[swimorsinka]

Hmm, I use Redshift every day and I've also used BigQuery. You can't have a valid benchmark without adding sort keys and dist keys to Redshift. It really isn't meant to work without them.

I understand it's more work that way, which is why BigQuery is so nice. But if you're trying to do lots of queries, BigQuery is also more expensive.

[georgewfraser]

We've gotten a lot of passionate feedback in this category, and we actually did a comparison with sort and dist keys in the extended version. They don't make as much difference as people think, these charts would be ~15% shifted with sort and dist keys.

[vgt]

Did you guys implement cluster keys in BigQuery as well?

(BQ PM)

[georgewfraser]

No we just did a comparison between Redshift with no tuning, conservative tuning, aggressive tuning. The point was just to assess "how much does tuning matter anyway?"

You are correct that if we wanted to do a full-on "tuned comparison", we would need to implement partitioned tables in BQ, Snowflake and Presto.

[vgt]

(and clustering in BQ! [0])

[0] https://cloud.google.com/bigquery/docs/clustered-tables

[maldeh]

That actually sounds like a solid improvement. Do the head-to-head comparisons change with sort- and dist-keys? What was the tuning used?

[georgewfraser]

You can see the whole back-and-forth here:

https://github.com/fivetran/benchmark/issues/1

The tuning was pretty aggressive, some of the sort keys were a little unrealistic for a real-world scenario where you're doing a lot of different queries.