TimeSeries data

• Components

  • Object

  • Metric

  • Tag

• Access patterns

• Storage in HBase

  • Rowkey

    • Why rowkey is important

    • Row key design

      • Benefits

      • Support tag aggregation

    • Scaling

      • Vertical sharding

      • Horizontal partitioning

      • Downsampling

        • Pre-downsampling

        • Post-downsampling

• References

  • 百度

  • Write time series DB from scratch

  • ELK

  • Uber M3

  • Datadog

  • Aggregation

Components

• Time series = Object + Tag + Metrics + actual data

Object

• Monitoring object could be in three categories:

  • Machine level: Physical machine, virtual machine, operation system

  • Instance level: Container, process

  • Service level (logical object): Service, service group, cluster

Metric

• Metrics are numeric measurements. Metrics can include:

  • A numeric status at a moment in time (like CPU % used)

  • Aggregated measurements (like a count of events over a one-minute time, or a rate of events-per-minute)

• The types of metric aggregation are diverse (for example, average, total, minimum, maximum, sum-of-squares), but all metrics generally share the following traits:

  • A name

  • A timestamp

  • One or more numeric values

Tag

• Annotated key value pairs

Access patterns

• Sequential read: Read by time range

• Random write: Different time series data

  • Usually each object has a write sampling frequency is per 5s/10s.

• Much more write than read

• Lots of aggregating dimensions

Storage in HBase

Rowkey

Why rowkey is important

• If rowkey could be designed properly, then data could be distributed evenly into HRegions. And different HRegions could be located in different server nodes.

Row key design

ts = (object, tags) + metric + [(timestamp, value), (timestamp, value), …]

// entity_id is hashed result of combination (object, tags)
// metric_id is hashed result of metric
// timebase is the result of Unix timestamp % 3600, 
//             4 byte length, Rowkey represents 1 hour data. 
RowKey = entity_id + metric_id + timebase

Benefits

• entity_id and metric_id makes data evenly distributed.

• timebase makes continous data next to each other.

Support tag aggregation

• HBase does not have native support for index. This makes it impossible to find all entity_ids given a tag.

• In the example below:

  • Give a tag: K1=V1, it could find all entities containing the tag: entity_id1, entity_id2, entity_id3

Scaling

Vertical sharding

• Each product has a different database

Horizontal partitioning

• Slice name is Product - data - {starttime}

  • startime is the data starting time in table.

  • It will help remove data in batch.

Downsampling

Pre-downsampling

• The longer the retention period is, the less data could be stored.

Post-downsampling

• At query time, dynamically downsample data based on user assigned query range.

References

百度

• 百度大规模时序数据存储（一）| 监控场景的时序数据
• 百度大规模时序数据存储（二）| 存储选型及数据模型设计
• 百度大规模时序数据存储（三）| 核心功能设计

Write time series DB from scratch

• https://fabxc.org/tsdb/

ELK

• https://www.twosigma.com/articles/building-a-high-throughput-metrics-system-using-open-source-software/

Uber M3

• https://eng.uber.com/m3/

Datadog

• https://www.infoq.com/presentations/datadog-metrics-db/

Aggregation

• https://www.youtube.com/watch?v=UEJ6xq4frEw&ab_channel=HasgeekTV