideally made from real usage data rather than a guess.

A lake isn’t slow. A lake where something got put on the wrong tier for its actual access pattern is slow, and that’s not a property of being a lake. It’s a property of someone picking the wrong tier.

So: next time someone tells you a data lake is “obviously” the cheap-but-slow option, which of the two questions are they actually answering?

Muhammad Mairaj is a Senior Backend Engineer with 8 years building payment and data systems. He writes about data engineering, measured. More about him.