1.1 Time-Series Database

This is a historical learning note and may contain outdated or incomplete understanding.

1. What Is a Time-Series Database

A database optimized specifically for timestamp or time-series data. Compared with many traditional business tables that only keep the current state, a time-series database keeps historical data that changes over time. Time-series queries also often use time as a filter condition.

2. What Is Time-Series Data

Time-series data is a series of data based on time.

3. Why a Time-Series Database Is Needed

3.1. Recording with a Relational Model

If a database is used to record time-series data, it looks like this:
- Disadvantages:
  - Writes: large amount of data,
  - Queries: low efficiency when aggregating by time
  - Cost: high cost

3.2. Recording with a Time-Series Model

metric: the metric name and identifier of the current data. It is equivalent to a table in a relational database.
data point: a data point, equivalent to a row in a relational database.
timestamp: the timestamp representing when the data point was generated.
field: different fields under a metric. In general, these store attribute information that changes as the timestamp changes. For example, a location metric can have longitude and latitude as two fields.
tag: a label or additional information. In general, these store attribute information that does not change as the timestamp changes. A timestamp plus all tags can be regarded as the table’s primary key.
In the example above:
- The metric is Wind, and every data point has a timestamp.
- There are two fields: direction and speed.
- There are two tags: sensor and city. The first and third rows both store the device whose sensor number is 95D8-7913, and whose city attribute is Shanghai. As time changes, both wind direction and wind speed change: the direction changes from 23.4 to 23.2, while the speed changes from 3.4 to 3.3.

4. Time-Series Database Implementation

4.1. LSM Tree

LSM.md

4.2. Distributed Storage

Time-series databases face massive data writing, storage, and reading workloads that a single machine cannot handle, so multiple machines are needed for storage, that is, distributed storage.

4.2.1. Distribution Algorithm

Distributed system partitioning

4.2.2. Shard Key

Shard by metric + tags. Because queries are often performed over a time range, data with the same metric and tags can be assigned to one machine and stored continuously. Sequential disk reads are fast.