[configuration] Add exception for 'false default' when there is a double negation

[mx-psi]

Relates to #4344.

Changes

Adds exception to naming and default value rule for environment variables.

This is a backwards compatible change since the original statement is phrased as a 'SHOULD'. The OTEL_SDK_DISABLED deviates from this new norm (this is not a spec violation but a lack of consistency).

The wording tries to be similar to the one in footnote [1] here: https://github.com/open-telemetry/opentelemetry-specification/blob/v1.40.0/specification/protocol/exporter.md

There are several OpenTelemetry artifacts that are either not bound by the specification or have decided, for usability reasons, to not follow this rule in this case. This change would, I argue, bring more consistency and less confusion to end users.

I plan to submit a follow-up PR to change MeterConfig.disabled to MeterConfig.enabled if this PR is accepted.
I don't plan to submit a PR to migrate OTEL_SDK_DISABLED, but this could be done if there is interest.

Thanks to @pellared for guidance on this :)

For non-trivial changes, follow the change proposal process.

• Related issues [metrics] MeterConfig disabled does not follow Collector enabled convention #4344
• ~~Related OTEP(s) ~~
• Links to the prototypes (when adding or changing features) (not applicable)
• CHANGELOG.md file updated for non-trivial changes
• spec-compliance-matrix.md updated if necessary

[trask]

I support but would recommend that it be left open through at least the Tue, Jan 21 spec meeting given the contentious discussion around the topic previously.

[MrAlias]

This is introducing ambiguity and inconsistency to a clear and well litigated part of the specification. It is only substantiated with a biased subjective opinion about an undefined user-base.

The prior decision for this part of the specification was made on logical reasoning. There are neither objective data nor facts being presented here that add to the prior discussion. Do these exist? Can you provide them to motivate this change?

The added ambiguity and inconsistency deteriorate the specification and should not be accepted without better motivation.

[mx-psi]

with a biased subjective opinion

Could you clarify what is biased about my or others opinion here?

about an undefined user-base.

I think I have been very clear about the user base that motivated this decision: Collector users. The discussion that followed also showed other artifacts within OpenTelemetry that do not follow this approach. What is undefined about this?

There are neither objective data nor facts being presented here that add to the prior discussion.

#4344 (comment), #4344 (comment) and #4344 (comment) are objective data showing how this convention is not followed elsewhere in OpenTelemetry.

I am happy to dig up more, but before I spend any further effort on this, I would like to step back and clarify what evidence would you consider as 'objective data'. What kind of evidence would make you change your mind?

The added ambiguity and inconsistency deteriorate the specification

The sentence this PR modifies in the spec is inconsistent with several software artifacts end users use that are under the OpenTelemetry umbrella. I disagree that this adds any inconsistency. Furthermore, I think the sentence has around the same level of ambiguity as the existing text. Could you give examples where this is more ambiguous than before?

and should not be accepted without better motivation.

Again, I think we need to step back and think: what motivation would we need? I don't want to spend effort on this if there is no motivation that would justify this change for you.

[MrAlias]

Could you clarify what is biased about my or others opinion here?

You represent a specific demographic with inherent bias.

[MrAlias]

about an undefined user-base.

I think I have been very clear about the user base that motivated this decision: Collector users. The discussion that followed also showed other artifacts within OpenTelemetry that do not follow this approach. What is undefined about this?

This is too narrow of a scope to claim it is representative of OpenTelemetry users. What about all the other languages being used, do you have data representing their expectations?

[MrAlias]

There are neither objective data nor facts being presented here that add to the prior discussion.

#4344 (comment), #4344 (comment) and #4344 (comment) are objective data showing how this convention is not followed elsewhere in OpenTelemetry.

These are all instances where different choices were made. These choice, again, are framed with biased subjective opinions. They do not address any reasoning from #2755 or add anything new to that discussion.

[MrAlias]

I am happy to dig up more, but before I spend any further effort on this, I would like to step back and clarify what evidence would you consider as 'objective data'. What kind of evidence would make you change your mind?

I would appreciate that. I'm looking for user feedback, not OTel developer opinions, that represent this view. I would like the data to show not only those that are motivated to complain in an issue about the current state of things, but also shows how many users appreciate the current form of our environment variables.

[MrAlias]

I am happy to dig up more, but before I spend any further effort on this, I would like to step back and clarify what evidence would you consider as 'objective data'. What kind of evidence would make you change your mind?

I would appreciate that. I'm looking for user feedback, not OTel developer opinions, that represent this view. I would like the data to show not only those that are motivated to complain in an issue about the current state of things, but also shows how many users appreciate the current form of our environment variables.

An OTel user-survey would help in this regard.

[MrAlias]

The added ambiguity and inconsistency deteriorate the specification

The sentence this PR modifies in the spec is inconsistent with several software artifacts end users use that are under the OpenTelemetry umbrella. I disagree that this adds any inconsistency. Furthermore, I think the sentence has around the same level of ambiguity as the existing text. Could you give examples where this is more ambiguous than before?

Adding an exception that directly contradicts and existing recommendation is ambiguous. It leaves the reader with an unclear understanding about what they need to do. Especially since we already have prior art where the non-normative "exception" has been ignored.

[MrAlias]

Again, I think we need to step back and think: what motivation would we need? I don't want to spend effort on this if there is no motivation that would justify this change for you.

I appreciate that fact. However, I would ask that it also be understood that much effort has already gone into the decision that resulted in our current state. This will reverse that decision. It seems reasonable that we require some level of effort to provide objective facts and data to motivate the change. Without such we will not be making an effective decision and likely just find ourselves back here in a few months.

[yurishkuro]

I tend to agree with @MrAlias here - the existing wording is very clear and simple, the added caveat makes it worse. The "double negation" argument tries to fix poor data modeling / naming by introducing a caveat to the spec that breaks the fundamental principle - unset values are defaulted to "zero" value, which for Booleans is false. I would rather fix the naming. I agree that x_disabled=false is harder to grok, but the solution is to change the name, not the default treatment, e.g. if x is some action (like "sampling" or "reload") then replace it with skip_x, which has no double negative readability issues.

[austinlparker]

I think @yurishkuro brings up a good point -- the existing guidance is clear and definitive, and loosening this requirement adds a lot of potential confusion. Is there some motivating examples for why we couldn't change names in order to allow for false defaults?

[trask]

Is there some motivating examples for why we couldn't change names in order to allow for false defaults?

from the Java agent perspective (#4344 (comment)):

just sharing a data point, I suspect we won't change OTEL_INSTRUMENTATION_[NAME]_ENABLED in the Java agent, since some of the instrumentations are enabled by default and some are disabled by default, and so it would feel weird (and inconsistent in a different way) to have some specific instrumentations using OTEL_INSTRUMENTATION_[NAME]_ENABLED and other specific instrumentations using OTEL_INSTRUMENTATION_[NAME]_DISABLED

(and an additional wrinkle is that we have occasionally changed the default for specific instrumentations from enabled to disabled in major version bumps)

[austinlparker]

Is there some motivating examples for why we couldn't change names in order to allow for false defaults?

from the Java agent perspective (#4344 (comment)):

just sharing a data point, I suspect we won't change OTEL_INSTRUMENTATION_[NAME]_ENABLED in the Java agent, since some of the instrumentations are enabled by default and some are disabled by default, and so it would feel weird (and inconsistent in a different way) to have some specific instrumentations using OTEL_INSTRUMENTATION_[NAME]_ENABLED and other specific instrumentations using OTEL_INSTRUMENTATION_[NAME]_DISABLED
(and an additional wrinkle is that we have occasionally changed the default for specific instrumentations from enabled to disabled in major version bumps)

Yeah, that is kinda annoying... is this something we can ensure that's normalized in the new config file stuff and just preserve the existing implementations?

[trask]

Yeah, that is kinda annoying... is this something we can ensure that's normalized in the new config file stuff and just preserve the existing implementations?

I think that's the goal of this PR, but maybe it's not expressed well (or in the right place in the spec?), would appreciate any concrete suggestions on how to achieve this goal in specification language @MrAlias @yurishkuro

[yurishkuro]

if we're talking specifically about the situation where multiple instrumentations can be enabled by user while at the same time some are automatically enabled, a good approach that is often used in linters is to have a named "preset" (rather than implicit defaults) that controls which parts are enabled by default via that preset. In other words, the enabled defaults (which contradict the existing guidance for default being false) are not enabled because individual settings are defaulted to true, but because another umbrella mechanism is applied, and then individual settings can override it.

[trask]

if we're talking specifically about the situation where multiple instrumentations can be enabled by user while at the same time some are automatically enabled, a good approach that is often used in linters is to have a named "preset" (rather than implicit defaults) that controls which parts are enabled by default via that preset. In other words, the enabled defaults (which contradict the existing guidance for default being false) are not enabled because individual settings are defaulted to true, but because another umbrella mechanism is applied, and then individual settings can override it.

yeah, this sounds like a nice approach when configuration properties are limited to key/value pairs, but I think the Declarative Configuration SIG has ruled out any kind of built-in merge/overlay support for declarative (yaml) configuration

[yurishkuro]

the Declarative Configuration SIG has ruled out any kind of built-in merge/overlay support for declarative (yaml) configuration

the presets approach does not contradict declarative configuration (e.g. golangci yaml config uses it)

[pellared]

As of today .NET Automatic Instrumentation has a lot of _ENABLED env vars defauling to true: https://opentelemetry.io/docs/zero-code/net/configuration. We failed to find a better name that would not be confusing. Our users and ourselves did not like the _DISABLED with default false proposal (it was discussed several times during SIG meetings).

Related issue: open-telemetry/opentelemetry-dotnet-instrumentation#1633

CC @open-telemetry/dotnet-instrumentation-maintainers (please double-check)