I initially wanted to make a sub-millisecond log parser in C++ but that blew into a embeddable decision engine, that can run YAML defined rules on incoming data. The rules are executed in a vectorized format on incoming data by reprojecting into a columnar format first, if it's not already. Depending on the payload size and rules complexity, the performance goes from 200K records/s to more than million records/sec, in terms of througput this would be around 200 MiB/s to 3 GiB/s on average.
Rules can be sql expressions too, or onnx models (numeric), window ops and quite a few more operations are supported.
It's comparable to DuckDB but for streaming data and on the fly decisions
0 comments