JetStream speeds up known queries inside existing databases
In plain words
Databases handle many kinds of queries, but that flexibility can make known queries slower. JetStream is a system that adds generated, query-specific accelerators to an existing database, which still handles storage, writes, and other queries.
Some accelerators keep extra data that saves work when answering a query. Across 22 tested queries, the stateful version had much faster reads on average than DuckDB. That result describes the tested queries, not every database or query.
The research
Can known database queries run faster without moving data into a separate system or giving up writes and less predictable queries? JetStream addresses that tension by attaching generated query accelerators to an existing database, which remains responsible for storage and other operations.
JetStream generates query accelerators beside an existing database, with optional maintained summaries and proofs, to speed known queries while preserving the host’s other work.
Read-only TPC-H results compare JetStream with DuckDB across all 22 queries at scale factor 20. Stateless accelerators had a 2.1x geometric-mean read speedup at 32 threads; stateful accelerators had 315x, and verified ones had 93x.
These geometric means combine per-query speedup ratios, so they summarize the tested queries rather than describe every query. With one refresh every 15 seconds, the geometric-mean workload speedup across the 22 queries was 120.5x over DuckDB, counting reads and maintenance costs.
The results cover the tested TPC-H workload and six SEC-EDGAR queries, not every database workload. The verified-answer guarantee also depends on host input and commit-order contracts, as well as trusted SQL translation and host interfaces.
Analysis
Deep
- Original paper authors
- Akhilesh Balasingam, Amadou Ngom, Geoffrey X. Yu, Tim Kraska
- Submitted to arXiv
- License
- arXiv non-exclusive