Metrics summary quantiles are incorrect

[dwlbnks]

Summary

query_log_execution_time latency quantiles are not computed correctly

Description

Samples are displayed in the wrong latency buckets when only a few samples are present. When a query is run twice for example, the faster execution time should be displayed in the 0 and 0.5 quantile, while the slower should be displayed in the upper quantiles. instead the faster is displayed in all quantiles up to 0.999 and the slower is displayed in the 1 quantile.

This is likely an issue with metrics crate that ReadySet depends on.

Expected behavior

Latencies appear in correct quantiles.

Actual behavior

Latencies are not binned correctly

Steps to reproduce

Run a query twice with metrics enabled and query the metrics endpoint (within the 1 min sliding window)

ReadySet version

main (1b7efc9ff93e1579a0ebc585d3c3c884fc19b7dc)

Upstream DB type and version

N/A

Instance Details

N/A

Deployment Details

N/A

Logs

N/A

[dwlbnks]

Bug introduced in mheffner/rust-sketches-ddsketch#5. Unit tests in https://github.com/metrics-rs/metrics/tree/main/metrics-exporter-prometheus will probably have to be updated.

[tbjuhasz]

There was an outstanding question from Mike, the person who maintains the Crate. I have answered his question and waiting for him to merge the change.

mheffner/rust-sketches-ddsketch#10

[gvsg-rs]

@tbjuhasz - Can you update this ticket and close this as "Won't do".