Mar 5, 2021
Druid vs Rockset Revisited
Alan Li
Rockset recently published a blog post that compared the performance of Apache Druid 0.18 versus Rockset using the SSB benchmark. Druid 0.18 is about 9 months out of date at this point, so we wanted to revisit the benchmark based on the latest version of Druid (0.20.1), which includes several performance improvements we’ve been doing over the last few months.
Our findings are below:
| Query | Imply3.3.0 | Imply 2021.02 | Rockset |
|---|---|---|---|
| Q1.1 | 1245 | 462 | 944 |
| Q1.2 | 155 | 154 | 254 |
| Q1.3 | 92 | 160 | 296 |
| Q2.1 | 662 | 139 | 161 |
| Q2.2 | 591 | 102 | 136 |
| Q2.3 | 319 | 79 | 129 |
| Q3.1 | 292 | 333 | 626 |
| Q3.2 | 514 | 102 | 598 |
| Q3.3 | 417 | 91 | 343 |
| Q3.4 | 100 | 72 | 32 |
| Q4.1 | 883 | 179 | 384 |
| Q4.2 | 389 | 109 | 132 |
| Q4.3 | 384 | 87 | 41 |
| 6043 | 2069 | 4076 |
We can see a remarkable improvement in performance with the latest version of Druid. Two main things drove most of the change:
1) Many of these queries in the SSB benchmark use expressions. In older versions of Druid, we did not have vectorized expressions implemented, so many of the queries couldn’t vectorize. In the latest version, 100% of these queries vectorize.
2) The schema used to ingest data was changed to match what Rockset is doing. Similar to how Rockset “specified some keys for column-based clustering”, we also used column based clustering (in Druid it’s called partitioning). No further tuning was done in Druid.
In our next release of Druid, we are releasing another set of performance improvements. Stay tuned for more information.