T0

An experimental high performance real-time in-memory distributed timeseries database.

In current, it is just a demo.

Targets

• kindly inspire by InfluxDB IOx, Google Monarch, ScyllaDB and much more open-source projects.
• use new experimental features to make the fastest insertion / query TSDB
• OLAP / OLTP fusion
  • support massive get / set operation for recent mutable data.
  • immutable old data & use Apache Arrow / Parquet ecosystem
    • boost analytical query (push-down more piplinable calculator with SIMD)
    • support zero-copy transportation to easily integrate to other analytical project
  • massive distributed query
• load-on-demand component & easily scaling on component level
• optional WAL / distribution backup
• columnar format & rich-type column

Core Concepts

• Data must has a timestamp
• Tracking data transmutation of unique object in various of times (timeseries)
  • Has a set of labels to identify unique object in different timestamps
  • Has a set of data(scalar) that are computable
• Data distribution is related to the time
  • Can always find a minimal size of time interval (unit), in this level, data in each fragment of a single unique object are continuous
  • Can always find a large enough size of timer interval, in this level, both unique object and data are sparse
• Insertions are always happened in recent rather than before
• OLTP / OLAP fusion
  • Recent data for single data query: alerting, monitoring
  • History data for analysis: attribution analysis, machine learning
• Data aggregation are always group by timestamp
• Data can be merged with neighboring data on time
• No transaction required

Installation

Dependencies

• rustc (1.60.0-nightly+)
• clang (13.0.0+)

Build

git clone https://github.com/Homebrew-TSDB-Club/t0.git
cargo build --release

Get Started

./target/release/t0 --address=0.0.0.0:1108 --server-cores=24 --storage-cores=16

Example

Prometheus Remote Write

Features

• core
  • coroutine runtime
    • CPU core-affinity coroutine runtime
    • epoll I/O
    • Linux aio / io_uring API
    • query language
      • uniform logical expression
      • PromQL parser
      • custom query language syntax & parser
    • asynchronous & multiplexing server
      • Tokio(work-stealing coroutine) based HTTP/2(gRPC) server
      • core-affinity coroutine based HTTP/2(gRPC) server
      • FlatBuffers over QUIC
  • function level tracing
  • load-on-demand component: insertion / storage / query / config
  • decentralized federation deployment
  • self metrics
• insertion
  • Prometheus remote write protocol
  • custom protocol over FlatBuffers
• storage
  • column-oriented
  • rich-type column
  • chunking data
  • data format
    • mutable in-memory chunk with custom format
    • immutable in-memory Apache Arrow format
    • immutable Apache Parquet format file storage
  • shared-nothing insertion based on CPU core-affinity coroutine
  • query calculator push-down
    • projection
    • filter
    • time range
    • limit
    • pipeline compute
  • data archive pipeline: mutable -> immutable -> file
• query
  • basic PromQL support
  • transport
    • Apache Arrow Flight over HTTP/2(gRPC)
    • DPDK / RDMA
  • shared-nothing mutable chunk query
  • inverted index
  • sparse index

Benchmark

Insertion

Conclusion

Over 30% overhead in croaring bitmap, and still have lots of optimize ways.