Hi Marcell,
It'd be helpful to see the table DDL and the query too along with an idea of how many regions might be involved in the query. If a query is a commonly run query, usually you'll design the row key around optimizing it. If you have other, simpler queries that have determined your row key, then another alternative is to add one or more secondary indexes. Another common technique is to denormalize your data in ways that precompute the join to avoid having to do it at run time.

With joins, make sure to order your tables from post filtered largest (on LHS) to smallest (on RHS). Also, if you're joining on the PK of both tables, you should use the USE_SORT_MERGE_JOIN hint. Another common tuning exercise is around determining the best parallelization to use (i.e. guidepost width) or even disabling parallelization for more than an entire region's worth of data.

It'd also be interesting to see the raw HBase code for a query of this complexity.

Thanks,
James 

On Thu, Mar 8, 2018 at 1:03 PM, Marcell Ortutay <mortutay@23andme.com> wrote:
Hi,

I am using Phoenix at my company for a large query that is meant to be run in real time as part of our application. The query involves several aggregations, anti-joins, and an inner query. Here is the (anonymized) query plan: https://gist.github.com/ortutay23andme/1da620472cc469ed2d8a6fdd0cc7eb01

The query performance on this is not great, it takes about 5sec to execute the query, and moreover it performs badly under load. If we run ~4qps of this query Phoenix starts to timeout and slow down a lot (queries take >30sec).

For comparison, I wrote a simple Go script that runs a similar query talking directly to HBase. The performance on it is substantially better. It executes in ~1.5sec, and can handle loads of ~50-100qps on the same cluster.

I'm wondering if anyone has ideas on what might be causing this difference in performance? Are there configs / optimizations we can do in Phoenix to bring the performance closer to direct HBase queries?

I can provide context on the table sizes etc. if needed.

Thanks,
Marcell