Define observability requirements for stable components (#11772)

## Description

This PR defines observability requirements for components at the
"Stable" stability levels. The goal is to ensure that Collector
pipelines are properly observable, to help in debugging configuration
issues.

#### Approach

- The requirements are deliberately not too specific, in order to be
adaptable to each specific component, and so as to not over-burden
component authors.
- After discussing it with @mx-psi, this list of requirements explicitly
includes things that may end up being emitted automatically as part of
the Pipeline Instrumentation RFC (#11406), with only a note at the
beginning explaining that not everything may need to be implemented
manually.

Feel free to share if you don't think this is the right approach for
these requirements.

#### Link to tracking issue
Resolves #11581

## Important note regarding the Pipeline Instrumentation RFC

I included this paragraph in the part about error count metrics:
> The goal is to be able to easily pinpoint the source of data loss in
the Collector pipeline, so this should either:
>   - only include errors internal to the component, or;
> - allow distinguishing said errors from ones originating in an
external service, or propagated from downstream Collector components.

The [Pipeline Instrumentation
RFC](https://github.com/open-telemetry/opentelemetry-collector/blob/main/docs/rfcs/component-universal-telemetry.md)
(hereafter abbreviated "PI"), once implemented, should allow monitoring
component errors via the `outcome` attribute, which is either `success`
or `failure`, depending on whether the `Consumer` API call returned an
error.

Note that this does not work for receivers, or allow differentiating
between different types of errors; for that reason, I believe additional
component-specific error metrics will often still be required, but it
would be nice to cover as many cases as possible automatically.

However, at the moment, errors are (usually) propagated upstream through
the chain of `Consume` calls, so in case of error the `failure` state
will end up applied to all components upstream of the actual source of
the error. This means the PI metrics do not fit the first bullet point.

Moreover, I would argue that even post-processing the PI metrics does
not reliably allow distinguishing the ultimate source of errors (the
second bullet point). One simple idea is to compute
`consumed.items{outcome:failure} - produced.items{outcome:failure}` to
get the number of errors originating in a component. But this only works
if output items map one-to-one to input items: if a processor or
connector outputs fewer items than it consumes (because it aggregates
them, or translates to a different signal type), this formula will
return false positives. If these false positives are mixed with real
errors from the component and/or from downstream, the situation becomes
impossible to analyze by just looking at the metrics.

For these reasons, I believe we should do one of four things:
1. Change the way we use the `Consumer` API to no longer propagate
errors, making the PI metric outcomes more precise.
We could catch errors in whatever wrapper we already use to emit the PI
metrics, log them for posterity, and simply not propagate them.
Note that some components already more or less do this, such as the
`batchprocessor`, but this option may in principle break components
which rely on downstream errors (for retry purposes for example).
3. Keep propagating errors, but modify or extend the RFC to require
distinguishing between internal and propagated errors (maybe add a third
`outcome` value, or add another attribute).
This could be implemented by somehow propagating additional state from
one `Consume` call to another, allowing us to establish the first
appearance of a given error value in the pipeline.
5. Loosen this requirement so that the PI metrics suffice in their
current state.
6. Leave everything as-is and make component authors implement their own
somewhat redundant error count metrics.

---------

Co-authored-by: Pablo Baeyens <[email protected]>
Co-authored-by: Pablo Baeyens <[email protected]>

Author: jade-guiton-dd

docs/component-stability.md

@@ -66,6 +66,115 @@ Stable components MUST be compatible between minor versions unless critical secu
 component owner MUST provide a migration path and a reasonable time frame for users to upgrade. The same rules from beta
 components apply to stable when it comes to configuration changes.
 
+#### Observability requirements
+
+Stable components should emit enough internal telemetry to let users detect errors, as well as data
+loss and performance issues inside the component, and to help diagnose them if possible.
+
+For extension components, this means some way to monitor errors (for example through logs or span
+events), and some way to monitor performance (for example through spans or histograms). Because
+extensions can be so diverse, the details will be up to the component authors, and no further
+constraints are set out in this document.
+
+For pipeline components however, this section details the kinds of values that should be observable
+via internal telemetry for all stable components.
+
+> [!NOTE]
+> - The following categories MUST all be covered, unless justification is given as to why one may
+>   not be applicable.
+> - However, for each category, many reasonable implementations are possible, as long as the
+>   relevant information can be derived from the emitted telemetry; everything after the basic
+>   category description is a recommendation, and is not normative.
+> - Of course, a component may define additional internal telemetry which is not in this list.
+> - Some of this internal telemetry may already be provided by pipeline auto-instrumentation or
+>   helper modules (such as `receiverhelper`, `scraperhelper`, `processorhelper`, or
+>   `exporterhelper`). Please check the documentation to verify which parts, if any, need to be
+>   implemented manually.
+
+**Definition:** In the following, an "item" refers generically to a single log record, metric point,
+or span.
+
+The internal telemetry of a stable pipeline component should allow observing the following:
+
+1. How much data the component receives.
+
+    For receivers, this could be a metric counting requests, received bytes, scraping attempts, etc.
+
+    For other components, this would typically be the number of items received through the
+    `Consumer` API.
+
+2. How much data the component outputs.
+
+    For exporters, this could be a metric counting requests, sent bytes, etc.
+
+    For other components, this would typically be the number of items forwarded to the next
+    component through the `Consumer` API.
+
+3. How much data is dropped because of errors.
+
+    For receivers, this could include a metric counting payloads that could not be parsed in.
+
+    For receivers and exporters that interact with an external service, this could include a metric
+    counting requests that failed because of network errors.
+
+    For processors, this could be an `outcome` (`success` or `failure`) attribute on a "received
+    items" metric defined for point 1.
+
+    The goal is to be able to easily pinpoint the source of data loss in the Collector pipeline, so
+    this should either:
+    - only include errors internal to the component, or;
+    - allow distinguishing said errors from ones originating in an external service, or propagated
+        from downstream Collector components.
+
+4. Details for error conditions.
+
+    This could be in the form of logs or spans detailing the reason for an error. As much detail as
+    necessary should be provided to ease debugging. Processed signal data should not be included for
+    security and privacy reasons.
+
+5. Other possible discrepancies between input and output, if any. This may include:
+
+    - How much data is dropped as part of normal operation (eg. filtered out).
+
+    - How much data is created by the component.
+
+    - How much data is currently held by the component, and how much can be held if there is a fixed
+        capacity.
+    
+        This would typically be an UpDownCounter keeping track of the size of an internal queue, along
+        with a gauge exposing the queue's capacity.
+
+6. Processing performance.
+
+    This could include spans for each operation of the component, or a histogram of end-to-end
+    component latency.
+    
+    The goal is to be able to easily pinpoint the source of latency in the Collector pipeline, so
+    this should either:
+    - only include time spent processing inside the component, or;
+    - allow distinguishing this latency from that caused by an external service, or from time spent
+        in downstream Collector components.
+    
+    As an application of this, components which hold items in a queue should allow differentiating
+    between time spent processing a batch of data and time where the batch is simply waiting in the
+    queue.
+    
+    If multiple spans are emitted for a given batch (before and after a queue for example), they
+    should either belong to the same trace, or have span links between them, so that they can be
+    correlated.
+
+When measuring amounts of data, it is recommended to use "items" as your unit of measure. Where this
+can't easily be done, any relevant unit may be used, as long as zero is a reliable indicator of the
+absence of data. In any case, all metrics should have a defined unit (not "1").
+
+All internal telemetry emitted by a component should have attributes identifying the specific
+component instance that it originates from. This should follow the same conventions as the
+[pipeline universal telemetry](rfcs/component-universal-telemetry.md).
+
+If data can be dropped/created/held at multiple distinct points in a component's pipeline (eg.
+scraping, validation, processing, etc.), it is recommended to define additional attributes to help
+diagnose the specific source of the discrepancy, or to define different signals for each.
+
 ### Deprecated
 
 The component is planned to be removed in a future version and no further support will be provided. Note that new issues will likely not be worked on. When a component enters "deprecated" mode, it is expected to exist for at least two minor releases. See the component's readme file for more details on when a component will cease to exist.