SysPad › Lessons › database lessons › Timestream

How Timestream works

Serverless time-series database. Recent data lives in a fast memory store, historical data tiers to cheap magnetic storage. Writes billed per record; queries billed per GB scanned. Ideal for IoT and metrics.

Time-series data has a peculiar shape: a firehose of timestamped points, where last week’s data is queried constantly and last year’s almost never. Timestream is built around that reality, it keeps recent data in fast memory and quietly slides older data to cheap storage, so you’re not paying RAM prices to hold a year of sensor readings you rarely touch.

What you will learn

What makes time-series data different enough to need its own database?
How does automatic tiering cut cost without you moving any data yourself?
Why does a single query still see both recent and ancient data?
When is a general-purpose database the wrong choice for metrics?

The shape of time-series data

Metrics, IoT readings, and events share a pattern: huge write volume, always append-only, always stamped with a time, and queried by time ranges("average CPU over the last hour"). Recent data is hot; old data is cold but worth keeping.

Automatic tiering: fast now, cheap later

Timestream keeps two tiers: a memory store for recent data (fast writes and queries) and a magnetic store for history (much cheaper). You set how long data lives in each, and Timestream moves it between them for you.

No archival jobs, no manual exports, the lifecycle is a configuration value, not code you maintain.

⚓ IoT fleets, application and infrastructure metrics, clickstreams, industrial telemetry.

One query, both tiers

You query with SQL that includes time-series functions (interpolation, smoothing, time-bucketed aggregation). A query transparently spans the memory and magnetic tiers, you don’t target one or the other.

Serverless, and when not to use it

Timestream (for LiveAnalytics) is serverless: no instances to size, it scales with your ingest and query load and bills for what you use, ideal for spiky or growing telemetry.

Availability note: as of 2025 AWS closed Timestream for LiveAnalytics to new customers. The concepts here (tiering, time-series querying) still apply to time-series databases generally; check current AWS availability before building on it.

Flashcards

What access pattern defines time-series data?: High-volume, append-only, time-stamped writes, queried by time range, with recent data hot and old data cold.
How does Timestream hold a year of data without paying RAM prices for all of it?: Two tiers: a fast memory store for recent data and a cheap magnetic store for history. It auto-moves data between them by age.
What job do you write to archive aging Timestream data to cheaper storage?: None. Tiering is automatic based on your retention settings, no archival job to build or run.
Does querying both recent and old data require two queries?: No. A single SQL query (with time-series functions) transparently spans the memory and magnetic tiers.
Your workload is relational transactions across customers and orders. Timestream?: No, it’s purpose-built for time-stamped telemetry. Relational/transactional work belongs in RDS/Aurora.

Sources

Open the SysPad canvas · Official Timestream documentation