LocustDB

An experimental analytics database aiming to set a new standard for query performance and storage efficiency on commodity hardware. See How to Analyze Billions of Records per Second on a Single Desktop PC and How to Read 100s of Millions of Records per Second from a Single Disk for an overview of current capabilities.

Usage

Download the latest binary release, which can be run from the command line on most x64 Linux systems, including Windows Subsystem for Linux. For example, to load the file test_data/nyc-taxi.csv.gz in this repository and start the repl run:

Bash ./locustdb --load test_data/nyc-taxi.csv.gz --trips

When loading .csv or .csv.gz files with --load, the first line of each file is assumed to be a header containing the names for all columns. The type of each column will be derived automatically, but this might break for columns that contain a mixture of numbers/strings/empty entries.

To persist data to disk in LocustDB's internal storage format (which allows fast queries from disk after the initial load), specify the storage location with --db-path When creating/opening a persistent database, LocustDB will open a lot of files and might crash if the limit on the number of open files is too low. On Linux, you can check the current limit with ulimit -n and set a new limit with e.g. ulimit -n 4096.

The --trips flag will configure the ingestion schema for loading the 1.46 billion taxi ride dataset which can be downloaded here.

For additional usage info, invoke with --help:

```Bash $ ./locustdb --help LocustDB 0.1.0-alpha Clemens Winter clemenswinter1@gmail.com Massively parallel, high performance analytics database that will rapidly devour all of your data.

USAGE: locustdb [FLAGS] [OPTIONS]

FLAGS: -h, --help Prints help information --mem-lz4 Keep data cached in memory lz4 encoded. Decreases memory usage and query speeds. --reduced-trips Set ingestion schema for select set of columns from nyc taxi ride dataset --seq-disk-read Improves performance on HDD, can hurt performance on SSD. --trips Set ingestion schema for nyc taxi ride dataset -V, --version Prints version information

OPTIONS: --db-path Path to data directory --load ... Load .csv or .csv.gz files into the database --mem-limit-tables Limit for in-memory size of tables in GiB [default: 8] --partition-size Number of rows per partition when loading new data [default: 65536] --readahead How much data to load at a time when reading from disk during queries in MiB [default: 256] --table Name for the table populated with --load [default: default] --threads Number of worker threads. [default: number of cores (12)] ```

Goals

A vision for LocustDB.

Fast

Query performance for analytics workloads is best-in-class on commodity hardware, both for data cached in memory and for data read from disk.

Cost-efficient

LocustDB automatically achieves spectacular compression ratios, has minimal indexing overhead, and requires less machines to store the same amount of data than any other system. The trade-off between performance and storage efficiency is configurable.

Low latency

New data is available for queries within seconds.

Scalable

LocustDB scales seamlessly from a single machine to large clusters.

Flexible and easy to use

LocustDB should be usable with minimal configuration or schema-setup as: - a highly available distributed analytics system continuously ingesting data and executing queries - a commandline tool/repl for loading and analysing data from CSV files - an embedded database/query engine included in other Rust programs via cargo

Non-goals

Until LocustDB is production ready these are distractions at best, if not wholly incompatible with the main goals.

Strong consistency and durability guarantees

small amounts of data may be lost during ingestion
when a node is unavailable, queries may return incomplete results
results returned by queries may not represent a consistent snapshot

High QPS

LocustDB does not efficiently execute queries inserting or operating on small amounts of data.

Full SQL support

All data is append only and can only be deleted/expired in bulk.
LocustDB does not support queries that cannot be evaluated independently by each node (large joins, complex subqueries, precise set sizes, precise top n).

Support for cost-inefficient or specialised hardware

LocustDB does not run on GPUs.

Compiling from source

Install Rust: rustup.rs
Clone the repository

Bash git clone https://github.com/cswinter/LocustDB.git cd LocustDB

Compile with RUSTFLAGS="-Ccodegen-units=1" and CARGO_INCREMENTAL=0 for optimal performance:

Bash RUSTFLAGS="-Ccodegen-units=1" CARGO_INCREMENTAL=0 cargo run --release --bin repl -- --load test_data/nyc-taxi.csv.gz --reduced-trips

Running tests or benchmarks

cargo test

RUSTFLAGS="-Ccodegen-units=1" CARGO_INCREMENTAL=0 cargo bench

Storage backend

LocustDB has support for persisting data to disk and running queries on data stored on disk. This feature is disabled by default, and has to be enabled explicitly by passing --features "enable_rocksdb" to cargo during compilation. The database backend uses RocksDB, which is a somewhat complex C++ dependency that has to be compiled from source and requires gcc and various libraries to be available. You will have to manually install those on your system, instructions can be found here. You may also have to install various other random tools until compilation succeeds.

LZ4

Compile with --features "enable_lz4" to enable an additional lz4 compression pass which can significantly reduce data size both on disk and in-memory, at the cost of slightly slower in-memory queries.