Michael Kleen

``` compare_run.py --v1 branch:master --v2 branch:aw/indexer/per-ref-value-lookup --spec specs/select.toml --env CRATE_HEAP_SIZE=4g Q: select * from uservisits where "sourceIP" = '25.193.131.52' C: 1 | Version | Mean ± Stdev | Min |...

``` compare_run.py --v1 branch:master --v2 branch:aw/indexer/per-ref-value-lookup --spec specs/queries.toml --env CRATE_HEAP_SIZE=4g Q: select * from uservisits limit 100 C: 1 | Version | Mean ± Stdev | Min | Median |...

export uservisits is around 20 % slower. ``` compare_run.py --v1 branch:master --v2 branch:aw/indexer/per-ref-value-lookup --spec specs/export.toml Q: COPY uservisits TO DIRECTORY '/tmp/' with (wait_for_completion=true) C: 1 | Version | Mean ±...

Note: Recovery source was manually disabled on the branch ``` compare_run_disk_usage.py --v1 branch:master --v2 branch:aw/per-ref-valueno-no-recovery --spec specs/queries.toml Version1: branch:master Version2: branch:aw/per-ref-valueno-no-recovery Description Version 1 Unit Version 2 Unit Diff --------------------...

This one is looking much better now: ``` compare_run.py --v1 branch:master --v2 branch:aw/indexer/per-ref-value-lookup --spec specs/select.toml V1: 5.10.0-dea6e9b5ff707dd58ab9cebf9219d4565dd42e50 V2: 5.10.0-dffebbb8893579e47f14fc84f74cc29e853285f6 Q: select * from uservisits where "sourceIP" = '25.193.131.52' C: 1...

Here are still regressions: ``` compare_run.py --v1 branch:master --v2 branch:aw/indexer/per-ref-value-lookup --spec specs/queries.toml # Results (server side duration in ms) V1: 5.10.0-dea6e9b5ff707dd58ab9cebf9219d4565dd42e50 V2: 5.10.0-dffebbb8893579e47f14fc84f74cc29e853285f6 Q: select * from uservisits limit 10...

export is still having a regression: ``` compare_run.py --v1 branch:mkleen/with-source --v2 branch:master --spec specs/export.toml ## Running Query: Statement: COPY uservisits TO DIRECTORY '/tmp/' with (wait_for_completion=true) Concurrency: 1 Iterations: 10 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████|...

We said we'd only allow overriding the cluster setting with a table/index setting. Anything else in regards to storage/node paths is out-of-scope.

retest this please

The underlying [t-digest library](https://github.com/tdunning/t-digest) only supports double precision for percentiles. Therefore we cannot provide percentiles for numeric directly. What is the exact usecase ? Would a cast to double sufficient...