ESQL: Introduce TimestampAware interface/contract #136772
Description
Description
Timeseries related functions require access to the timestamp field (typically @timestamp). Currently, the functions declare a UnresolvedAttribute(@timestamp) inside their constructor:
public MyTimeFunction(Source source, Expression field) {
// delegate to the actual constructor using an UnresolvedAttribute
this(source, field, new UnresolvedAttribute(source, field, MetadataAttribute.TIMESTAMP_FIELD);
}
public MyTimeFunction(Source source, Expression field, Expression timestamp) {
...
}so that the Analyzer injects it however this has several disadvantages:
- it hardcodes the field name which makes the resolution fail in case of a rename for example
ts indices-* | rename x = @timestamp | ... - if the public constructor is called after the analysis phase, the plan becomes invalid
An improved approach is to be explicit about the dependency and instruct the planner to deal with across the entire query in a holistic fashion.
One approach is to introduce a dedicated TimestampAwareBuilder (similar to ConfigurationAware ) to handle the constructor reference or simply rely on the TimestampAware interface to mark the consumer and (potentially a (a TimestampSource for the provider to clearly indicate the @timestamp either through a dedicated method or through its type/"flag" similar to MetadataAttribute but not by name since it's an actual user facing field).
// consumer interface
public interface TimestampAware {
Expression timestamp();
}In the function registry define
protected interface TimestampAwareBuilder<T> {
T build(Source source, Expression field, Expression timestamp);
}
// which would be fed into the function definition:
def(MyTimeFunction.class, time(MyTimeFunction::new, "myTimeFunction"))An alternative to TimestampAwareBuilder is to check if the target function implements TimestampAware interface and if so, look at the constructor field of time Expression based on the @Param annotation information. The former should be more reliable since it applies to all constructors whereas @Param is used for documentation purposes.
The advantage of this approach is the resolution and supplying of the timestamp is externalized from the function into the registry which can wire the field directly similar to the approach taken with Configuration.
It's up to the instantiators of the function to supply the timestamp instead of the function itself, the current situation.