Hacker News Reader

Comment by riku_iki | Hacker News Reader

it depends on your query obviously.

In general, I did very deep benchmarking of pg, clickhouse and duckdb, and I sure didn't make stupid mistakes like this: https://news.ycombinator.com/item?id=36990831

My dataset has 50B rows and 2tb of data, and I think columnar dbs are very overhiped and I chose pg because:

- pg performance is acceptable, maybe 2-5x times slower than clickhouse and duckdb on some queries if pg is configured correctly and run on compressed storage

- clickhouse and duckdb start falling apart very fast because they specialized on very narrow type of queries: https://github.com/ClickHouse/ClickHouse/issues/47520 https://github.com/ClickHouse/ClickHouse/issues/47521 https://github.com/duckdb/duckdb/discussions/6696

View on HN

"2-5x times slower" can mean the difference from 2 seconds to 4 to 10 seconds. Two seconds is still (barely) acceptable for interactive usage, ten seconds: not so much. You're also going to need less beefy servers, or fewer servers.

I also "just" use PostgreSQL for all of this by the way, but the limitations are pretty obvious. You're much more limited in what you can query with good performance, unless you start creating tons of queries or pre-computed data and such, which have their own trade-offs. Columnar DBs are "overhyped" in the sense that everything in programming seems to be, but they do exist for good reasons (the reason I don't use it are because they also come with their own set of downsides, as well as just plain laziness).

ClickHouse can do large GROUP BY queries, not limited by memory: https://clickhouse.com/docs/en/sql-reference/statements/sele...

as explained in https://github.com/ClickHouse/ClickHouse/issues/47521#issuec... it can't, that parameters only applies on pre aggregation phase but not aggregation.

Feature request is not implemented yet: https://github.com/ClickHouse/ClickHouse/issues/40588

ClickHouse uses "grace hash" GROUP BY with the number of buckets = 256.

It can do size about 256 times larger than a memory because only one bucket has to be in memory while merging. It works for distributed query processing as well and is enabled by default.

About the linked issue - it looks like it is related to some extra optimization on top of what already exists.

> only one bucket has to be in memory while merging.

its hard for me to judge about implementation details, but per that person reply memory is also multiplied by number of threads which do aggregation.

Do you have happen to have any documentation about your benchmarking? I'm also considering these options at the moment (currently using pg+timescaledb) and interested in what you found.

I don't have documentation.

I just created large tables, and tried to join, group by, sort them in pg, clickhouse, duckdb, looked what failed or being slow, and tried to resolve it.

I am happy to answer specific questions, but I didn't use timescaledb.

Reply on news.ycombinator.com