June 29, 2026
Benchmarks, bragging, and side-eye
How we made WINDOW JOIN parallel and vectorized
QuestDB says it made a painful data task way faster—and the crowd wants receipts
TLDR: QuestDB says it turned a messy, slow market-data lookup into a simpler command that runs much faster on giant datasets. Commenters liked the comparisons, but the big mood was: cool story, now show the results in a trusted public benchmark so everyone can argue properly.
QuestDB just rolled in with a classic 'look what we built' flex: a new way to answer a very trading-desk kind of question—matching each trade with the average buy and sell prices around it—without forcing users to write a giant spaghetti bowl of database code. In plain English, they took something that used to be annoyingly complicated and turned it into a shorter command, then said it runs much faster on huge piles of market data. The company even brought benchmark bragging rights, claiming big wins over ClickHouse, DuckDB, and Timescale.
But the real action? The comments. One reader, pantsforbirds, gave the post a polite gold star, basically saying the comparison was actually useful because it showed how a purpose-built database behaves differently from general favorites like DuckDB and ClickHouse. That’s the calm side of the room. Then tosh showed up with the sneaky little comment that always starts a bigger argument: where’s the time-series version of ClickBench, the famous public speed test people love to fight over? Translation: nice claims, but can we get a neutral showdown before we crown anybody king?
So the mood is half applause, half 'drop the benchmark league table'. No full-on flame war yet, but the energy is absolutely there: one camp is impressed by the cleaner approach, while the other is already reaching for independent tests, spreadsheets, and probably popcorn. Tech post? Sure. But in the comments, it’s really a story about trust, bragging rights, and the eternal internet demand for receipts.
Key Points
- •QuestDB introduces WINDOW JOIN as a dedicated SQL operator for aggregating rows from one table over a time window around rows in another table.
- •The article contrasts WINDOW JOIN with a more complex workaround using ASOF JOIN, range join, UNION ALL, and GROUP BY to compute time-windowed averages.
- •QuestDB says the old pattern is inefficient because it scans the prices table independently for two join phases and requires hash aggregation over about 50 million groups in the test data.
- •The implementation combines left-hand-side data parallelism with a low-cardinality fast path that enables reuse of SIMD aggregation kernels.
- •On a benchmark with 50 million trades rows and 150 million prices rows, QuestDB reports its parallel plus SIMD path is 5.0x faster than its single-threaded fallback and 25x faster than ClickHouse's best rewrite.