Exponential histogram downscaling can go into an infinite loop

[axw]

Describe the bug

When configuring a base-2 exponential bucket histogram with a maximum of 1 bucket, DoubleHistogram.record may spin and never return.

Steps to reproduce

Run this example: axw/opentelemetry-java-examples@a1bf2e7

What did you expect to see?

I expect the the program not to spin.

I don't know what the correct logic should be. Require > 1 max buckets? Or otherwise automatically increase the zero threshold to 1.0 as needed? This probably needs clarification in https://opentelemetry.io/docs/specs/otel/metrics/sdk/#base2-exponential-bucket-histogram-aggregation

What did you see instead?

The program spins. Attaching a debugger shows that it's stuck in an infinite loop in DoubleHistogram.record, due to the start index being -1 and the end index shifting down to 0 -- so the delta converges on 2, which is always greater than a maximum of 1 bucket.

What version and what artifacts are you using?
Artifacts: opentelemetry-api, opentelemetry-sdk
Version: 1.33.0

Environment
Compiler: OpenJDK 17
OS: Ubuntu 22.04

[jack-berg]

Thanks for the report and the repro. Will dig into this.

[jack-berg]

I don't know what the correct logic should be. Require > 1 max buckets?

I think we need to require > 1 max buckets. Specifying a single bucket defeats the purpose of exponential bucket histogram since there is effectively no ability to rescale to accommodate the range of measurements. With a single bucket, the measurements MUST fit into that bucket, and because the formula for bucket boundaries always results in a boundary falling on 1, there is no scale which can accommodate a measurement <= 1 and any measurement >= 1 in a single bucket.

To demonstrate, try running your example with:

histogram.record(1.0);
histogram.record(1.00001);

This will hang, while the following will resolve:

histogram.record(1.00001);
histogram.record(Double.MAX_VALUE);

To your point, we could also increase the zero threshold to 1.0.

But what is the point of a single bucket exponential histogram? What information do you gain over an explicit bucket histogram with a single bucket (which records min, max, sum, count)?

As an additional data point, dot net requires the number of buckets to be >= 2: https://github.com/open-telemetry/opentelemetry-dotnet/blob/main/src/OpenTelemetry/Metrics/Base2ExponentialBucketHistogramConfiguration.cs#L29-L34

[axw]

But what is the point of a single bucket exponential histogram? What information do you gain over an explicit bucket histogram with a single bucket (which records min, max, sum, count)?

Agreed, it's fairly pointless. I was just messing around with the parameters and happened to notice this behaviour. I think it's fair to require more than one bucket - that should be simplest.