diff --git a/specification/error-handling.md b/specification/error-handling.md
new file mode 100644
index 00000000000..a134e256c9d
--- /dev/null
+++ b/specification/error-handling.md
@@ -0,0 +1,58 @@
+# Error handling in OpenTelemetry
+
+OpenTelemetry generates telemetry data to help users monitor application code.
+In most cases, the work that the library performs is not essential from the perspective of application business logic.
+We assume that users would prefer to lose telemetry data rather than have the library significantly change the behavior of the instrumented application.
+
+OpenTelemetry may be enabled via platform extensibility mechanisms, or dynamically loaded at runtime.
+This makes the use of the library non-obvious for end users, and may even be outside of the application developer's control.
+This makes for some unique requirements with respect to error handling.
+
+## Basic error handling principles
+
+OpenTelemetry implementations MUST NOT throw unhandled exceptions at run time.
+
+1. API methods MUST NOT throw unhandled exceptions when used incorrectly by end users.
+   The API and SDK SHOULD provide safe defaults for missing or invalid arguments.
+   For instance, a name like `empty` may be used if the user passes in `null` as the span name argument during `Span` construction.
+2. The API or SDK may _fail fast_ and cause the application to fail on initialization, e.g. because of a bad user config or environment, but MUST NOT cause the application to fail later at run time, e.g. due to dynamic config settings received from the Collector.
+3. The SDK MUST NOT throw unhandled exceptions for errors in their own operations.
+   For example, an exporter should not throw an exception when it cannot reach the endpoint to which it sends telemetry data.
+
+## Guidance
+
+1. API methods that accept external callbacks MUST handle all errors.
+2. Background tasks (e.g. threads, asynchronous tasks, and spawned processes) should run in the context of a global error handler to ensure that exceptions do not affect the end user application.
+3. Long-running background tasks should not fail permanently in response to internal errors.
+   In general, internal exceptions should only affect the execution context of the request that caused the exception.
+4. Internal error handling should follow language-specific conventions.
+   In general, developers should minimize the scope of error handlers and add special processing for expected exceptions.
+5. Beware external callbacks and overrideable interfaces: Expect them to throw.
+6. Beware to call any methods that wasn't explicitly provided by API and SDK users as a callbacks.
+   Method `ToString` that SDK may decide to call on user object may be badly implemented and lead to stack overflow.
+   It is common that the application never calls this method and this bad implementation would never be caught by an application owner.
+7. Whenever API call returns values that is expected to be non-`null` value - in case of error in processing logic - SDK MUST return a "no-op" or any other "default" object that was (_ideally_) pre-allocated and readily available.
+   This way API call sites will not crash on attempts to access methods and properties of a `null` objects.
+
+## Error handling and performance
+
+Error handling and extensive input validation may cause performance degradation, especially on dynamic languages where the input object types are not guaranteed in compile time.
+Runtime type checks will impact performance and are error prone, exceptions may occur despite the best effort.
+
+It is recommended to have a global exception handling logic that will guarantee that exceptions are not leaking to the user code.
+And make a reasonable trade off of the SDK performance and fullness of type checks that will provide a better on-error behavior and SDK errors troubleshooting.
+
+## Self-diagnostics
+
+All OpenTelemetry libraries -- the API, SDK, exporters, instrumentation adapters, etc. -- are encouraged to expose self-troubleshooting metrics, spans, and other telemetry that can be easily enabled and filtered out by default.
+
+One good example of such telemetry is a `Span` exporter that indicates how much time exporters spend uploading telemetry.
+Another example may be a metric exposed by a `SpanProcessor` that describes the current queue size of telemetry data to be uploaded.
+
+Whenever the library suppresses an error that would otherwise have been exposed to the user, the library SHOULD log the error using language-specific conventions.
+SDKs MAY expose callbacks to allow end users to handle self-diagnostics separately from application code.
+
+## Exceptions to the rule
+
+SDK authors MAY supply settings that allow end users to change the library's default error handling behavior.
+Application developers may want to run with strict error handling in a staging environment to catch invalid uses of the API, or malformed config.