Deep dive into the numbers behind the performance that makes Firebolt the next-gen cloud data warehouse for low-latency analytics.
A realistic benchmark that mirrors the real-world demands of production workloads, delivering results that truly matter.
A set of queries that represent customer usage patterns and reflect complex real-world data operations
Workload execution scenarios that simulate diverse operational conditions, such as varying concurrency levels and data sizes
Broad range of structural complexity, based on weighted query elements such as joins, filters, aggregation etc.
Query latency: 20% of queries (5 out of 25) have less than 25 ms latency, on single node engine across S/M/L/XL node types
Scale-up and scale-out: over 8x faster execution with single extra-large and 8-node setups
Multi-cluster: query throughput soared from 449 to 4213 QPS; low latency maintained in tens of milliseconds under 8000 concurrent requests
Bulk ingestion: the benchmark demonstrates near-linear scaling, with ingestion speeds above 15 TB/hr, as more nodes are added
Parquet (Snappy compressed): top performer with the fastest ingestion times (235 seconds), the best price-to-performance ratio ($2.46/TB), and the greatest throughput (15.3 TB/hour)
CSV (gzip) and CSV (uncompressed): show near-linear scaling, with an avg. cost of $4.54/TB across all 3 engine configurations (4, 8 & 16 small nodes)
INSERT: Singleton and micro-batch INSERT queries have P95 response times of 0.273 and 0.284 seconds respectively, only rising to 0.441 seconds when inserting 1,000 rows at a time
UPDATE: Singleton UPDATE queries have median and P95 latencies of 0.215 and 0.263 seconds respectively
DELETE: Singleton DELETE queries have median and P95 latencies of 0.194 and 0.313 seconds respectively
%20logo.png){kind=icon}
