A simple benchmark on real-world NYC Taxi dataset has been exercised:
| Query | ClickHouse (v21.2.5.5) | TensorBase (main branch) | Speedup Ratio of TB |
|---|---|---|---|
| select sum(trip_id) from trips_lite | 0.248 sec | 0.079 sec | 3.1 (TB is faster) |
| select date_part('year',pickup_datetime), count(1) from trips_lite group by date_part('year',pickup_datetime)* | 0.514 sec | 3.375 sec | 0.15 (TB is slower) |
Note:
- The adoption to ClickHouse has not been completed in that DataFusion supports a different dialect. The query sql in ClickHouse is: select toYear(pickup_datetime), count(1) from trips_lite group by toYear(pickup_datetime)
- Hardware: 1 Socket, Intel Xeon Platinum 8260, 24 cores / 48 hyperthreads, 6-channel DDR4-2400 ECC REG 192GB DRAMs
- trips_lite Dataset: column-stripped NYC TAXI Dataset, 1464781690 rows, two columns(trip_id, pickup_datetime)
- Measurement rules: run 3 times, pick up the best run time. This is an in-memory test. If runs from the cold disk, ClickHouse wins for compression.